Every weekday morning I host a Mid-Day Stock Market analysis stream at 11:30 AM Eastern and the prep starts hours before that. A significant part of my analysis focuses on different economic indicators and I would open Fred to pull the latest CPI series, look at the EIA site to see what is happening with energy, check AAA for gas prices, look at Zillow’s research for rents, open the Cleveland Fed for their inflation nowcast then check Kalshi to see where prediction markets had the next print. If that wasn’t enough I would look at the Treasury site for yields, have the BLS calendar open to see what’s dropping this week while having the Atlanta Fed open for their GDPNow data. Some mornings I would have 15-20 tabs open trying to make sense of the economic indicators before I finished my coffee.

The part that always annoyed me was that the actual CPI print was describing a month that ended six weeks ago. The rent data was on one lag, the gas data on another, the used-car data on a third. I was the aggregation layer manually trying to figure out what the actual landscape was. At some point the question stopped being how do I make this routine faster and became why doesn’t the thing I actually want exist. What I wanted was simple to describe and apparently impossible to find: one place where the economic data lives, current as of this morning, cross-referenced, with a single honest headline number for what inflation is doing right now. Not last month. Not the month before that. Now.

So I built it. It’s called Nowflation.com, it’s completely free, and this is the full story of why it exists and how it works.

The problem with the official number

Let me be clear up front: this isn’t a hit piece on the Bureau of Labor Statistics. The CPI is a serious, rigorous statistical product, and it’s the authoritative historical record of US consumer prices. Nowflation is measured against it, not pretending to replace it. In my opinion, the CPI has a structural feature that makes it nearly useless for the question I care about, and that feature is time.

Walk through the calendar with me. The June CPI report comes out on July 14th. The prices in that report were collected throughout June so on the day it prints the newest data point in it is already two weeks old and the oldest is six weeks old. Then we all sit around until mid-August for the next one. If something changes in the economy on July 15th the official record won't fully reflect it until September. In a world where gasoline reprices daily and rents reprice with every new lease that's driving by looking at a photograph of the road from last month.

Shelter makes it worse, and shelter is roughly a third of the index. The BLS measures rent by surveying a rotating panel of existing units and each unit only gets sampled about twice a year. Most tenants aren't signing a new lease in any given month so a lease signed today filters into the official shelter number over quarters not weeks. New-lease rent indexes from sources like Zillow and Apartment List lead the official shelter series by roughly a year. We all watched this play out in real time as market rents rolled over hard and official shelter kept printing hot month after month which dragged headline CPI along with it. Anyone watching new-lease data knew disinflation was coming long before the official number admitted it and anyone who only watched the CPI print was a year behind what was actually happening in the economy.

None of this is a conspiracy but rather a methodology. The CPI answers a specific question: what did the average urban household experience across the survey window? That's a legitimate question and the BLS answers it carefully. My question is different: what do prices look like today for the person signing the lease, filling the tank, and pushing the cart? Very few places were answering that question in one place for free every day and aggregating all of the different economic data which was the gap.

So I went and got every API I could

The build started the way most of my builds start and I started with a data audit. I sat down and mapped every public source of US price and macro data I could get programmatic access to. I began with a FRED API key, then direct feeds from the BLS, the BEA, and the Treasury. Then I added the EIA for energy, the USDA for food, and Zillow's research files and Apartment List for new-lease rents. I am using AAA for daily pump prices and Manheim for wholesale used-vehicle auctions. I even added in Kalshi for prediction-market odds on the actual prints. I added source after source and key after key until the pipeline covered the entire consumer basket and most of the macro picture around it. I published on nowflation.com every place where the data is being pulled in from to be as transparent as I can be.

Today the engine ingests 1,068 public series from 24 sources, more than 331,000 observations and counting which all flow into one database. Every observation gets stamped with the date it arrived and that little timestamp turns out to matter enormously which I will discuss when I talk about backtesting. This was a deliberate choice from day one because just about everything is public data. I believe that nowflation.com doesn’t create an edge from magical secret data. I believe that the edge is it’s breadth, speed, transparency, and a willingness to be graded in public. Every number on the site traces back to a source you could go pull yourself. All I did was save you the time of having 15-20 tabs open and aggregated all of the data into once place.

What Nowflation actually is

Strip away all the pages and the site is three products sitting on one pipeline.

The first is the Nowflation Gauge which is a daily CPI-comparable inflation index. As I write this on the morning of July 9, the gauge reads 1.78% year over year. The official CPI for the month of May read 4.2%. That gap of 2.42 percentage points is the whole reason the site exists and I’ll spend a full section below on what it means and why the two numbers disagree.

The second is the Nowflation Forecast which is a call on the next official print. This is published before the release and never touched afterward. Right now the model has June CPI tracking at 4.025% year over year against a Street consensus of 3.9%.

The third is the Scoreboard, where I grade that forecast in public after every single print, side by side with the Cleveland Fed, the Wall Street consensus, and the Kalshi prediction markets. Anyone can put a number on the internet. The scoreboard is what separates a forecast from a guess.

In addition to the Nowflation Gauge, Forecast and Scorecard I wanted to build a site based on how I wanted to view the data. I built a full macro hub that includes GDP nowcasts, labor, housing, money and credit, rates, the fiscal picture, recession odds, state and metro pages, country pages, and a long list of single-item trackers.

Where the numbers come from

Every component of the gauge is fed by the fastest credible public source I could find for it.

Motor fuel comes from daily national pump prices backed by EIA data and a forward model built off futures. As I write this, the national average sits at $3.796 a gallon, and the two-week forward read has it at $3.797, basically flat.

Shelter which is the biggest and most important piece, comes from new-lease rent data from Zillow and Apartment List. This data leads the official series by roughly a year in addition to utilizing home prices and mortgage rates for the ownership side.

I am using Manheim's wholesale auction data for used vehicles which historically front-runs the official used-car index by a couple of months.

Food at home is built from USDA and BLS shelf-price data item by item.

Energy beyond the pump which is electricity and utility gas is being pulled from the EIA.

The broad macro spine and the official series everything gets benchmarked against comes from FRED plus direct BLS, BEA, and Treasury feeds.

Kalshi's markets give me a live read on where real-money traders think the prints land which rounds out the pipeline.

Nowflation is built in the open from many different public sources courtesy of Uncle Sam.

How the gauge engine actually works

This is the part I get asked about most, so let me open the hood properly. The gauge is built on the CPI's own architecture. I take the basket and organize it into 14 components which includes:

• shelter rent

• shelter owned

• motor fuel

• used vehicles

• new vehicles

• food at home

• food away from home

• electricity

• utility gas

• medical care

• apparel

• recreation

• education and communication

• a residual for everything else.

Each component carries its published CPI relative-importance weight so the basket math mirrors the official construction. The index is anchored to a base period of January 2018 equals 100 which also produces one of my favorite stats on the site: cumulative prices are up 33.9% since January 2018.

Then comes the core idea which is the thing that makes it nowflation instead of just a re-plot of the CPI. Each component rides the official series through history and then at the front edge wherever a faster market-price source exists the live data takes over. Fuel reprices every single day from pump prices. Rents update monthly from new-lease data which means my shelter component turns when the market turns instead of four quarters later. Where no faster source exists yet the component honestly carries forward at its latest official reading until the next print.

I want to stress that last part since it's the difference between an honest daily index and a fake one. On a typical day the movement in the gauge is mostly fuel plus whatever the latest monthly rent update did. I'm not pretending to have a real-time feed on medical care prices and the reality is that almost nobody does. What I have is a framework where every component updates at the fastest honest frequency available for it and where the daily attribution tells you exactly what moved. Take a recent daily read where the gauge ticked up about half a basis point, it was motor fuel that contributed essentially all of the increase. That's written right on the page in plain English every morning.

There is another design decision that I want to explain because the gauge ships in two shelter variants and they answer two different questions. The CPI-comparable variant uses rental equivalence, the same conceptual approach the BLS uses just without the survey lag. That's the number you compare against the official CPI and it reads 1.78% today. The Cost-of-Living variant prices ownership the way an actual buyer experiences it which is house price times the current mortgage rate. That's the marginal buyer's reality and it reads 1.70% today. When rates are moving those two views of shelter can tell very different stories and I’d rather show you both than pick one and hide the other.

The headline gauge doesn't travel alone, it ships alongside companion reads for core CPI, supercore services excluding shelter, PCE, and core PCE which is the measure the Fed actually targets. The core CPI forecast for the June print sits at 2.87% against the official 2.9% from May and core PCE tracks at 3.4% against 3.41%. Right on top of the official numbers where the lags are small and far away from them where the lags are large. That's exactly the pattern you'd expect if the methodology is doing its job.

Why my number and the official number disagree

So: 1.78% versus 4.2%. A 2.42-point gap is enormous and I want to be very precise about what it does and doesn't mean.

It does not mean the BLS is lying and it does not mean inflation is really 1.78% in some cosmic sense. The two numbers are measuring different windows of time with different shelter clocks. The official 4.2% describes May weighted toward leases signed over the past couple of years. My 1.78% describes today weighted toward leases being signed right now.

If you look at the component detail the gap explains itself. On rent, the official series has 2.9% year over year. My new-lease-based measure has 0.8%. On utility gas, the official read says up 3.0% while current market data says down 5.6%. Meanwhile, on electricity I'm actually above the official number, 8.4% versus 5.9%, so this isn't a machine built to print a lower figure. It prints what market prices say in whichever direction they point.

Here's how I'd translate the gap: the pipeline is cooler than the record. Market prices today are running well below what the trailing official record shows which historically means the official prints have room to come down as the lagged components catch up to reality. The gauge led on the way up in 2021 and 2022 when anyone watching new leases could see the official number was about to run hot. It leads on the way down too.

Now the honest nuance which matters. That 4.2% May print was up from 3.8% and is the hottest official CPI reading since April 2023. My forecast for the very next print for June which will be reported on July 14th is 4.025% which slightly above the 3.9% consensus. How does a warm near-term forecast square with a 1.78% gauge? Easily once you separate the two products. The forecast is arithmetic about one specific report : what's already baked into June's collection window, base effects and all. The gauge is a statement about where prices are today. In the near term the model says the June print comes in warm. Beyond that the gauge says the direction will trend lower. Holding both of those at once isn't a contradiction. It's the entire point of having both tools.

The forecast, and why I publish a blend instead of my ego

The in-house nowcast model does exactly what you’d expect as it takes everything in the pipeline, the daily fuel data, the rent trajectory, the component carry-forwards, the seasonal factors, and produces a call for the upcoming print. Right now it thinks June runs a touch warmer than consensus and I publish that transparently. I don't nudge my model toward the crowd so I can look smarter after the fact.

The number that I actually put on the tape as the headline forecast isn't my raw model. It's an ensemble which is an inverse-error-weighted blend of the forecasters who've earned a real, graded track record. The mechanics are simple and a little ruthless as each forecaster's weight is 1 divided by its own historical error so more accurate forecasters count for more and the weights recompute themselves after every print grades. Nobody gets grandfathered in and a source like Kalshi stays display-only until it accrues enough graded prints to earn a seat so the headline never leans on a forecaster with no evidence behind it.

Grading myself in public, and never rewriting the record

This is the section I care about most because the financial internet is full of people making calls and quietly forgetting the ones that missed. I built Nowflation to make that impossible for me where three mechanisms do the work.

First is the backtest. When I test how the model would have done historically, the engine replays each past print seeing only the data that existed the day before that print. Later revisions are hidden and the future data is hidden. That's the real reason every observation gets stamped with its arrival date: so the backtest can honestly reconstruct what was knowable at the time, with zero look-ahead. A backtest that peeks isn't a test it’s a marketing document.

Second, the scoreboard is live and automatic. On print mornings the system captures the official number the moment it drops and grades every forecaster against it. The public leaderboard shows each forecaster's average error, its bias, and its win count, mine included. Nowflation is graded against the Cleveland Fed, the Street, and Kalshi. When Nowflation misses, the miss goes on the board and it stays there.

Third, every forecast gets frozen before the print into a dated timestamped receipt committed to the site's code repository. The locked timestamp, the scheduled release time, and every forecaster's number is collected and published and cannot be edited afterward. The same discipline applies to the gauge's own history: values are stored and never silently revised. When the methodology changes, and it will, since this is a living project, the change ships in a dated public changelog. You will never load the site and discover that the past has been improved.

The engineering: one machine and a lot of guardrails

People sometimes assume there’s a team behind this but there isn’t. The engine runs on a Mac mini sitting on my desk, Node.js and a SQLite database, chewing through 24 sources every morning before the market opens.

What the public touches is a static site rebuilt each morning and pushed to a global CDN. That's why every page loads in under a second and it’s why a hiccup on my end can’t take the site down. The engine and the site are decoupled on purpose. The publish itself has to earn its way out the door every morning. Before anything deploys the system runs a self-test with a dozen checks, six of which can block the entire publish on their own. Every series carries a day-over-day sanity guardrail so one bad API response can’t poison the index. I built a drift guard on the outputs themselves: if the gauge moved more than three quarters of a point in a day, or the forecast more than half a point, the deploy gets blocked, the last good version of the site stays live, and my phone lights up. Bad data cannot quietly reach the air.

You don't have to take my word for the pipeline's health, either. There's a public status page showing the live self-test, the freshness of every single source, and publish health. The observability is itself part of the trust. If something's stale you can see that it’s stale.

What Nowflation is not

Nowflation is not the official number and it’s not trying to be. The CPI remains the authoritative record. The gauge is benchmarked against it, leads it, argues with it, and never overrules it.

It's not a magic tick-by-tick feed of the whole basket. Day to day, the honest movement is fuel plus the monthly rent update with the slower components carried forward from the last print. A few overlays like the Manheim used-car splice are still warming up while they accrue enough clean history, and until then they fall back to carry-forward.

It’s not a black box. The methodology page lists every series, every source, and every weight, all 1,068 series across all 24 sources, and every one of them is public.

For now it runs on one machine in my home. That's a single point of failure which I am honest about and I’m actively hardening it. The static architecture insulates you from it, but the engine is the engine.

The other half of the reason: one place for my community

Everything above is the analytical case. Here’s the personal one. I didn’t build this just so my own mornings would be easier. I built it so the people in my community would have one place to go. For years, the answer to where do I check this stuff was a list of nine government websites and a shrug. Now it’s one URL, nowflation.com.

The site has grown well past the gauge as it now includes:

• A heat check that scores the whole economy on a heating-versus-cooling scale.

• A page that compares my numbers to the BLS line by line.

• A personal inflation calculator where you re-weight the basket to your own life, your rent, your commute, your grocery cart, and get your inflation rate instead of the average one.

• A GDP nowcast

• Jobs and labor pages

• Housing and affordability

• Money and credit, rates

• The fiscal picture

• Recession odds

• State pages

• Metro pages

• Country pages for more than 200 economies

• Single-item trackers: eggs, gas, coffee, milk, bread, bacon, ground beef, chicken, etc

All in, it’s more than 60 pages plus a long programmatic tail which is organized so you can get from the headline to the detail in two clicks. Everything is also free with no paywall, no login, no unlock-the-last-30-days gate. There are free CSV downloads on the data pages, an open API, and embeddable widgets so you can drop any chart onto your own site. The robots file even welcomes AI crawlers on purpose: when someone asks an AI assistant what inflation is doing today, I want the answer to be citable, sourced, and current.

The whole thing updates every morning before the open with today’s read written straight from the data: what moved, by how much, and why. One page instead of 20 tabs. That was the mission statement the entire time.

Kick the tires

Here’s how I’d suggest you use it. Check the gauge against the official number and watch the gap. Read the daily attribution so you know what’s actually moving instead of guessing. Before a print, look at the forecast and where it sits against the Street and Kalshi. After the print, go straight to the scoreboard and see who was right, me included.

June CPI drops on July 14. My number is already locked, timestamped, and sitting on the board at 4.025% against a 3.9% consensus. Whatever happens, the grade goes up in public, right next to everyone else’s.

That’s Nowflation. Built out of one guy’s frustration with 20 browser tabs, running on a small computer with strict rules, graded in the open, and free on purpose. Come see what the data says today: nowflation.com.

The Fed held the funds rate at 3.5% to 3.75% and that part was fine. What wasn’t fine was the dot plot. Nine of the eighteen officials who submitted projections now pencil in at least one rate hike before year-end. Of those nine officials, five of them believe we need 2 hikes in 2026 and one of them actually thinks we need three hikes this year. The median dot for year-end jumped to 3.8% from 3.4% back in March.

The market got the message. The 2-year Treasury yield ripped about 16 basis points higher to 4.21% which is its highest in over a year. The dollar had its best day in almost a year, gold fell more than 2%, and the Dow gave back 507 points. Traders now put the odds of a September hike near 49%, up from 27% the day before.

The Inflation They’re Fighting Is Mostly Oil, and Oil Just Collapsed

The hawks are anchoring on a 4.2% CPI print which is the highest since April 2023. Yes, this is double their 2% target but that number is a lagging photograph of a fire that’s already being put out. Truflation which rebuilds the CPI basket every day from more than 30 real-time data sources and tends to lead the official print by roughly 45 days has a YoY inflation print at 1.81%. The real-time read is 1.81% and the official read is 4.2% which is a gap of 2.39% and the real-time number is the one pointing where we’re headed.

The main reason for the gap is oil. The entire inflation scare of the last few months was built on the Iran war and the threat to the Strait of Hormuz. WTI crude ran up more than 45% during the conflict and that fed straight into gasoline, freight, and the broad goods basket. Now look at what oil is doing. WTI is below $76 a barrel which is a three month low and down nearly 40% from its conflict peak. The US and Iran are set to sign an interim deal Friday in Switzerland, the Strait is reopening, Iranian barrels are coming back online, and more than 100 ships carrying crude are about to be released into the market. The IEA is now warning about a glut rather than a shortage with 2027 supply set to grow roughly four times faster than demand.

Here’s the part the hawks are missing, and Warsh said it himself in the room. The Fed can’t do much about the price of oil or a dozen eggs and a move in those prices does not have first order consequences for what the Fed is doing according to him. The Fed’s job is to stop a supply shock from broadening into second and third round effects. The whole question is whether the oil spike is leaking into everything else. Truflation at 1.81% is the answer and the answer is no. You don’t raise rates to fight a supply shock that’s already reversing. That’s the textbook policy error.

A Strong Jobs Report With Calm Wages Isn’t Inflationary

The bears want to use 172,000 jobs as a reason to fear inflation and that chain breaks at the second link. Jobs don’t show up in prices, labor costs do and labor costs are quiet. The report itself was was strong as 172,000 jobs were added in May against the consensus of 80,000. March and April were revised up by a combined 93,000 and that was the strongest three month stretch of hiring in over two years. Unemployment held steady at 4.3%.

Here’s what the bears skip over. Wages are the bridge between jobs and prices and wages are behaving. Average hourly earnings rose 0.3% on the month and just 3.4% over the year. With CPI at 4.2% real wages are slightly negative which means workers are losing ground not bidding pay up faster than prices. That last point is the one that matters since rising real wages are the single ingredient a wage-price spiral can’t do without.

The whole bear case leans on the Phillips curve and that foundation has been cracking for decades. The idea that low unemployment has to produce inflation keeps getting disproven. In the late 1990s unemployment was under 4% while inflation fell. 2019 hit a 50-year low near 3.5% with core inflation under 2%. The thread running through every one of those low-inflation booms is productivity and we’ve got it again.

Services turn the argument inside out because two-thirds of consumer spending goes to services. When it comes down to it services prices are mostly just labor costs passed through. Every new hire adds a unit of supply at the same moment it adds a paycheck of demand. Inflation only shows up when demand exceeds supply and a jobs report like this one grows both sides at once.

On top of that pricing power is gone. Frito-Lay has cut some prices by as much as 15% and Walmart and Costco have been absorbing input costs through automation instead of passing them along. Core PPI has run hotter than core CPI for months and that gap is the tell. Companies are eating the cost in their margins rather than handing it to the consumer.

I want to be clear about one soft spot since it cuts the opposite way from what the hawks assume. The long-term unemployed are up 524,000 over the past year and now make up 27.5% of everyone out of work and financial-sector payrolls are down 107,000 since last May. The labor market is solid but it’s fraying at the edges. That isn’t an argument for a hike its an argument that there’s room to cut.

Productivity Is Rising, and Jobs Don’t Show Up in CPI

This is the structural reason the old models are continuously wrong. Output per hour is climbing so the same wage growth produces less price pressure than it used to. Unit labor costs are the only labor number that actually feeds prices and they’re tiny. Unit labor costs rose just 0.5% over the last four quarters. That’s wage growth net of productivity and it’s one of the lowest-pressure labor-cost backdrops in decades. The entire labor-cost pressure in this economy right now is half a percent.

Productivity is running well above the prior cycle. Output per hour is up 2.8% YoY and across the current cycle since late 2019 it’s grown at a 2.1% annual pace versus 1.5% in the prior one. More output per worker means a given wage gain gets diluted across more production. Workers are also taking the smallest slice on record as labors share of output has fallen to about 54% which is the lowest since the series began in 1947. When labor can’t capture its own productivity gains it has no leverage to drive a wage-price spiral. Those gains are flowing to margin and to owners not into a self-reinforcing pay loop.

Here’s the CPI mechanism point the bears keep skipping. A strong jobs number isn’t an input to the price index. CPI measures what consumers pay. The only path from jobs to prices runs through wages and then through unit labor costs and both of those are calm. The bears are reacting to a number that doesn’t even feed the thing they’re afraid of and a big reason is AI. The productivity increase correlates to AI tools getting embedded across business functions and it’s also tempering wage demands because workers feel they have less leverage. Both of those push disinflation not inflation. This is the supply-side story Warsh was gesturing at and its showing up in the hard data.

Standing Still Is Already Tightening

This is the piece many individuals get backwards and Bill Ackman has been one of the clearest voices on it. What hits the economy isn’t the nominal rate it’s the real rate which is the nominal rate minus inflation. As inflation falls and the Fed holds nominal rates flat the real rate climbs all by itself. The Fed gets tighter while doing nothing. Here is the actual math behind it. The effective fed funds rate is about 3.63% and if you use the real-time inflation read of 1.81% the real fed funds rate is already running near 1.8% and it’s rising as oil-driven inflation rolls off. That is a restrictive setting and it gets more restrictive every week the disinflation continues.

So the framing of “hold versus hike” is wrong from the start. Holding is a passive tightening. If the Fed wants policy to stay roughly where it is in real terms while inflation comes down it has to cut just to stand still. Two cuts in the back half don’t really ease policy so much as keep it from grinding tighter into a slowing economy.

The Real Recession Risk Is the Refinancing Wall

A lot of debt was termed out at near-zero during the pandemic and now it has to be refinanced into a much higher rate. The wall is real and it’s hitting now as roughly $1.35 trillion of non-financial corporate debt matures in 2026 and has to be refinanced. S&P counts about $7.3 trillion of rated corporate debt maturing over the next three years with close to $1.5 trillion of that speculative-grade which is the most exposed.

In todays environment the rate gap between where debt was issued and where it is refinanced at is brutal. Much of this paper was issued in the ultra-low window of 2020 and 2021. Refinancing it today means rolling low-coupon debt into 6% and higher. On the commercial real estate side loans struck at roughly 4.1% to 4.7% are refinancing near 6.5% and over $1.26 trillion of CRE debt matures through 2027.

Based on Goldmans math, for every extra dollar a company spends on interest as it refinances it tends to cut about 10 cents of capital spending and about 20 cents of labor costs. Goldman has framed forced refinancing at elevated rates as something that can run into thousands of job losses a month and it scales higher the longer rates stay up. Now follow the chain to the recession call because higher refi cost leads to cuts in capex and headcount which causes unemployment to increase and rising unemployment is the single most reliable leading indicator of recession. Recessions are what produce real bear markets so holding rates too high here isn’t a neutral wait-and-see stance. It’s actively loading risk onto the most leveraged third of corporate America.

The Consumer Is Tapped Out at Record Carrying Costs

The same rate that squeezes corporate refinancing lands directly on a household that’s already stretched thin. Total credit card debt sits at $1.25 trillion and the average APR on cards carrying a balance is north of 21% which is roughly 500 basis points above where it was before COVID. Nearly half of cardholders carry a balance and 61% of the ones that carry debt have been stuck there for at least a year. Credit Card rates, auto loans near 7.5%%, and floating-rate consumer debt all move with the Fed. A hike pulls more money out of household budgets and hands it to interest payments which slows spending and feeds right back into the slowdown that’s already underway. You don’t tighten the screws on a consumer who’s putting groceries on a 22% card to fight an oil spike that’s already deflating.

Maximum Employment Matters More Than 2% Right Now

This is the philosophical core of my position and it’s a dual-mandate argument not a dovish wish. The Fed is charged with both price stability and maximum employment. With underlying inflation near target and the labor market on the fragile side of the ledger employment is the binding constraint right now. The overshoot is also the wrong kind to fight with rates. Rate hikes work by crushing demand. They do nothing about a closed shipping strait or a war premium in crude. Holding rates punishingly high to offset an energy spike just destroys demand and jobs while doing nothing to the actual source of the price pressure which is already reversing on its own.

The asymmetry is the whole point. An energy spike self-corrects when supply returns and supply is returning. A cracked labor market does not self-correct. Once layoffs start and unemployment turns up it feeds on itself. The costlier mistake by far is staying too tight and letting the maturity wall push unemployment higher not easing a few months early into a supply shock that’s fading.

Keep in mind that two percent is a target not a ceiling to defend at any cost. Underlying inflation by trimmed-mean PCE, the kind of core measure Warsh has favored is already running around 2.3%. Treating a temporary energy-driven headline number as a reason to risk the employment mandate gets the cost-benefit exactly backward. With pay up 3.4% and prices up more than that households are losing ground in real terms and the fastest way to ease that squeeze is to let energy roll over and stop tightening into a slowdown.

The Case for Two Cuts in the Back Half of 2026

When you put all of this together the path is obvious. As the oil shock rolls off and the real-time data confirms inflation running near or below 2% the real fed funds rate is climbing on its own. Two 25 basis point cuts in the back half which would take the funds rate to roughly 3.0% to 3.25% would do three things at once. They’d keep real policy from grinding tighter into a slowdown, take pressure off the refinancing wall before it forces layoffs, and give a labor market that’s already cooling at the edges room to firm up instead of cracking.

This is insurance and the timing is the whole point. The Sahm-rule dynamic is brutal because once unemployment starts climbing it tends to climb fast. By the time the data confirms a recession it’s too late to prevent it. You cut before the margin-to-layoffs chain starts not after. Waiting for the official CPI to prove disinflation means waiting 45 days past the point the real-time data already showed it and acting only once the jobs are already gone. There’s also a tailwind that gives the Fed cover to ease and that’s productivity. Warsh framed AI as American ingenuity and spent real time on the supply side today. David Sacks has been pounding the table that AI was 75% of GDP growth in the first quarter and that AI capex alone is a 2.5% tailwind to growth this year that is likely going to climb past 3% next year. This is before you even count the productivity gains downstream. A real productivity boom raises potential output and lowers the neutral rate which is exactly the 1990s setup where the Fed could run easier policy without lighting up inflation. Warsh said it himself: we don’t face a cruel choice between jobs and prices. You can have strong growth, low inflation, and strong employment together. That’s an argument for easing into the boom, not choking it.

Give Warsh Credit, This Was a Good for the Institution

I thought Warsh ran a strong first meeting and the direction he’s taking the Fed is the right one. The single best thing he announced was the data task force. Warsh said out loud what people like me have been saying for years. Most of the data the Fed steers by comes from old-fashioned survey methods which is a national accounts picture that in his words looks very little like the U.S. economy in 2026 with response rates that have collapsed and questions written for a different era. He wants real contemporaneous data instead of what he called an echo of history. This is the entire ballgame because if the Fed had been looking at real-time inflation today they’d have seen 1.81% not reacted to a 4.2% echo. The fact that Truflation already publishes a Warsh trimmed mean index tells you the real-time world is ready for him.

Bringing in outside experts is the right instinct also. Warsh was explicit that he isn’t outsourcing decisions and he’s widening the aperture starting from first principles and forcing what he called a family fight among smart people before the committee decides. That’s how you fight groupthink and it’s a sharp break from a Fed that too often moved as a herd. Dropping forward guidance and letting markets price off data is also correct even if traders hated it. Warsh’s logic is right because when the markets just reflect the Fed’s words back at the Fed the central bank binds itself to the single most valuable signal it has which is market prices reacting to the real economy. We need less hand holding and more information from industry experts.

I support his decision to decline submitting his own dot as dots create false precision and fake commitment. Warsh noted that everyone submitted them with pencils and big erasers while not hearing much conviction in the room. Treating the dots as low-conviction guesses rather than promises is honest and it quietly undercuts the nine hawks more than anything I could write. By the new Chair’s own account those dots are pencil marks not commitments.

The Experts I’d Ask To Participate on These Task Forces

Since Warsh is recruiting here’s who I’d want in the building and exactly where.

Jamie Dimon belongs on the data task force. Nobody on earth has better real-time economic data than JPMorgan. They see card spending, deposit flows, and loan demand across tens of millions of households and businesses as it happens. That’s the actionable data Warsh keeps describing and Dimon has preached real-time data discipline for years. Dimon does lean hawkish on inflation. He’s the one calling it the skunk at the party and warning that deficits and global remilitarization are structurally inflationary. You want a rigorous skeptic who stress-tests the optimistic case and you want his data pipes in the building. He’s also been clear that chipping away at Fed independence backfires and pushes rates higher which is the right principle to have in the room.

Brad Jacobs belongs right next to Dimon on the data task force considering he spent spent his whole career building real-time read-throughs on the industrial economy, from United Rentals to XPO to GXO and now QXO in building-products distribution. Freight volumes, equipment rentals, and construction-materials demand are some of the cleanest leading indicators we have and they move weeks before the official series catch up. Jacobs is also one of the best capital allocators alive so he understands exactly how a rate move ripples through the cost of capital for real companies making real decisions. Put him where the Fed is hunting for faster, ground-level signals on activity.

David Friedberg belongs on the data task force with them considering he is a data scientist at his core. He built The Climate Corporation by turning real-time weather and agricultural data into priced risk and then sold it for around a billion dollars. He thinks in predictive models and live data feeds not quarterly surveys which is precisely the upgrade Warsh is after. If the Fed wants new analytic methods and new sources Friedberg has spent his career building exactly that.

Bill Ackman belongs on the inflation frameworks task force. Ackman’s real-rate insight is precisely the frame that group needs. As inflation falls and nominal rates hold, real rates are quietly rising so the framework has to think in real terms not nominal ones. He also called the oil-driven inflation correctly as a war-driven, weeks-not-months phenomenon, and he’s been right that the back half sets up well on AI and infrastructure spending once the war resolves. He thinks about the rate that reaches the real economy which is what a modern inflation framework should be built around.

Mark Cuban belongs on the inflation frameworks task force too and not for the reason people would guess. With Cost Plus Drugs he rebuilt drug pricing from the ground up by stripping out the middlemen and showing exactly how much opacity sits between a product’s true cost and what Americans actually pay. Healthcare is one of the stickiest and hardest-to-model pieces of core services inflation and Cuban has hands-on knowledge of how those prices get built. Add the visibility he has into pricing and costs across hundreds of small businesses from his investing and he brings a view of inflation that no academic model captures.

Chamath Palihapitiya belongs on the inflation frameworks task force. For years he’s hammered the same pointthat you can’t talk about inflation without talking about the money supply and the liquidity the Fed pumps into the system. Whatever you make of his individual calls, that monetarist lens, where the money comes from and where it goes is exactly the first-principles question this task force is meant to wrestle with. He’ll push the group to start at the plumbing instead of the symptoms.

David Sacks belongs on the productivity and jobs task force. This one is almost too obvious because that task force is literally about AI’s impact on output and employment. Sacks now co-chairs the President’s Council of Advisors on Science and Technology after running AI policy for the administration. He’s the sharpest voice on AI as a supply-side productivity force which is the same lens Warsh used when he called AI American ingenuity. Put him where the Fed is trying to figure out how a general purpose technology reshapes potential output and let him make the supply-side case in full.

Elon Musk belongs alongside Sacks on the productivity and jobs task force and he might be the single most important voice in the room for it. Nobody is pushing the frontier of automation, robotics, and applied AI harder from Tesla’s factories to Optimus to xAI. If the Fed wants to understand how a general purpose technology actually reshapes output per worker and the demand for labor he’s living it at industrial scale every single day. He also runs one of the largest real-time information networks on the planet in X, which the data task force could borrow from as well.

Peter Thiel belongs on the productivity and jobs task force. His entire intellectual project is whether technology is actually delivering progress or whether we’ve been stuck in stagnation dressed up as innovation and that is the productivity debate in a single sentence. Between Founders Fund and Palantir he’s been funding and building the frontier of applied technology for two decades. Drop him in the room where the Fed is trying to decide whether AI is a real productivity break or just more capex and let him pressure-test the optimism.

Stanley Druckenmiller belongs on the balance sheet task force and frankly the Fed should have been listening to him for years. He’s one of the greatest macro investors who ever lived. He reads the bond market and liquidity conditions better than almost anyone breathing and he’s been blunt that the Fed misjudged both the easing and the tightening cycle. The balance sheet group is wrestling with the ample-reserves regime and how the Fed’s footprint moves markets, and Druckenmiller has traded around exactly that plumbing his whole career. You want his read on what the balance sheet is doing to financial conditions.

Larry Fink belongs right next to Druck on the balance sheet task force. He runs BlackRock which is the largest asset manager on earth and the Fed’s balance sheet is mostly Treasuries and mortgage-backed securities. These are the exact markets BlackRock lives in every day. When the Fed needed help running its asset purchases in 2020 it called BlackRock so none of this is theoretical for Fink. Add the real-time risk lens of the Aladdin platform and he brings both the market plumbing and the data this group needs.

Patrick Bet-David belongs on the communications task force. Warsh kept circling back to the problem of reaching regular people, the person in the milk aisle watching prices climb faster than their paycheck, and that’s exactly the gap PBD fills. He built one of the biggest finance and business media platforms in the country by explaining money in plain language and he came up through financial services so he knows the material cold. A central bank that wants to talk to households instead of only to bond traders should have someone in the room who knows how to do it.

That’s a deliberately mixed bench by background and worldview and that’s the point. Warsh said he wants task forces that can have a real family fight before the committee decides. This group would absolutely give him one.

Bottom Line

The Fed held rates in place which was fine. The nine officials reaching for hikes are making a classic late-cycle error. They’re tightening into a supply shock that’s already reversing, treating a strong jobs number as inflationary when unit labor costs are running at half a percent, and threatening to detonate the refinancing wall that’s the real recession risk. The real-time data says inflation is at 1.81%, underlying inflation is near 2.3%, oil is in free fall, wages are trailing prices, and real rates are rising on their own. The right move isn’t one hike or two. It’s two cuts in the back half of 2026 taken as insurance before the margin-to-layoffs chain starts with a productivity boom giving the Fed cover to ease.

Warsh got the big things right including the real-time data push, the outside experts, the honesty about forward guidance and the dots. Now he has to win the argument inside the room against nine people reading last quarter’s oil prices off a lagging chart. Now we need to see what he can accomplish behind closed doors over the next several months.

Disclaimer

This article is for informational and educational purposes only and reflects the personal opinions and analysis of Steven Fiorillo as of the date of publication. Nothing in this letter constitutes investment advice, financial advice, trading advice, tax advice, legal advice, or a recommendation to buy, sell, or hold any security, commodity, or other financial instrument. Readers should not treat any of the content as such.

The author is not a registered investment advisor, broker-dealer, or financial planner. The analysis presented here is based on publicly available information, including public company filings, earnings releases, third-party research, and other sources believed to be reliable at the time of writing. However, no representation or warranty, express or implied, is made as to the accuracy, completeness, timeliness, or reliability of the information presented. Forecasts, projections, and forward-looking statements are inherently uncertain and subject to revision. Actual results may differ materially from any projections discussed.

Past performance is not indicative of future results. Any investment in securities discussed or referenced in this letter involves risk of loss, including the possible loss of principal. Readers should perform their own independent research and consult with a licensed financial advisor, tax professional, and legal counsel before making any investment decision.

Since Friday I have been arguing that the market didn’t sell off on the jobs report and the pushback always lands in the same place. The rhetorical methodology is that strong jobs mean a hot economy, a hot economy means inflation, and inflation means the Fed can’t cut. I think that chain of logic breaks at the second link and I don’t think it’s particularly close. After I lay out my reasoning I will explain why I believe the market fell and it wasn’t the job report.

The selloff is about a model, not the data

Lets start with why the market reacted the way it did. Kevin Warsh hasn’t said a word since taking over the Fed and he may not until the June 17 decision. In that vacuum the hawks on the committee are whispering, and traders are pricing in the possibility that the Fed looks at 172,000 jobs and concludes it needs to lean against the economy. That fear comes straight out of the Phillips curve which is the old framework that says when growth and employment go up inflation follows.

The problem with the Phillips curve is that it keeps failing the test of history. In the late 1990s unemployment fell below 4% while inflation fell alongside it. In 2019 unemployment hit 3.5% which was a 50-year low and core inflation still ran under the Fed’s 2% target. The most vibrant growth periods in American history have come with lower inflation than anyone expected and the reason is always the same, productivity.

For what it’s worth Kevin Warsh comes from the supply-side school and he doesn’t view strong employment as bad news that needs to be punished. The market is trading like the old framework still runs the Fed. I don’t think it does anymore.

The only labor market number that matters for inflation

Here’s the question I ask anyone who tells me jobs cause inflation: what’s the transmission mechanism? Jobs don’t show up in the CPI, labor costs do. Labor costs aren’t just wages, they’re wages relative to what each worker produces. That’s unit labor costs and it’s the number this entire debate should be fought over.

Here is the math from Friday’s job report. Average hourly earnings rose 0.3% on the month which annualizes to about 3.6%. Productivity is running at 2.8% and likely climbing as AI and robotics work through the economy. The official unit labor cost data which uses a broader compensation measure than hourly earnings, has that pressure at roughly half a percent. Even the crude subtraction from Friday’s prints only gets you to about 0.8%. Roughly half a percent is the labor cost pressure in this economy right now.

Top provide some context the 1970s wage-price spiral that everyone’s framework was built on ran with unit labor costs in the double digits. Oil kept rising, unions struck, workers quit for higher pay, and companies bid up compensation just to stop the turnover. Every one of those mechanics is absent today. If anything the psychology runs the other way as workers are worried about AI and they’re tempering wage demands as a result.

You can see the same story in the corporate data. Revenue growth at nonfinancial companies is accelerating while private payroll growth decelerates. The pre-tax domestic profits as a share of GDP are pushing toward record highs. Companies are producing more without proportionally more labor cost. That’s a productivity boom, and productivity booms are disinflationary by definition.

When someone tells me 172,000 jobs is inflationary my answer is that 172,000 jobs with unit labor costs growing under 1% puts less pressure on prices than 88,000 jobs would have at the 1970s cost structures.

We’re a services economy, and that breaks the bear case

This is the part of the argument I think gets ignored the most. Roughly two-thirds of consumer spending in this country is services, not goods. Services prices are overwhelmingly labor costs passed through to the customer. From your haircut to your childs daycare, the price tracks what the people providing it get paid relative to what they produce. That means the single biggest chunk of the price index lives and dies on the unit labor cost math I just discussed. With that spread under 1% the dominant part of the inflation basket has almost no cost impulse behind it.

There is also a second piece that people miss. A new hire in services adds supply and demand at the same time. The newly employed nurse, mechanic or analyst earns a paycheck earns a paycheck but they also produce services. Hiring expands the economy’s capacity to deliver the exact things consumers spend money on. Inflation happens when demand outruns supply. An employment report where supply and demand grow together with increased productivity doesn’t get you there.

The composition point also flips the bear argument on its head. If strong hiring is going to show up as inflation it has to come through wages. Wages just printed 0.3%. The bears are betting on the one data point that came in soft.

Companies are cutting prices, not raising them

Inflation requires pricing power and the evidence says it’s gone. Frito-Lay which is owned by PepsiCo cut prices on some products by 15%. Walmart and Costco have both said they’re not passing cost increases through at anywhere near the rate you’d expect and they’re crediting efficiency gains from AI and automation for the room to do it.

The producer price data backs them up as Core PPI has been running above core CPI for several months. This means input costs are rising faster than the prices consumers pay. Companies are eating the difference through productivity or through margin. Yes, PPI and CPI track different baskets so the spread is a noisy signal on its own but when you pair it with what management teams are saying out loud the read is straightforward. Companies tested higher prices over the past few years then watched the demand fall off which produced an lesson. Consumers did the disciplining that the Fed thinks is its job.

Oil is a war premium with an expiration date

The honest counterpoint to everything I discussed above is headline CPI sitting in the mid 3s. This is an energy issue as inflation was running in the low 3% range then dipped below 3% before the Iran conflict started and crude is up 55% since then. None of that traces back to American hiring. It traces to a war plus manufacturers building precautionary inventory in case the conflict drags on which pulls demand forward and then gives it back.

The supply response is already forming as Venezuela has raised production and the UAE is pumping more after effectively walking out of OPEC. The last time quota discipline broke down like this was 1986 when Saudi Arabia decided to match every cheater barrel for barrel and crude collapsed from the low $30s to around $10. I’m not predicting a two-thirds crash but the direction once this conflict settles isn’t a mystery as energy math is mechanical. Once the price stops rising its YoY contribution to CPI fades to zero. If your inflation thesis depends on a war premium you’re not arguing that the jobs report causes inflation, you’re arguing that geopolitics does and geopolitical premiums mean-revert.

Ask the bond market

The deepest most inflation-sensitive market on the planet looked at all the same data and shrugged. The 10-year yield has held a 4.4% to 4.6% range through the entire oil spike and it did that while absorbing serious foreign selling. Japan dumped $76 billion of Treasuries in the latest month alone while China, India, and Turkey also trimmed. The outcome was that yields barely moved.

The shape of the curve says even more. The 10-year minus 2-year spread is flattening from a peak of about 70 basis points this cycle. A bond market pricing an inflationary boom steepens. A flattening curve says the long end sees disinflation maybe even deflationary undercurrents despite crude being up on the war premium. Gold is telling the same story as it peaked the day Warsh was appointed and hasn’t looked back. The market’s oldest inflation hedge topped out the moment a credibility-focused chairman got the job.

There’s no monetary fuel either

When you strip everything else away inflation still needs money behind it. M2 is growing at 4.7% to 4.8% as of April. Since the 1990s and CPI has typically run below M2 growth. When I compare this to 2021 M2 was expanding north of 25% a year at the peak of the stimulus era which became the fuel. What happened in 2021 is not happening in 2026.

Conclusion

Friday’s report described an economy adding workers, growing output faster than labor costs, and watching inflation decelerate in real time. That’s not a problem the Fed needs to fix. The risk isn’t that the Fed is behind on inflation, the risk is that it repeats 2022 when it treated a supply shock with rate hikes which restricts the supply you need. This would likely stretch a shock from months to years.

Hiking into this backdrop could slow the exact CapEx boom driving the productivity gains. Manufacturing construction has broken above a ceiling that held since the dot-com era, and tech CapEx near 2% of GDP has room to run toward the 5% to 6% which would mimic what was achieved in the industrial revolution. I believe that Warsh understands all of this better than the market is giving him credit for. Strong jobs with half a percent unit labor cost growth isn’t a resemblance of the 1970s. It’s actually the opposite and at some point the market is going to trade it that way.

What I believe is a more likely scenario for why the market violently sold off on Friday is that many people are trying to front run the SpaceX IPO to raise capital. There is more than $8 trillion sitting in money markets on the sideline which will not be earmarked for this event. If people were going to deploy that capital they would have done it already as we have experienced a strong bull market with the market reaching new all-time highs. Investors likely sold down assets to free up capital heading into the weekend and then due to how quicky the market reacts it probably caused algos to see an opportunity to trade off of it. As prices moved lower at a rapid pace it is likely that stop losses were triggered and people also got margin called which forced them to sell into weakness.

I believe the selloff was less about the jobs report and more about freeing up capital for a landmark IPO.

Disclaimer

This article is for informational and educational purposes only and reflects the personal opinions and analysis of Steven Fiorillo as of the date of publication. Nothing in this letter constitutes investment advice, financial advice, trading advice, tax advice, legal advice, or a recommendation to buy, sell, or hold any security, commodity, or other financial instrument. Readers should not treat any of the content as such.

The author is not a registered investment advisor, broker-dealer, or financial planner. The analysis presented here is based on publicly available information, including public company filings, earnings releases, third-party research, and other sources believed to be reliable at the time of writing. However, no representation or warranty, express or implied, is made as to the accuracy, completeness, timeliness, or reliability of the information presented. Forecasts, projections, and forward-looking statements are inherently uncertain and subject to revision. Actual results may differ materially from any projections discussed.

Past performance is not indicative of future results. Any investment in securities discussed or referenced in this letter involves risk of loss, including the possible loss of principal. Readers should perform their own independent research and consult with a licensed financial advisor, tax professional, and legal counsel before making any investment decision.

Here’s the question: when you click buy on a stock, what did you actually purchase?

Most people would likely answer “shares of Apple” or “some Nvidia” and move on. There are many people who call themselves investors that are simply buying a ticker symbol hoping it goes up. What you actually bought is fractional ownership of a real operating business and that ownership comes with a claim on every dollar of revenue the company generates and every dollar of profit it keeps. Once that idea clicks your entire relationship with the stock market should change.

A Share Is a Slice of a Business Not a Lottery Ticket

Every public company divides its ownership into a set number of shares. Apple has roughly 14.7 billion shares outstanding and when you purchase just 1 share your purchasing one fourteen-point-seven-billionth of the entire enterprise. If you purchase 100 shares you own roughly 0.0000007% of Apple. That percentage sounds ridiculously small until you think about what it’s a percentage of. Apple is a $4.6 trillion business that sold $416.2 billion worth of iPhones, Macs, services, and wearables in 2025. You don’t need a big slice of that pie for your slice to be valuable.

Owning common stock entitles you to three things which are a proportional claim on the company’s earnings and assets, a vote on directors and major corporate actions, and your cut of any dividend the board declares. The voting rights get the headlines during proxy season while the dividend gets the attention from income investors. The claim on earnings is the one that actually drives long-term returns and it’s the one most investors rarely stop to think about.

Your Claim on Revenue and Earnings, in Real Dollars

In Apple’s 2025 fiscal year, and I am using the words fiscal year because they do not report on a traditional calendar year, Apple generated $416.2 billion in revenue and $112 billion in net income across roughly 15 billion diluted shares. When you do the math each share was backed by $27.84 of revenue and $7.46 in earnings. When you apply this to a 100 share position the business generated roughly $2,774 in sales on your behalf and earned about $746 in profit. Every iPhone 17 sold in Shanghai, every App Store subscription billed in São Paulo, every AppleCare plan renewed in Cleveland, a sliver of each transaction accrued to your shares.

Here’s where people get tripped up: nobody mailed you a check for $746. So is the claim real?

Yes, its absolutely real, and the confusion comes from assuming a claim on earnings means a cash payment. What the claim means is the profit either gets sent to you directly or it stays inside a business you own a portion of which should make your stake more valuable. A dollar of retained earnings reinvested at a high return on capital is often worth more to you than a dollar paid out. Depending on what type of account your holding the shares in you would owe taxes on the dividend and no matter what you would have to find somewhere else to deploy the capital being returned. This is the entire reason stocks have value in the first place. A share is worth something for one reason, it entitles the holder to a stream of current and future earnings. If we get rid of that component a stock certificate is just paper.

Where Your $746 Actually Went

Management has four options for every dollar of profit they generate. They can elect to reinvest it in the business through R&D and capital spending, acquire other companies, buy back stock, or pay dividends. Apple does all four and the capital allocation mix tells you a lot about how they prioritize their capital allocation plan.

Apple currently pays a dividend of $1.08 per share which is a payout ratio of 12.58% on their earnings. The dividend has an average growth rate of 4.69% over the past 5-years and they have raised the dividend for 13 consecutive years.

The bigger story is the buyback as Apple repurchased about $90.7 billion of its own stock in the 2025 fiscal year. This is where the ownership math gets fun because every share Apple retires makes your slice of the pie bigger without you spending another dime. The long-term numbers are staggering. Apple’s share count has fallen from roughly 25.2 billion in fiscal 2013 to about 14.7 billion today which is a reduction of around 42%. If you bought 100 shares back then and never added another share to your position, your percentage claim on Apple’s revenue, earnings, and assets is roughly 71% larger today.

Since the 2016 fiscal year Apple’s revenue per share has increased from $9.85 to $30.53 while its earnings has increased from $2.09 to $8.25. This is what I mean by the earnings belong to shareholders even when they aren’t distributed. Apple took profits you had a claim on and used them to permanently increase the size of your claim.

The Price You Pay Sets the Terms of the Deal

Understanding what you own is half the equation the other half is what you pay for it since buying a stock is really buying a stream of earnings at a price. Apple currently trades for $307.34 per share and has generated $8.25 in earnings over the trailing twelve months. Apple trades at a current valuation of 37.25 times earnings. When you flip that multiple upside down you get the earnings yield which is about 2.68%.This means that for every $100 you invest in Apple today it buys you a claim on roughly $2.68 of their current annualized profit.

Now the question becomes if that is a good deal or not. This depends on what happens to the earnings and your perspective on valuing a business. If your paying 37.25 times for a business that compounds profits at 15% a year for a decade it may be a great deal but if the earnings were to remain stagnate you basically purchased an expensive bond with extra steps. The multiple is the market telling you how much future growth is already baked into the price. Your job as an owner is deciding whether that assumption is reasonable.

This framing also separates price from value, and the past year is a perfect case study. Apple’s 52-week range runs from about $195 to over $316.94. The stock was roughly 60% more expensive at the high than at the low. Did your claim on the business change 60% in twelve months? Of course it didn’t but Apple’s revenue did increase 16.6% year over year in the March quarter. Apple’s underlying. While the revenue grew by double digits the underlying earnings stream didn’t swing anywhere near as violently as the price of that stream. The business and the quote are two different things and the quote is just what other people are currently willing to pay for your claim to Apple’s revenue and earnings.

Why the Ownership Mindset Changes Everything

When you internalize that a stock is a claim on earnings your behavior changes. A 10% drawdown stops feeling like losing money and starts feeling like the market repricing the exact same claim you had on a company’s business last month. You owned 0.0000007% of Apple before the dip and you own 0.0000007% of Apple after it. The revenue per share didn’t fall 10% on a Tuesday afternoon, what declined was what people were willing to pay for that revenue stream, which is reflected in the share price. You also start asking better questions. Instead of “where’s the next support level,” you ask “is this earnings stream growing, and what am I paying per dollar of it?” Instead of chasing whatever moved 8% today you look at the income statement and figure out whether the business behind the ticker is earning more money this year than last year.

Over short stretches stocks trade on sentiment, flows, and headlines. Over long stretches they follow earnings per share with remarkable consistency. The price eventually goes where the profits go. None of this means valuation is easy or that great businesses can’t be terrible stocks at the wrong price. Apple at 38 times earnings is a very different proposition than Apple at the mid-20s multiple it averaged over the past decade. The point is that this framework forces you to anchor every decision to the only thing a share ultimately represents which is a fractional claim on a company’s revenue, earnings, and assets.

The Bottom Line

A stock isn’t a ticker, a chart, or a vibe, It’s a deed. When you buy a share, you’re buying a fixed fraction of everything a business sells and everything it earns this year and every year you hold it. The revenue and earnings are partly yours while the decisions about reinvesting it, buying back stock, or paying it out are made on your behalf and they compound in your favor when management allocates capital well. The next time you’re considering making a purchase skip the chart for a second and ask the owner’s questions. How much revenue and profit does this share entitle me to? Is that stream growing? What am I paying for each dollar of it? Get comfortable answering those three questions and you’ve separated yourself from most of the people trading the same ticker.

When your investing rather than speculating your not buying a stock, your buying a small ownership in a business one tiny slice at a time.

Full disclosure: I am a shareholder in Apple (AAPL) which I used as the example in this article

Disclaimer

This article is for informational and educational purposes only and reflects the personal opinions and analysis of Steven Fiorillo as of the date of publication. Nothing in this letter constitutes investment advice, financial advice, trading advice, tax advice, legal advice, or a recommendation to buy, sell, or hold any security, commodity, or other financial instrument. Readers should not treat any of the content as such.

The author is not a registered investment advisor, broker-dealer, or financial planner. The analysis presented here is based on publicly available information, including public company filings, earnings releases, third-party research, and other sources believed to be reliable at the time of writing. However, no representation or warranty, express or implied, is made as to the accuracy, completeness, timeliness, or reliability of the information presented. Forecasts, projections, and forward-looking statements are inherently uncertain and subject to revision. Actual results may differ materially from any projections discussed.

Past performance is not indicative of future results. Any investment in securities discussed or referenced in this letter involves risk of loss, including the possible loss of principal. Readers should perform their own independent research and consult with a licensed financial advisor, tax professional, and legal counsel before making any investment decision.

Barron’s put it on the cover, which tells you the “is this a bubble” conversation has officially left the group chat. The framing is that the chip rally has gone parabolic and it’s time to separate the real pillars from the pretenders, and they’re not wrong that it’s stretched. The information technology sector is up roughly 30% on the year, the Nasdaq just booked its best two-month stretch in decades, and names like Dell and IBM are putting up single-session moves of 13% to 33% that we usually only see around earnings.

• IT sector: up about 30% year to date.

• Nasdaq: best two-month run in decades, up 8% in May alone.

• This week’s outliers: IBM about +13%, Dell about +33%, both in a single session.

There are several takes ranging from it’s all a bubble to this time is different. My take is that the answer lives in the cash flows. Many of these names are printing real free cash flow and raising guidance, Dell just lifted its guidance for the fiscal year. There may be some that are trading on a story and a multiple but there is a big difference between these two groups and viewing the entire narrative the same is just lazy.

2. The “RAMpocalypse” is the realest story almost nobody’s pricing

Of everything here, this is the one I think the market is underpricing. The memory shortage isn’t a forecast anymore as it’s showing up in prices right now. Samsung, SK Hynix, and Micron make more than 90% of the world’s DRAM, and they’ve been redirecting capacity toward high-bandwidth memory (HBM). HBM is specialized memory that wraps around Nvidia’s AI chips. The reason is simple math as an HBM3E module sells for roughly $60 to $100 versus about $5 to $10 for a comparable slice of conventional DDR5. Micron has talked about a 3-to-1 trade-off, meaning every bit of HBM it makes costs it three bits of normal memory. So the AI build-out is literally cannibalizing the memory that goes into phones, laptops, and data-center SSDs.

• There is real supply concentration as Samsung, SK Hynix, and Micron control over 90% of DRAM.

• HBM3E runs about $60 to $100 per module vs roughly $5 to $10 for comparable DDR5.

• DRAM contract prices spiked about 90% in Q1 2026 versus the prior quarter, with some high-performance DDR5 up over 110%, and TrendForce sees another 50% or so this quarter.

I have two initial thoughts on this. This is a clean tailwind for the memory makers for as long as it lasts which is exactly why that side of the chip trade has been on fire. This is also a real-economy cost that flows straight into PC and phone prices, and eventually into the inflation data. The setup is great until it’s a cycle top, and memory is the most cyclical corner of tech there is. That tension is the whole story.

3. Jamie Dimon called the market “exuberant,” but with an asterisk

At the Reagan National Economic Forum on Friday, Jamie Dimon described this market as exuberant, then did something he hasn’t always done lately: he said that’s “not bad.” That’s a real shift in tone. For most of this year he’s been the cautious voice, warning about too much exuberance, frothy AI and Big Tech valuations, and took a page out of Greenspan’s book describing “irrational exuberance.” On Friday he stopped short of calling valuations unreasonable, framing the froth as defensible when it’s underpinned by genuine earnings growth rather than pure speculation. For context, April CPI ran at 3.8%, a touch hotter than expected, so the inflation worry hasn’t gone anywhere.

I believe that when the most-quoted bank CEO in the world moves from “be careful” to “elevated but reasonable,” it’s a sentiment data point worth paying attention to. It doesn’t mean he’s right but it could mean that the bar for what counts as too expensive just got nudged higher by someone the whole market listens to.

4. Perpetual futures, crypto’s biggest product, are finally coming to the US

This one slipped under the radar on a busy Friday, and it’s a genuinely big structural change. The CFTC cleared the way for perpetual futures in the US for the first time. Two approvals landed: Kalshi got the green light to launch a perpetual futures contract tied to bitcoin, the first time a perpetual has gotten the agency’s blessing on US soil, and Kalshi said it’ll seek approval for more than a dozen other crypto contracts. Separately, the CFTC cleared Coinbase to give US clients access to crypto options and perpetual trading on Deribit, the offshore (Dubai-regulated) derivatives venue Coinbase owns that already does billions of dollars a day in volume.

If you’ve never traded one, a perpetual is a futures contract with no expiration date and a funding mechanism that keeps it tethered to spot, and it lets traders put heavy leverage on everything from bitcoin to meme coins. It’s the single most-traded product in global crypto, and until now it lived almost entirely offshore.

I believe there are two large reasons this matters. First, it’s a real revenue and volume catalyst for Coinbase and a credibility milestone for Kalshi as a derivatives exchange. Second, it’s another brick in the “everything market” wall, the same trend behind Robinhood’s agentic tools, where 24/7, highly leveraged access to any asset keeps creeping toward the mainstream retail account. This could become a major shift in how people trade no matter where you stand on crypto itself.

5. The IPO window just blew wide open, and SpaceX is the whale

Reports this weekend say SpaceX has taken a concrete step toward going public, and it’s the marquee name in what’s shaping up to be a summer that is defined by the IPO market. The SpaceX numbers are all over the map with the most recent valuations coming in north of $1.75 trillion. The xAI merger from earlier this year impacts the financials and the valuation of SpaceX. Starlink is now past 10 million subscribers and doing the revenue heavy lifting and going public as soon as this summer would likely make SpaceX the largest IPO in history by deal size.

• The IPO targets for SpaceX that are being floated have reached $2T and we’re going to find out in the coming weeks what the actual number is.

• Goldman is projecting that the 2026 US IPO proceeds could top $160 billion which would beat the 2021 record of about $156 billion.

• Cerebras debuted near a $70 billion valuation and popped about 68% on day one, Fervo Energy set a record for the largest renewable IPO, and OpenAI and Anthropic are both reportedly weighing 2026 listings.

Traditionally a wide-open IPO window is usually a risk-on signal and a read on liquidity and animal spirits. The other side is that mega-IPOs soak up a lot of capital and it needs to come from somewhere. The combination of SpaceX followed by OpenAI and Anthropic could cause capital to be pulled from all areas of the market which matters for all investors because its next to impossible to handicap how this will impact the stock market.

6. A bill in Congress could push Mercedes-Benz out of the US

This is a quieter story with real read-through for the US-China auto fight. A House bill could, depending on how it’s read, bar Mercedes-Benz from the US market over its Chinese ownership. The bill is the Motor Vehicle Modernization Act of 2026, from House Energy and Commerce Chair Brett Guthrie. It would prohibit any automaker with direct or indirect equity held by a foreign-adversary government (China, Russia, North Korea) from importing, selling, or building cars for the US market, for a five-year window. Mercedes gets caught because its two largest individual shareholders are Chinese.

• BAIC (Chinese state-owned): about 9.98%.

• Li Shufu / Geely (via Tenaciou3 Prospect Investment): about 9.69%.

• Combined: about 19.67%, above the bill’s 15% threshold.

• The kicker in my opinion is that the usual carve-out for companies that have built cars in the US for years explicitly doesn’t apply when there’s foreign-adversary ownership. The bill cleared a House committee only

My opinion is that these bills get rewritten constantly and anything can happen. It seems like Washington is moving to wall off the US auto market from Chinese ownership and Chinese-connected-car tech. That’s a multi-year theme touching autos, supply chains, and the broader decoupling trade, and Mercedes getting swept up shows how blunt the early drafts are.

Disclaimer

This article is for informational and educational purposes only and reflects the personal opinions and analysis of Steven Fiorillo as of the date of publication. Nothing in this letter constitutes investment advice, financial advice, trading advice, tax advice, legal advice, or a recommendation to buy, sell, or hold any security, commodity, or other financial instrument. Readers should not treat any of the content as such.

The author is not a registered investment advisor, broker-dealer, or financial planner. The analysis presented here is based on publicly available information, including public company filings, earnings releases, third-party research, and other sources believed to be reliable at the time of writing. However, no representation or warranty, express or implied, is made as to the accuracy, completeness, timeliness, or reliability of the information presented. Forecasts, projections, and forward-looking statements are inherently uncertain and subject to revision. Actual results may differ materially from any projections discussed.

Past performance is not indicative of future results. Any investment in securities discussed or referenced in this letter involves risk of loss, including the possible loss of principal. Readers should perform their own independent research and consult with a licensed financial advisor, tax professional, and legal counsel before making any investment decision.

On Wednesday after the bell Jensen Huang printed a number bigger than anything I had modeled and then he said something that I want everyone to think about:

“The buildout of AI factories — the largest infrastructure expansion in human history — is accelerating at extraordinary speed. Agentic AI has arrived, doing productive work, generating real value and scaling rapidly across companies and industries.”

That isn’t marketing language as its the the CEO of a $5 trillion company telling you in plain English that the cycle isn’t peaking, it’s expanding. I’m going to walk everyone through what Nvidia just confirmed which ties back to every leg of the thesis I’ve been writing about for the past six weeks. I’ll go through the print itself, the new customer mix, the agentic workload shift that drove most of this, the capital return story that nobody read correctly, and where I think the wallet goes next.

The Number That Settles It

Nvidia just delivered $81.6 billion in revenue during Q1 which was up 85% YoY and up 20% sequentially. To put that in context $81.6 billion is roughly 1.5x what Pfizer does in an entire year and Nvidia produced it in 90 days. Jensen came out and guided for $91 billion of revenue in Q2 plus or minus 2% and management explicitly stated this assumes zero China data center compute revenue. That’s roughly $4.2 billion above the published consensus of $86.8 billion. I thought I waws being bullish when I put out my Nvidia Q1 earnings prediction video on YouTube and projected that the Q2 guide would be $88-$90 billion (see below).

What was even more telling is what happened to FY27 estimates after this print. The sell-side was at roughly $310 to $320 billion in FY27 revenue heading into the call. With $81.6 billion already in the books and a Q2 guide at $91 billion, you’re looking at the first half of the year delivering $172.6 billion in revenue. If the second half is even flat sequentially, FY27 lands above $355 billion. If revenue grows in the back half of 2026 as Rubin ramps then NVDA’s revenue can exceed $375 billion. I argued in “Chips Gone Wild” that the buyers weren’t slowing down and that they were increasing their spend with Nvidia. The Q1 print on its own would have confirmed it but the Q2 guide ends the debate.

The Customer Base Just Got Wider, Not Narrower

Here’s the data point nobody is talking about and it might be the most important takeaway of the entire print. Nvidia changed how they report Data Center revenue this quarter. Instead of “Compute” and “Networking,” they now break it into “Hyperscale” (the big public clouds and consumer internet giants) and “ACIE” (AI Clouds, Industrial, and Enterprise, which captures sovereign AI projects, pharma, automotive, defense, and government).

The split came in at almost exactly 50/50:

• Hyperscale: $37.9 billion, up 115% YoY

• ACIE: $37.4 billion, up 74% YoY

Every bear case on Nvidia for the past two years has been some version of “the entire business depends on four hyperscalers and when their spend slows Nvidia will collapse.” The data just told them they’re wrong as half of Nvidia’s Data Center revenue is now coming from buyers other than the big public clouds. Nvidia has become critical across the value chain and the sovereign AI piece is bigger than most people realize. Saudi Arabia’s HUMAIN which is a subsidiary of the Public Investment Fund has committed over $100 billion to build 11 data centers with a combined 2,200 megawatts of capacity. It will be powered by several hundred thousand Nvidia GPUs. The UAE’s Stargate UAE project with G42 launched with roughly 100,000 chips in initial deployment and is scaling from there. The UK is deploying up to 120,000 Blackwell Ultra GPUs across the country by the end of 2026 which includes the Loughton supercomputer that ties 24,000 GB300 chips into Microsoft Azure. France signed for a 1 gigawatt phase one buildout while Germany is deploying tens of thousands of GPUs for AI factories focused on automotive and robotics.

This is the dynamic Jensen has been describing for two years and the market has been slow to price it in. From the earnings call it doesn’t seem like Sovereign AI is a 2028 story anymore but rather a story for today. Just looking at Saudi Arabia, if HUMAIN executes on the plan it would generate tens of billions in revenue for Nvidia over the next three years. That’s exactly the diversification story that I have been arguing was forming back in April when I wrote about the $1.66 trillion in combined backlog and RPO across the megacaps. The demand wasn’t concentrated into a few companies as it was spreading and the ACIE line item just put a number on it.

The Networking Tell

Inside the Data Center number was networking which grew 199% YoY to $14.8 billion. This is the single most important data point nobody is highlighting as networking revenue grows when cluster sizes grow. You don’t buy more InfiniBand and Spectrum-X Ethernet unless you’re connecting more GPUs into bigger training and inference fabrics. When I see 199% growth in networking it tells me that the physical signature of AI factories being assembled at a scale we have never seen. I believe that the networking moat is the part of the Nvidia story that’s hardest to displace with custom silicon.

The press release discussed three optical partnerships I want to flag because none of these are in consensus models yet. Nvidia has multi year strategic agreements with Coherent, Corning, and Lumentum on advanced optics for the Vera Rubin platform. There is also a partnership with Marvell through NVLink Fusion that lets custom silicon plug directly into Nvidia’s rack-scale fabric which means even when a hyperscaler runs Trainium or TPU silicon that the network around it is often still Nvidia’s. That’s a moat inside a moat that nobody is really discussing from what I am seeing. If you wanted to know whether the data centers being built are getting bigger or smaller, the networking line settled it.

Agentic AI Is What Actually Changed

Jensen used the word “agentic” 30+ times across the press release, CFO commentary, and the earnings call. That isn’t a buzzword choice but rather the explanation for why demand surged this quarter and why the Q2 guide is what it is. The underlying dynamic through 2024 and most of 2025 was that AI compute demand was driven by generative model training and one-shot inference. You prompt the model, it answers, and the cycle ends. That workload is real but it’s relatively predictable and CPU-light. The new wave is different as Agentic AI runs autonomous workflows.

The model plans, calls tools, and queries databases write code, evaluates results, retries when steps fail, and operates over multi-step processes without human intervention. This sequence is brutally more compute-intensive as a single agentic task can spawn thousands of inference calls. Coordination between agents requires orders of magnitude more memory bandwidth and interconnect throughput than a single prompt-response cycle. Traditional x86 CPUs choke on this orchestration which is why Nvidia announced the Vera CPU this quarter. This will be Nvidia’s first processor purpose-built specifically for agentic AI orchestration. Nvidia has indicated that when the Vera CPU is paired with Blackwell Ultra and the upcoming Rubin generation that it’s the architecture that runs the agentic factory.

The two things I am taking away from what Jensen has discussed at all the conferences and the earnings call is that the demand curve isn’t plateauing because we’ve already built enough capacity for generative AI. The demand curve is re-accelerating because agentic AI needs roughly an order of magnitude more compute per query than the workloads we’ve been provisioning for. The next thing is that this is likely why Nvidia is committing to capital return now as they have so much visibility into the next two to three years of demand that they’re comfortable returning $20 billion in a single quarter to shareholders. You don’t do that if you’re worried about the cycle.

The Bubble Argument Is Officially Dead

Let me put real numbers on the table about what this business generates.

• Operating income: $53.5 billion in 90 days, up 147% YoY

• Net income: $58.3 billion, up 211% YoY

• Free cash flow: $48.5 billion in a single quarter

• GAAP gross margin: 74.9%

• Cash returned to shareholders this quarter: $20.0 billion

• New buyback authorization: $80.0 billion, with no expiration

• Dividend: raised 25x, from $0.01 to $0.25 per quarter

When I wrote the bubble-debunk piece in April the core argument was simple as the Dot-com companies funded their buildouts with debt and equity issuance because they had no earnings. Today’s AI infrastructure cycle is funded out of operating cash flow at companies with 75% gross margins. Nvidia’s print is the receipt as they generated $50.3 billion in operating cash flow in 90 days. The hyperscalers that buy from Nvidia generated over $580 billion in operating cash flow on a trailing twelve months basis between them. There’s no debt-fueled speculation here as this is the most profitable companies in history funding the most expensive infrastructure cycle in history out of pocket.

Supply Is the Constraint, Not Demand

Buried in the CFO commentary is one line that I am not seeing discussed and its critical to the AI revolution: “We have strategically secured inventory and capacity to meet demand beyond the next several quarters.” Total supply-related commitments now sit at $119.0 billion while inventory rose from $21.4 billion to $25.8 billion sequentially. Multi-year cloud service commitments grew from $27 billion to $30 billion. Nvidia isn’t building inventory because demand is rolling over but rather its building it because they still can’t keep up. Nvidia was able to put $18.6 billion into non-marketable securities this quarter alone. That’s Nvidia investing directly into customers, AI infrastructure startups, and ecosystem partners. Their total non-marketable securities now sit at $43.4 billion on the balance sheet which is nearly double the level from three months ago. Nvidia isn’t just selling chips, their investing in the customers who buy them which deepens the lock-in and accelerates the deployment cycle.

The narrative going into earnings was “what if demand peaks?” The narrative coming out is “what if supply can’t catch up to where demand actually is?” That’s a fundamentally different story and it has the same shape as what I argued in “Chips Gone Wild” about the memory makers, AMD’s MI350 backlog, and the SanDisk data center print. The whole chip stack is supply-bound, not demand-bound.

The Capital Return Pivot Is a Phase Transition

The dividend hike is the most underappreciated part of Nvidia’s print because for the past three years the investment thesis on NVDA has been hypergrowth. You owned it because revenue was doubling, total addressable market estimates were being revised upward every quarter, and the multiple was justifiable on growth alone. That phase is likely getting old as nobody is bumping their TAM estimates by 50% every report. What management is signaling with this capital return policy is the next phase of the thesis. They are telling you cash generation is durable enough to commit to a permanent capital return program, not just opportunistic buybacks. They are telling you the operating profile has matured to the point where dividend growth mandates can finally initiate a position. Management is also signaling that they think the stock is undervalued because they would rather retire shares than sit on the cash.

A company with 75% gross margins, 65% operating margins, and $48.5 billion of quarterly free cash flow is not raising the dividend 25x because they ran out of better ideas. They are doing it because they want to expand the shareholder base. This is the same phase transition Apple went through in 2012 when they initiated their dividend, and Microsoft in 2003. In both cases, the stock didn’t outperform immediately but the buyer base permanently widened, the volatility decreased, and the multiple compressed less than the bears expected. Nvidia is now eligible for the same treatment as the capital return policy is the structural floor underneath the next leg of the thesis. It’s not a fireworks-level catalyst on day one but it changes the underwriting math for institutional capital over the next 12 to 18 months.

Where the $700 Billion Capex Wave Is Going

Combined 2026 capex from the five biggest builders is tracking past $700 billion as Amazon is at $200 billion while Alphabet is expected to spend around $190 billion. Meta forecasted between $125 to $145 billion and Microsoft is trending toward $120 billion. Oracle rounds out the group with around $50 billion being allocated towards CapEx. This is an increase of 3.5x increase in two years across the most profitable companies on earth. The question heading into Nvidia’s print was simple, is that capex actually landing in Nvidia’s revenue line or is it being diverted to custom silicon? The answer the print gave was that it’s absolutely landing in Nvidia’s coffers. Hyperscale revenue at Nvidia grew 115% YoY which means Nvidia’s share of the wallet is going up, not down.

The custom silicon story (Trainium 3, TPU v7 Ironwood, MTIA v3, Maia 200) is real for inference workloads, and I won’t pretend it isn’t. When it comes to training Nvidia confirmed its still capturing the lion’s share with margins being high and switching costs remaining very sticky. The networking print tells us the rest as inference at scale still runs through Nvidia’s fabric even when the chip itself is custom. The NVLink Fusion partnership with Marvell I mentioned earlier is the structural answer to that threat.

When this is layered on top of the hyperscaler capex the trillion-plus dollars in compute commitments that Anthropic and OpenAI have signed with Nvidia, Oracle, Microsoft, Amazon, Google, and AMD over the next five to seven years they opportunity expands. Then layer on top of that the sovereign AI buildouts I described earlier. We also are seeing another vertical with Meta’s next training cluster, Tesla’s Dojo successor, Apple’s rumored data center buildout, and the enterprise migration that hasn’t even ramped yet. The total addressable market for AI compute over the next five years is north of $4 trillion by Jensen’s own framing, and at this point, that estimate looks conservative.

Where the Buildout Goes From Here

I closed “Chips Gone Wild” by saying the second-derivative trades become interesting from here including power generation, transformers, cooling, and fiber. The picks-and-shovels names that benefit from the buildout’s duration rather than its peak and Nvidia’s print sharpens that thesis. Nvidia has the chips ready and TSMC is expanding CoWoS advanced packaging capacity to 150,000 wafers per month by the end of 2026. The bottleneck is now electrons, substations, liquid cooling, fiber, switchgear, and transformers. The companies that supply those likely have a multi-year tailwind that doesn’t depend on guessing which chip wins the inference market in 2028.

Some examples for instance are GE Vernova as the pure play on grid infrastructure and gas turbines. Every AI factory needs power, and a meaningful percentage of new data center capacity is now being sited near gas peaker plants and small modular reactor pilot zones. GEV makes the equipment that makes that work. Eaton is the electrical components story which includes switchgear, transformers, and substation equipment. The backlog on Eaton’s electrical segment has been growing double digits for six quarters straight and the AI data center buildout is the single biggest contributor. Vistra and Constellation are the regulated and merchant nuclear/gas generation companies that are benefiting from the buildout. Constellation locked in a 20-year power agreement with Microsoft last year to restart Three Mile Island. Vistra has done similar long-term deals with hyperscalers.

Bottom Line

If you read “Chips Gone Wild” two weeks ago and wondered whether I was being too aggressive Nvidia’s Q1 results was the answer as I wasn’t bullish enough. The customer base is wider than the bears claim as Sovereign AI is a real revenue stream now and the ACIE line is the proof. The networking number tells you the clusters are getting bigger not smaller, and the silicon photonics partnerships extend the moat. The supply commitments indicate that demand is still ahead of supply, and the strategic investments tell you Nvidia is funding the demand it can’t yet fulfill. The capital return policy indicates that management knows what’s coming over the next two to three years and they’re positioning the equity for the next phase of the buyer base. ACIE growing 74% YoY tells you sovereign AI, enterprise, and industrial demand isn’t a slide deck, its real revenue expanding the metrics. The AI infrastructure cycle isn’t in its peak innings and I believe its at the start of the middle of the cycle. The agentic workload is just beginning to ramp and the receipts keep getting bigger. For now, the chips aren’t just wild, they’re proving the thesis in real time.

Disclaimer

This article is for informational and educational purposes only and reflects the personal opinions and analysis of Steven Fiorillo as of the date of publication. Nothing in this letter constitutes investment advice, financial advice, trading advice, tax advice, legal advice, or a recommendation to buy, sell, or hold any security, commodity, or other financial instrument. Readers should not treat any of the content as such.

The author is not a registered investment advisor, broker-dealer, or financial planner. The analysis presented here is based on publicly available information, including public company filings, earnings releases, third-party research, and other sources believed to be reliable at the time of writing. However, no representation or warranty, express or implied, is made as to the accuracy, completeness, timeliness, or reliability of the information presented. Forecasts, projections, and forward-looking statements are inherently uncertain and subject to revision. Actual results may differ materially from any projections discussed.

Past performance is not indicative of future results. Any investment in securities discussed or referenced in this letter involves risk of loss, including the possible loss of principal. Readers should perform their own independent research and consult with a licensed financial advisor, tax professional, and legal counsel before making any investment decision.

Full Disclosure: I am a shareholder in Nvidia, Amazon, Alphabet, Meta Platforms, Tesla, Apple, and Microsoft

Here’s the headline change in one table the EIA put right at the front of the report:

This is basically the entire energy thesis getting rewritten in a single month.

What Actually Happened

The Strait of Hormuz has been effectively closed to shipping traffic since February 28^th. Roughly 20% of global oil supply moves through the Strait and the EIA is now estimating that Iraq, Saudi Arabia, Kuwait, the UAE, Qatar, and Bahrain collectively shut in 7.5 million barrels per day (bpd) of crude production in March. The agency expects shut-ins to peak at 9.1 mb/d in April before slowly recovering through the back half of the year. To outline how detrimental this is I believe its important to look at how much oil is actually produced by nation and how many nations actually exceed 1 million bpd.

The base-case assumption in the report is that the conflict doesn’t extend past April and that traffic through the Strait gradually resumes. Even under that assumption he EIA doesn’t expect a return to pre-conflict production levels until late 2026 The EIA is building a risk premium into prices in their base case through 2027 because backlog and trade-flow disruption take time to unwind.

There are several aspects of the energy market I want to flag before going forward which include:

• Brent averaged $103/b in March and touched almost $128/b on April 2. The EIA’s monthly average forecast peaks at $115/b in Q2 2026 before easing.

• The Brent-WTI spread blew out to $12/b in March (from $6/b in February) and the EIA expects it to peak at $15/b in April before grinding back to $4/b by Q4.

• An attack on Qatar’s Ras Laffan LNG export facility on March 18 damaged two liquefaction trains representing 17% of Qatari export capacity. QatarEnergy estimates repairs take up to five years. That’s not a temporary disruption as its structural and can take an extended period to fix.

• U.S. crude oil production is forecast at 13.5 mb/d in 2026 and 13.8 mb/d in 2027. The U.S. announced an SPR release on March 11 and a 60-day Jones Act waiver on March 17.

The setup matters for several reasons. Obviously it’s the middle of May and this base case from the EIA assumed that the conflict would end before April concluded. This didn’t happen and energy prices are still hovering around $100 per barrel.

Upstream Exploration & Production (E&P)

The E&P segment of the oil patch is the cleanest beneficiary as U.S. producers are selling barrels into a market where WTI averaged $72.74/b in Q1 2026 and the EIA now forecasts $101.63/b for Q2. The full-year 2026 WTI forecast is $87.41 per barrel which increased 34% from the 2025 average of $65.50 per barrel. Pure-play Permian and Eagle Ford operators get the most leverage because their cost structure is fixed and every incremental dollar above breakeven drops to the bottom line. The asterisk is OPEC+ surplus capacity because the EIA was forecasting OPEC+ production of 43.9 mb/d in 2026 which has now been cut to 41.7 mb/d. When the Strait reopens and those barrels come back the same producers that just got a windfall will be facing a supply wall heading into 2027.

The EIA’s own 2027 Brent forecast of $76 per barrel tells me that they expect this to normalize at a higher than they previously projected for. The upstream trade is asymmetric in a specific way in my opinion where the next two to three quarters are a clear earnings tailwind but it will be harder to maintain that earnings power past that timeline.

Refiners: Diesel Is Doing The Heavy Lifting

If I was only going to glance through the STEO I would have focused on the refiners because the EIA is forecasting retail diesel at $4.80 per gallon in 2026 which is up from $3.66 in 2025. This is an increase of 31% and the wholesale diesel forecast sits at $3.40 per gallon compared to $2.32 in 2025. In the EIA’s model wholesale diesel hits $4.25 per gallon in Q2 2026. This is the asymmetric refining trade as gas is being forecast to hit $3.70 per gallon in 2026 compared to $3.10 in 2025 which is an increase of 19%. Diesel is up considerably more because distillate inventories in the U.S. are below the five-year average and Europe banned seaborne Russian diesel imports back in 2023 which is a major impact because both Europe and Asia rely on Middle East diesel flows that just got disrupted. The EIA is forecasting that diesel refining margins decline in the back half of 2026 but that they will remain well above the 2025 levels due to continued global supply tightness for diesel fuel.

Natural Gas and LNG: The Story I’m Most Interested In

This is where the EIA report gets the most interesting in my opinion and where I think the medium term thesis is the cleanest. I will be using the term LNG a lot so I should outline what it is. LNG is liquefied natural gas which is natural gas that has been cooled to liquid form making it easier to transport globally. LNG takes up approximately 1/600th the volume of natural gas.

U.S. LNG exports hit 17.9 Bcf/d in March 2026 which was the second-highest monthly export volume on record. The EIA is forecasting that the 2026 LNG export volume will be 17.0 Bcf/d and the 2027 export volume will come in at 18.6 Bcf/d. The previous record was set in 2025 with 15.1 Bcf/d. The spread is what makes this so interesting.

• Henry Hub to TTF (Europe) spread: $14.89/MMBtu in March, up 83% from February.

• Henry Hub to JKM (Asia) spread: $15.23/MMBtu, up 98% from February.

Henry Hub itself is forecast at $3.67/MMBtu in 2026 which means that U.S. LNG is being sold into international markets at multiples of the domestic price. This spread is likely to compress when the Strait reopens but the Qatar Ras Laffan damage doesn’t go away as the repair timeline is expected to be roughly 5-years. This is critical because these facilities represented 17% of Qatari export capacity and Qatar exported 20% of global LNG in 2025. In my opinion this means that we will see a structurally tighter global LNG market for the rest of this decade which could benefit U.S. LNG exporters.

There is also additional capacity coming online which is also an aspect to factor in within the LNG segment. Corpus Christi Stage 3 (Train 5) reached substantial completion in March. The Golden Pass Train 1 is set to begin exports in Q2 2026 while the Plaquemines just got DOE approval for an additional 0.5 Bcf/d of export authorization to non-FTA countries. That’s roughly 0.9 Bcf/d of incremental capacity coming online in Q2 alone which flows into a market that has been constrained due to the conflict overseas.

These are the areas that I see benefiting from the current environment:

• The LNG exporters themselves. Their contract structures lock in tolling fees regardless of the spread

• Upstream natural gas producers. The EIA expects marketed natural gas production to grow 2% in 2026 and 3% in 2027 with a substantial amount of that incremental growth being fed into the export terminals rather than domestic consumption

• Midstream operators. Pipelines moving feed gas from the basins to the Gulf Coast LNG facilities should benefit as their volumes on the take or pay contracts increase.

The Data Center Build Is The Other Half Of The Natural Gas Story

I want to spend real time on this because I don’t think the EIA report fully captures what’s happening here, and I think it’s the single most important structural shift in U.S. energy demand of the next five years. The Q1 2026 hyperscaler earnings prints were released at the end of April and the numbers were dramatically higher than anything they had previously guided for. Microsoft, Alphabet, Amazon, and Meta now collectively plan to spend roughly $715 billion on capex in 2026 which is up from $399 billion in 2025 and $230 billion in 2024. That’s a 79% YoY increase on top of a 73% increase the prior year.

That trajectory deserves a closer look at the company level because the spending isn’t evenly distributed and the YoY growth varies.

In 2026 the hyperscalers have gone past what some would consider all-in on the AI buildout for their CapEx allocations:

• Amazon: $200 billion

• Microsoft: $190 billion (CFO Amy Hood attributed $25 billion of that to memory chip and component cost inflation)

• Alphabet: $185 to $190 billion

• Meta: $125 to $145 billion (raised from prior range, citing memory pricing and data center cost inflation)

The bulk of this capex is data center infrastructure which is bullish for energy because data centers are physical buildings that need physical power which is where the EIA report and the hyperscaler story collide. The translation from CapEx dollars to electricity demand isn’t subtle as the EIA is forecasting 5.8% commercial sector electricity demand growth by the summer of 2027. The commercial sector category the EIA is accounting for is where most data center load gets placed because data centers are leased commercial real estate.

Here’s where it gets interesting for the energy sector because the hyperscalers can’t actually wait for the grid to catch up. The Bloom Energy data center survey from March 2026 found that time-to-power for utility interconnection now runs roughly 1.5 to 2 years longer than developers had previously expected. Grid interconnection queues in major data center markets exceed four years in some cases. If you’re spending $200 billion to build AI infrastructure you can’t have your GPUs sitting in boxes waiting for the local utility to add a substation.

The solution the industry has converged on is behind-the-meter natural gas generation. There was a Cleanview analysis that showed 46 data centers with a combined capacity of 56 gigawatts are using behind-the-meter generation. To give this some perspective this is the equivalent to roughly 27 Hoover Dams of dedicated generation capacity and most of that equipment is natural gas-powered. The hyperscaler positioning on this is telling as some are explicitly framing natural gas as a bridge fuel while small modular reactors get built. There are some that are pairing natural gas with battery storage in hybrid configurations to manage the ramping limitations of combined-cycle turbines. Blackstone announced a $25 billion Pennsylvania data center and natural gas plant investment in 2025 which included a $1 billion acquisition of the 620-megawatt Hill Top Energy Center. Private equity is putting massive checks behind the thesis that data centers and gas plants get co-located.

What I believe this actually means:

• Natural gas turbine manufacturers see increased demand they can’t keep up with

• Natural gas producers benefit from a structural step-up in domestic gas demand that the EIA report only partially captures

• Midstream operators with capacity in the right basins benefit twice as they feed both LNG export facilities and the data centers coming online

• Independent power producers and natural gas producers see increased capacity value go from a rounding error to a meaningful contributor to earnings

What everyone should keep in mind about what I have outlined

• The hyperscaler CapEx numbers aren’t guaranteed and the guidance can get cut for a number of reasons

• If the AI demand curve disappoints or if the productivity gains don’t show up we could see CapEx guidance get revised

• Memory pricing is a major component of the CapEx increase which offsets the data center construction to a degree

• Both Microsoft and Meta explicitly called out higher memory and component pricing as the driver of guidance increases the prior week when they reported earnings

• The behind-the-meter trend is partly a regulatory arbitrage as some of these projects avoid certain interconnection requirements and approvals because they bypass the utility grid. If that regulatory permissiveness tightens it could cause build timelines to extend

My opinion is that the net effect is that the natural gas demand picture is structurally tightening in a way that the EIA forecast probably understates. Whether that flows through to Henry Hub prices, to midstream cash flows or capacity pricing depends on which constraints bind first but the direction is clear. The AI CapEx cycle is the largest single new demand source for U.S. natural gas in this decade.

Midstream: The Quiet Compounder

Midstream tends to get overlooked when oil prices are moving because the business model isn’t direct commodity exposure but this is the type of environment where the segment shines. Volumes are growing across the board as U.S. crude production is projected to be 13.8 mb/d in 2027 while natural gas production hits 112.6 Bcf/d in 2027. All of those barrels and molecules need to move from their point of origination to the immediate destination and sometimes the final destination. The EIA forecasts U.S. natural gas plant liquids production growing from 6.99 mb/d in 2025 to about 7.59 mb/d by 2027. NGL exports have been a tailwind for fractionation operators for years and that isn’t expected to change in this scenario. The way I’d think about midstream in this environment is that you’re getting paid a fee per molecule for infrastructure that’s increasingly hard to replace in a market where demand for that infrastructure is structurally growing. The capital programs at the master limited partnerships (MLPs) that I follow are increasingly tied to specific LNG and petrochemical demand growth rather than speculative basin development. The midstream sector has tended to grind but in an environment where the EIA is forecasting LNG exports growing from 15.1 Bcf/d in 2025 to 18.6 Bcf/d in 2027 the cash flow trajectory becomes more durable.

Electricity Demand: The Story That Doesn’t Stop

The EIA forecasts total electricity demand at 4,108 BkWh in 2026 which is an increase of 1.2% YoY and 4,244 BkWh in 2027 which is up 3.3% YoY. Residential electricity prices are forecast to grow 5.1% in 2026 and 2.4% in 2027. The IPP and merchant power names with exposure to growing load centers probably benefit the most here because they can repower existing assets, sign data center supply agreements at premium rates, and capture scarcity pricing during peak times. The regulated utility play is a different viewpoint because it becomes a combination of rate base growth from grid investment and the cost recoverable through customer rates. Merchants can provide leverage to scarcity pricing while regulated utilities provide predictable rate base compounding without much commodity exposure.

Renewables: Solar’s Year, Quietly

This is the part of the report that surprised as the EIA is forecasting that U.S. solar generation will grow 17% in 2026 and 22% in 2027. In the summer of 2027 solar generation is expected to reach 178 BkWh which surpasses wind generation by almost 30% for the summer months. That solar number isn’t a function of new policy as its a function of project pipelines that were approved years ago coming online combined with utility-scale solar being the cheapest new capacity addition in most parts of the country. Solar’s share of total U.S. electricity generation is expected to increase to 9% in 2027. I think the cleaner are to investigate isn’t the solar manufacturers considering the import dynamics and tariff backdrop but rather the utilities and IPPs that own the solar capacity and are selling power at prices that keep rising.

My thoughts in a nutshell

I keep coming back to the same observation as the April STEO didn’t just revise some price assumptions. It reframed the entire setup for energy in 2026 and 2027. The cyclical piece is real and the spike in Brent, the diesel crack expansion, and the windfall quarter for upstream operators are happening right now and showing up in the Q1 prints that are starting to cross the tape. That cyclical trade has a built-in expiration date because the EIA’s own base case assumes the Strait of Hormuz reopens and OPEC+ production recovers. The agency is already forecasting Brent at $76 in 2027 which tells you where they think this normalizes. What I find more interesting is the structural piece because parts of the April report do not unwind when the conflict ends. The Qatar Ras Laffan damage is going to take up to five years to repair and that takes 17% of Qatari LNG export capacity off the global market for the rest of this decade. The U.S. LNG export capacity expansion is moving forward whether Brent is at $96 or $76 because those projects were sanctioned years ago and the contracts are tolling-based. The AI CapEx cycle isn’t tied to oil prices at all as the hyperscalers committed to $715 billion of 2026 CapEx before the Hormuz situation started. The demand pull on natural gas, grid infrastructure, and generation is going to keep building whether Brent settles back to $76 or stays above $90.

That’s why I keep coming back to natural gas, LNG, midstream, and the IPPs as the segments where the time horizon makes the most sense to me. The setup isn’t perfect because if hyperscaler CapEx guidance gets revised lower or if behind-the-meter regulatory permissiveness tightens then the thesis will become weaker. The direction for U.S. natural gas demand is structurally higher because of LNG exports and AI infrastructure which are two demand sources that reinforce each other rather than competing for the same molecules. The upstream and refining beneficiaries are real as the midstream operators, the LNG exporters, the gas producers feeding the export terminals, and the IPPs with capacity in the right markets seem to be expanding their durability. The April STEO is the moment where the cash flow trajectory for those segments got rewritten higher and I don’t think the broader market is looking as closely as I am at the setup.

Full disclosure: I am invested in Microsoft, Amazon, Alphabet, and Meta Platforms which are mentioned in this article. I am also invested in energy companies that are not mentioned in this article.

Disclaimer

This article is for informational and educational purposes only and reflects the personal opinions and analysis of Steven Fiorillo as of the date of publication. Nothing in this letter constitutes investment advice, financial advice, trading advice, tax advice, legal advice, or a recommendation to buy, sell, or hold any security, commodity, or other financial instrument. Readers should not treat any of the content as such.

The author is not a registered investment advisor, broker-dealer, or financial planner. The analysis presented here is based on publicly available information, including the U.S. Energy Information Administration’s April 2026 Short-Term Energy Outlook, public company filings, earnings releases, third-party research, and other sources believed to be reliable at the time of writing. However, no representation or warranty, express or implied, is made as to the accuracy, completeness, timeliness, or reliability of the information presented. Forecasts, projections, and forward-looking statements, including those produced by the EIA and referenced in this letter, are inherently uncertain and subject to revision. Actual results may differ materially from any projections discussed.

The energy sector is subject to substantial risks, including but not limited to commodity price volatility, geopolitical events, regulatory changes, supply and demand imbalances, weather events, technological disruption, and macroeconomic conditions. Past performance is not indicative of future results. Any investment in securities discussed or referenced in this letter involves risk of loss, including the possible loss of principal. Readers should perform their own independent research and consult with a licensed financial advisor, tax professional, and legal counsel before making any investment decision.

AMD’s print: this isn’t a one-quarter move

AMD reported their Q1 2026 results on May 5^th and their revenue came in at $10.3 billion which was up 38% YoY. The heart of the story is data centers as AMD’s data center revenue increased 57% YoY to $5.8 billion. That single segment is now larger than AMD’s entire revenue base from a few years ago. On the earnings call Lisa Su (AMD CEO) called it “a clear inflection in our growth trajectory and a structural shift in our business,” and she’s right. AMD’s data center business is now the primary driver of the entire company’s revenue and earnings growth which is an indication of where their clients are spending their CapEx dollars.

What really matters is their Q2 guidance came in at $11.2 billion in revenue at the midpoint which is a 46% YoY growth rate. This is a sequential acceleration on a base that’s already running hot. The Instinct ramp is the rocket fuel with AMD’s MI400 series is coming in the second half of this year. The MI455X has 432GB of HBM4 per GPU and 19.6 TB/s of memory bandwidth which is 50% more memory capacity than the MI350 and 145% more memory bandwidth. Anyone running frontier AI workloads cares about memory throughput right now because that’s where inference workloads bottleneck. AMD just got a memory spec advantage for the first time in this cycle.

Then there’s the customer roster. Inside the Q1 release AMD highlighted Meta’s plan to deploy up to 6 gigawatts of AMD Instinct GPUs with the first 1 GW being a custom MI450-based design. Meta will also be the lead customer for AMD’s 6th gen EPYC. This is in addition to OpenAI’s 6 GW MI450 commitment from October 2025 which carries warrants for OpenAI to buy 160 million AMD shares and represents an estimated $80 to $100 billion in hardware revenue through 2030. Oracle is taking 50,000 MI450 GPUs starting in Q3 2026 while AWS, Google Cloud, Microsoft Azure all announced new EPYC instances during the quarter. AMD has 12 GW of customer-committed Instinct deployments between Meta and OpenAI alone, with the first GW shipping in the second half of this year. That’s not an aspirational pipeline, that’s contracted hardware backlog and it’s happening while Nvidia is still selling everything it can build. If chips were going to slow down, this isn’t the print you’d see.

Memory and storage: the super cycle nobody saw coming

Everyone has been laser focused on GPUs because they’re the headline grabber. The bigger story for the duration of this buildout might actually be memory and storage, because those segments are producing the growth rates GPU’s were previously producing.

Micron reported their Q2 2026 earnings on March 18^th and they produced was $23.86 billion in revenue, up 196% YoY and 75% sequentially from $13.6 billion in Q1 2026. Micron told the Street to expect $33.5 billion in revenue during Q3 with non-GAAP gross margins of approximately 81% and EPS of $19.15. An 81% gross margin in a memory business is mind blowing from a business that is supposed to be the most cyclical commodity business in tech. The reason is it’s high bandwidth memory. Micron has price and volume agreements locked in for its entire calendar 2026 HBM supply including HBM4. They’re shipping HBM4 36GB 12-high parts now which is aligned with Nvidia’s Vera Rubin platform. CEO Sanjay Mehrotra told investors HBM TAM is going from approximately $35 billion in 2025 to roughly $100 billion in 2028 which has been pulled forward by two years from their prior view. Two years pulled forward in a single year of forecasting is how fast the demand curve is moving.

CapEx is the tell and Micron raised fiscal 2026 capex from $18 billion to $20 billion. They broke ground in January on a $100 billion campus in New York with wafer output expected by H2 2028. Idaho is starting initial production by mid-2027 while the Singapore HBM packaging will produce meaningful contributions in 2027. They’re spending money like they have a multi-year demand book locked in and they probably do. Micon isn’t going to ramp their CapEx if the future demand isn’t visible or locked in already. They could allocate that capital towards buybacks but they are telling the market a better use of that capital is coming from expanding their infrastructure.

Next is SanDisk which delivered a showcase quarter with their Q3 2026 results. Their revenue came in at $5.95 billion which was up 251% YoY and 97% sequentially. Data center revenue increased 645% YoY and 233% sequentially as they recorded $1.47 billion in purchases. Their edge revenue which includes smartphones and PCs increased 295% to $3.66 billion. SanDisk guided for $7.75 to $8.25 billion in revenue with 79 to 81% gross margin in their Q4. What’s happening at SanDisk is the missing piece in the AI hardware story. Training data needs to live somewhere while inference workloads need fast SSDs to feed the GPUs. The big three NAND makers cut production a few years ago when prices collapsed and shifted focus to DRAM. Then the AI super cycle hit and you’ve got demand outstripping supply by such a wide margin that NAND prices have gone vertical and customers are signing multi-year contracts to lock in supply. SanDisk’s CEO David Goeckeler disclosed that SanDisk has signed five new long-term contracts with $11 billion-plus in financial guarantees and a $42 billion revenue backlog over the term of those agreements. He raised his calendar 2026 data center sector growth forecast to the mid 70s up from the previous guide of in the 60s.

These aren’t single-quarter pops as Micron and SanDisk are both signing customers to multi-year contracts because the customers want supply locked in. That’s the behavior of buyers who are scared they won’t have enough memory and storage to feed their data centers in 2027 and 2028.

Microsoft RPO: a $627 billion contract book

Now to the cloud. This is where the duration argument gets serious. Microsoft reported their fiscal Q3 2026 numbers on April 29^th. Their revenue came in at $82.9 billion, up 18% YoY. Microsoft Cloud crossed $50 billion in a single quarter for the first time at $54.5 billion which increased 29% YoY. Azure grew 40% YoY which beat their own guidance of 37-38% which ended a deceleration trend that had the market worried. The metric that matters the most in my opinion is the commercial remaining performance obligation. This is the contracted-but-not-yet-recognized revenue and it just reached $627 billion. This is up 99% YoY and roughly 25% of it is expected to convert to revenue in the next 12 months and that 12-month component grew 39% YoY. This is roughly 7.5x Microsoft’s quarterly revenue run rate and it doubled in a year.

Capex is the matching commitment as Microsoft spent $31.9 billion in Q3 alone with two-thirds of it on short-lived assets like GPUs and CPUs. Q4 capex is guided to over $40 billion while the calendar 2026 capex is now expected to be roughly $190 billion, up 61% from 2025, with about $25 billion of that coming from higher component pricing because memory chips have gone vertical. Amy Hood told the Street they expect to remain capacity-constrained at least through 2026. Microsoft has GPUs sitting in inventory because they don’t have the electrical capacity yet to install them. That’s the infrastructure layer that supply doesn’t fix in 90 days. Power, substations, transmission, transformers, cooling. The biggest software company on Earth is being supply-constrained by physical infrastructure. If you wanted evidence the AI buildout has years left, that’s it because Microsoft isn’t spending on the components to warehouse them for years on end. They are likely to take a blank check approach to stand up data centers as quickly as possible to monetize these assets that are not spun up yet.

Amazon’s AWS book: $364 billion and counting

Amazon reported the same day and their Q1 2026 revenue was $181.5 billion which was up 17% YoY. AWS hit $37.6 billion in revenue which was up 28% YoY. This was the fastest AWS growth rate in 15 quarters which is proof that AWS isn’t decelerating but rather reaccelerating. The AWS backlog is now $364 billion and Andy Jassy (Amazon CEO) stated on the call that this figure does not include the $100 billion-plus deal Amazon signed with Anthropic. So call it closer to $464 billion once that’s included for next quarters results. Amazon also said the backlog has “reasonable breadth” and isn’t concentrated in one or two customers which is an indication to me that more deals are coming outside of Anthropic and OpenAI.

When I layer in the Trainium business the numbers are more impressive. Jassy said Amazon’s custom silicon (Trainium, Graviton, Nitro) is now generating over $20 billion in annualized revenue growing triple-digits YoY, and would be a $50 billion run-rate business if it sold chips externally. Trainium has over $225 billion in revenue commitments. Trainium2 is “largely sold out.” Trainium3 just started shipping in early 2026 and is “nearly fully subscribed.” Much of Trainium4, which is still 18 months from broad availability, has already been reserved.

You can’t read those words and tell yourself AI hardware demand is rolling over. Customers are reserving chips that won’t ship for 18 months. That’s the opposite of softness.

Q1 capex was $43.2 billion and Amazon committed roughly $200 billion in 2026 capex which is the largest absolute number among the hyperscalers. Jassy laid out the calculus clearly: AWS has to pay cash for land, power, buildings, chips, servers, and networking 6 to 24 months before billing customers. Customer commitments cover a substantial portion of that capex. The data center assets have 30-plus year useful lives. The economics work over a multi-year window.

Google Cloud’s $462 billion moonshot

Then there’s Alphabet, which delivered the most striking single number of the entire earnings cycle in my opinion. Google Cloud revenue grew 63% YoY to $20.0 billion which was the first time the segment has ever crossed $20 billion in a quarter. The backlog is the wow moment as Google Cloud’s backlog hit $462 billion which was up from approximately $235 billion at the end of Q4. They nearly doubled the backlog sequentially. Their CFO Anat Ashkenazi said on the earnings call that over half of the backlog is expected to convert to revenue in the next 24 months.

That sequential doubling is the most aggressive backlog growth I’ve ever seen out of a cloud business. It tells you Google is signing massive AI compute contracts at a speed that would have been considered insane two years ago. CEO Sundar Pichai said Cloud’s enterprise AI solutions revenue grew nearly 800% YoY. They doubled the count of $100 million to $1 billion deals YoY and signed multiple billion dollar pus deals. Alphabet raised its full-year 2026 capex guide to $180 to $190 billion from $175 to $185 billion. The biggest tell is that Ashkenazi flagged that 2027 capex will “significantly increase” beyond 2026. If they are going to spend roughly $185 billion in 2026 I would think a significant increase would take them over $250 billion in 2027.

The compute deals that broke the model

Here’s the part that that in my opinion removes any remaining doubt about how long this lasts.

Anthropic announced on April 24 that Google was investing up to $40 billion at a $380 billion post-money valuation. That same week, the company expanded its agreement with Amazon to as much as $25 billion. Then, on May 5^th it was reported that that Anthropic had committed to spending $200 billion with Google Cloud over five years. That single contract accounts for more than 40% of Google’s freshly-disclosed $462 billion backlog.

In the same announcement window, Anthropic disclosed that its run-rate revenue had passed $30 billion, up from approximately $9 billion at the end of 2025. They have over 1,000 enterprise customers each spending more than $1 million annualized which doubled from the 500 figure they cited in February.

Now stack OpenAI on top of the recent news from Anthropic as the compute commitments is genuinely staggering:

• $250 billion with Microsoft for cloud services

• $300 billion with Oracle over five years (the Stargate flagship contract)

• $38 billion over seven years with AWS, signed November 2025

• $22.4 billion with CoreWeave through 2029

• $90 billion in committed AMD purchases (the 6 GW MI450 deal with warrants for 160 million AMD shares)

• An unknown but multi-tens-of-billions worth of Nvidia GPU spend tied to Nvidia’s $100 billion investment back into OpenAI

• A custom Broadcom inference chip in development, with first deployment in H2 2026

OpenAI raised $122 billion in March 2026, anchored by Amazon, Nvidia, SoftBank, and Microsoft, at a valuation reported at the trillion-dollar level. Their revenue is reportedly running around $2 billion per month and they’ve communicated a target of $280 billion in annual revenue by 2030 to potential IPO investors.

There are reports that Anthropic and OpenAI contracts together account for roughly $2 trillion in revenue backlog across Amazon, Google, Microsoft, and Oracle. That is $2 trillion in contracts across four companies from two AI labs and doesn’t include a single dollar from anyone else.

Why this lasts longer than the bears think

The bear case for the AI cycle has been that the economics don’t work, capex is outrunning revenue, and the labs will run out of money before the chips pay off. These arguments are incorrect because they assume Amazon’s situation will be replicated across the remaining hyperscalers and it does not incorporate how cash from operations has scaled over an extended period of time or what it will look like in the future.

What the bears conveniently leave out is that over the past decade (not including the TTM) the combination of Microsoft, Amazon, and Alphabet have produced $2.16 trillion in cash from operations while allocating $1.05 trillion towards CapEx. Microsoft has $627 billion of RPO while the AWS backlog is $364 billion (excluding Anthropic) and the Google Cloud backlog is $462 billion. That’s roughly $1.45 trillion in contracted future cloud revenue and almost all of it AI-driven. Anyone who thinks that these companies can’t afford the buildout or won’t monetize the spend hasn’t read their statement of cash flows closely enough. Over the past decade these three companies have generated $1.11 trillion in free cash flow. This is not something that can be looked past.

The other thing that struck me reading through these prints is how often “constrained” came up. Microsoft is constrained. Google is “compute constrained in the near term.” Amazon is selling Trainium chips that won’t exist for 18 months. Memory is on allocation. Power is the new GPU. The bottleneck moved from chips to substations, and substations don’t get fixed by Q3. Higher component prices are showing up in 2026 capex guides because the buyers can’t say no. Microsoft added $25 billion to their capex guide because of memory inflation and didn’t blink.

If you’re trying to time when the AI spend rolls over your likely to spend a tremendous amount of time being wrong because nobody knows. The reality is that the numbers are suggesting that it’s not in the next several years. The contracted RPO across these companies extends into the back half of the decade. The capex commitments are sized for capacity that gets installed in 2027 and 2028. The compute deals from OpenAI and Anthropic run to 2030 and beyond. None of this unwinds quickly.

The bottom line

I came into this earnings cycle focused on the numbers, and they’re not hype. The revenue from the memory and chip companies and the backlog from the hyperscalers are backed by contracts, booked revenue, capex commitments. In this cycle we’re seeing customers paying for compute that doesn’t exist yet because the supply chain can’t move fast enough. The people calling for a slowdown were assuming the buyers would tap the brakes as the bills came in, and they’re not. That tells me the buildout isn’t just continuing, it’s extending. I’ll be tracking the next leg through Nvidia’s earnings, where I expect more of the same with bigger numbers.

For now, the chips are wild. The data says they have plenty of room to run.

Disclosure: This article is for informational and educational purposes only and should not be treated as investment advice. I am not a licensed financial advisor. Every investor should conduct their own due diligence and consider their own financial situation before making any investment decisions. Past performance does not guarantee future results.

If you got this email and the masthead looks different, that’s because it is. I have rebranded to The Fiorillo Letter.

Putting my own name on the door felt like the right move. The work has always been mine, the calls have always been mine, the track record has always been mine. The brand should reflect that.

Nothing else changes. Same cadence, same focus, same voice. I'm still tearing apart 10-Qs, 10-Ks, and earnings transcripts the same way I have been. Hyperscaler coverage continues. PLTR, IBM, SOFI, NVDA, META, QXO, UBER, MSFT, GOOG, AMZN — all the names I've been writing about for years are still on the desk.

A few practical updates.

The new URL is thefiorilloletter.substack.com. Your subscription, paid status if you have one, and reading history are all intact. You don't need to do anything.

If you follow me on Seeking Alpha, YouTube, or X, those are all still going. I'll be updating bios and links on each platform as well. For anyone new here through the rebrand: I'm Steven Fiorillo and I discuss my research that is grounded in business and the stock market. I was recently ranked #3 of 14,748 financial bloggers on TipRanks with a 78% success rate over the trailing year. I run a daily livestream at 11:30 AM ET on YouTube and have written more than 1,500 articles on Seeking Alpha.

That's it. Thanks for sticking around through a name swap. A new article will drop later this week.

Thank you,

Steven Fiorillo

How big the bet actually is

I constructed the chart below to showcase what the hyperscalers are expected to spend in the current fiscal years on CapEx. Amazon alone is going to spend $200 billion on infrastructure and when you combine the spend from these 4 hyperscalers you get $609 billion at the midpoints of guidance.

In the chart below I am showing the YoY growth on CapEx and it’s increasing by more than 70% and could roughly double if we get increased guidance on Wednesday. Microsoft is sitting on $625 billion in remaining performance obligations which are signed contracts for cloud services not yet delivered. That number grew 110% last quarter. Amazon’s AWS backlog is $244 billion, up 40%. These aren’t speculative bets on future demand as customers are pre-committing dollars because the capacity doesn’t exist yet and they need it. The market doesn’t doubt the hyperscalers should be spending and the question this week is whether they’re earning the spend back fast enough.

Microsoft (Wednesday after close)

The consensus estimates have Microsoft printing $4.06 in non-GAAP EPS with revenue coming in at $80.65–81.75 billion. Microsoft closed Monday’s session at $424.82 which is roughly 24% off its 52-week high of $555.45 so the setup going in is about as hostile as you can engineer for a company that’s still printing record numbers. If you weren’t watching last quarter the recap is worth your time because it’s the entire reason Microsoft is siting where it is today. They beat the consensus estimates on every line that matters as revenue increased 17%, and profits were up 24%. Azure beat the companies own internal forecasts and the stock proceeded to drop 10% the following day shedding $357 billion in market cap in a single session. The reason was that management guided Azure growth at 37–38% for this current quarter which is fractionally below the 38% of growth posted in the current print. The small deceleration was enough to convince the market that the AI capex thesis was finally cracking. Whether that reaction was rational or not is its own debate.

Microsoft’s earnings will come down to Azure growth. If the print on Azure exceeds 38% we could see a meaningful rally because it tells the market the capacity constraints from last quarter are easing and re-acceleration is real. If it lands inside the 37–38% guided range the reaction could be muted with another quarter of investors sitting on their hands waiting to see whether the deceleration story is real. If Azure comes in below 37%, the bears are likely to point to this as evidence that the AI buildout isn’t supported by the underlying business fundamentals.

The other line items I’ll be watching closely are RPO and CapEx. The RPO figure of $625 billion is the cleanest piece of demand evidence we have anywhere in the AI complex. If we get another quarter of growth above 80% I believe that it would do more to settle the bubble debate than almost anything else management could say. Microsoft is on track for roughly $104 billion in capex this fiscal year which is up from $64.6 billion in fiscal 2025. If management raises that number Wednesday without a corresponding RPO acceleration to back it up we could see questions being raised again about their aggressive spending.

The cleanest outcome that we could see is Azure above 39%, RPO growing more than 80%, and no capex hike. What I don’t want to see is Azure dipping below 36%, capex raised by another $10–15 billion without a backlog tailwind, and any meaningful softening in the prepared commentary about customer commitments. Anything in between and the stock probably trades sideways while everyone waits for next quarter.

Alphabet (Wednesday after close)

Alphabet’s consensus is pegged at $2.63 in EPS on $106.89 billion in revenue. The EPS line is down about 6% YoY which sounds rough at first but Q1 2025 beat expectations by 40% and set a brutally hard comp. I believe that all eyes will be on Google Cloud as it grew 34% in Q3 and then jumped to 48% in Q4 which is the kind of acceleration you usually only see in a brand-new business segment not in something that’s already running at a $60 billion annual revenue rate. If Q1 comes in above 50% Alphabet basically writes the bull thesis for the entire AI cycle by itself and the $175–185 billion capex commitment for 2026 stops looking expensive against revenue and starts looking cheap against demand.

What changed for me on Alphabet over the last few quarters is the Cloud profitability ramp which I think a lot of investors still aren’t fully pricing in. Full-year 2025 Cloud revenue came in at $58.7 billion with operating income of $13.9 billion which was more than double the $6.1 billion they earned in the prior year. This is an incredibly profitable business growing 30%+ that happens to sit inside a $4 trillion company. If they hold that trajectory through Q1 Alphabet will be setting the tone with the offense in Q1.

The risk on the other side of the print is search which has been the supposed casualty of the AI revolution since ChatGPT launched at the end of 2022. Three years later search ad revenue is still growing in the mid-teens but this is the first full quarter where Gemini has been baked into the core search product and any meaningful slowdown would be a real negative regardless of how good the Cloud number looks. The market has been waiting patiently for evidence that AI is eating Google’s lunch and a soft search line is exactly the kind of thing that gets read as confirmation. The cleanest scenario Wednesday would be for Google Cloud to grow above 50% YoY with Search Ads growing more than 10% and no spike in CapEx allocation for 2026.

Amazon (Wednesday after close)

Amazon’s consensus numbers are $1.63 in EPS with revenue coming in at $177.2 billion. I believe that AWS is the entire ballgame here and everything else in the print is just supporting cast. AWS grew 24% in Q4 which was the fastest growth rate in 13 quarters and ended a long stretch of investors worrying that AWS had permanently lost its growth profile to Azure and Google Cloud. The backlog hit $244 billion up 40% and Andy Jassy was effectively telling investors on the call that capacity is the bottleneck right now not demand. Amazon is monetizing servers about as fast as it can plug them in and that’s the cleanest demand signal in the entire AI complex.

The problem and the reason this print is more risky than it looks on the surface is Amazon’s spend. We’re talking about $200 billion of CapEx in a single year up from $132 billion last year. Amazon’s free cash flow already cratered from $38.2 billion to $11.2 billion and the only thing keeping the market patient is the growth rate reaccelerating for AWS. Management needs to show that the backlog keeps converting into revenue at the same rate or one that is quicker. AWS operating margin is the second variable and it’s the one I think gets the least attention even though it might matter most. The Anthropic situation is the wild card on this print as Amazon is investing up to $25 billion into Anthropic alongside what could be $100 billion in AWS spend flowing back from Anthropic over time. The bears call this round-tripping and argue it’s artificially inflating AWS revenue. I view this as Amazon exchanging capital for equity in a single transaction then Anthropic taking that capital and buying additional compute in a separate transaction. Wednesday’s print is the first chance we get to see how that relationship actually shows up in the financials and how management chooses to characterize it.

Meta (Wednesday after close)

META’s consensus numbers are $6.71 in GAAP EPS on $55.49 billion in revenue and management guided for revenue to come in between $53.5–56.5 billion. The midpoint of that range implies 30% revenue growth which would be Meta’s fastest quarterly growth rate since Q2 2021. META’s 2026 capex commitment is $115–135 billion up from $70 billion last year which means infrastructure spend is going to roughly double in 12 months. The core business that the CapEx is supporting is advertising which is significantly different then Alphabet as Alphabet is supporting Ads and a Cloud infrastructure business. Q4 ad revenue hit $58.1 billion for META which was 97% of Meta’s total revenue. There is no AWS here, no Azure, and no Google Cloud cross-subsidizing the spend. The money comes from one place which makes the narrative harder to defend if the numbers aren’t stellar. Reality Labs is going to lose another $20 billion or so this year on top of the $83.6 billion in cumulative losses they’ve already absorbed since the segment was created. Investors have basically learned to ignore that line but it’s a narrative that continues to warrant a discussion.

What I do want to see on this print is Advantage+ continuing to compound as it now has a $60 billion annual run rate and the AI-driven targeting and creative automation that sit underneath it are producing measurable ROAS improvements for advertisers. This is justifying Meta’s ability to keep pushing pricing. As long as Advantage+ keeps growing the way it has been the model works and the CapEx spend pays for itself in advertising performance over time. I believe how Zuckerberg handles the operating income question may be the most important aspect of Q1 earnings for META. He told investors on the Q4 earnings call that 2026 operating income would exceed 2025’s despite the allocation towards CapEx. If he can keep this promise when depreciation alone is going to be a meaningful drag it could create a bullish setup for investors. I would like to see a print where META’s revenue comes in at $56.5 billion or higher and CapEx is held where it is rather than being raised. The print that gets Meta punished has revenue under $54 billion combined with CapEx exceeding $135 billion in 2026.

What this week actually decides

If I had to boil this down to five aspects to watch from the hyperscalers they would be cloud growth, backlog, margin, CapEx, and depreciation in my opinion.

The first and most important is cloud growth across Azure, Google Cloud, and AWS. If two of the three hit my marks 38%+ for Azure, 50%+ for Google Cloud, and 24%+ for AWS the AI CapEx thesis should stay intact for another quarter and the bear case would be negatively impacted. If two of the three miss the cycle could fall under extreme pressure and we’ll spend the next month watching every analyst on Wall Street recalibrate their growth assumptions for the back half of the year.

The second is backlog growth. RPO at Microsoft and AWS backlog at Amazon need to keep growing in line with CapEx. The combination of backlog deceleration paired with CapEx acceleration could be the most dangerous combination on these prints because its the configuration that tells you the build-out has gotten ahead of the demand.

The third is margin discipline and I believe that the margin for AWS needs to hold 35% or more. I also feel that Microsoft’s overall operating margin needs to stay where it is. If these start to crack then the market could start asking harder questions about ROI.

The fourth revolves around CapEx and whether any of the hyperscalers come out and increase guidance. After the January and February guides the next move is the tell. I believe CapEx acceleration would be welcomed by the market if the operating profitability and RPO is expanding but if CapEx is hiked without the financials to back it up we could see the market through a fit and punish these companies as they have done in the past.

The fifth and the one that could be underrated is the depreciation lag commentary. The reality is that new capacity hits the income statement immediately as depreciation while the revenue from that capacity ramps over 6–12 months. The gap could compress operating margins through 2026. How these management teams talk about this issue on their earnings calls could shape the consensus expectations for the rest of the year.

What would force me to rethink this narrative

The bear case I take seriously isn’t the one that says the AI build-out is fake. The customers are real, the contracts are real, the capacity demand is real, and you can see all of it in the financial statements if you know where to look. I would place the most emphasis on the idea of timing as the depreciation lag and the margin pressure that comes with it could compress earnings power for the next 12–18 months before the new infrastructure is fully revenue-productive. If that happens multiples can compress even when the underlying businesses are doing exactly what they’re supposed to do. That’s a different bear case than the one most people are arguing and it’s the one that I believe makes the most sense. The honest answer about this week is that I don’t know which way the prints are going to land because I don’t have a crystal ball. All of the numbers from companies such as Micron, Tiawan Semiconductor and Intel make me believe that the narrative and spend will continue to favor the bulls but we’re going to need these prints to be clean for the bear case not to overtake the narrative.

My take is based on following the money

Here’s where I land coming into Wednesday. The bear case has had three years to produce evidence and the closest thing it has to a smoking gun is the fact that CapEx is increasing. The hyperscalers are spending more because their customers are signing more contracts not because they’re guessing about future demand and hoping it materializes. Microsoft’s $625 billion in remaining performance obligations and Amazon’s $244 billion AWS backlog are the financial equivalent of the line out the door at a restaurant that hasn’t opened yet. You don’t ramp infrastructure spend by 70% in 12 months because you’re confused. You do it because the revenue on the other side of the build is already booked.

The single most underappreciated fact about this cycle is that the hyperscalers are funding it almost entirely out of operating cash flow. The reality is that in the previous fiscal year the combination of Microsoft, Alphabet, META, and Amazon generated $556.19 billion in cash from operations while allocating $357.51 billion on CapEx. This left them with $198.68 billion in free cash flow. These are some of the largest cash-generating businesses in corporate history choosing to reinvest at historically high rates because the returns on data centers are still well above their cost of capital. That’s not a bubble but rather capitalism working exactly the way it’s supposed to. I believe that the $600 billion plus of CapEx that is expected to be allocated this year will be defended on Wednesday and I am excited to dive into the new numbers and provide my thoughts on where we go from here.

Disclosure: This article is for informational and educational purposes only and should not be treated as investment advice. I am not a licensed financial advisor. Every investor should conduct their own due diligence and consider their own financial situation before making any investment decisions. Past performance does not guarantee future results.

April 15^th was the day the S&P 500 finally closed above 7,000 at 7,022.95. It had flirted with that level months earlier and hit an intraday high of 7,002.28 on January 28^th but it couldn’t hold onto this accomplishment into the close. The next thing we know is that the Iran conflict showed up and we were staring at 6,583 on the S&P \when the market closed on April 2nd. Ten trading days later we were at 7,041 as the markets rallied 6.96%. If you sold during the selloff and went to cash you watched nearly 7% rip happen without you. Now you’re probably waiting for a pullback to get back in. I’m going to tell you right now that the pullback you’re waiting for isn’t guaranteed to come at the time or at the level you want. I’m going to walk you through why throughout this article.

Here’s the short version, time in the market beats timing the market, not sometimes, not on average, but essentially always over an extended period of time. The data on this has been screaming at retail investors for four decades and people still don’t believe it. I will lay out the specific numbers on drawdowns, discuss the psychology that wrecks portfolios during them, look at how hard re-entry actually is, and on the math of dollar-cost averaging. My goal is to make the case that the most boring possible strategy which is automatic, unemotional, and permanent buying of the S&P 500 beats most strategies that retail investors actually try.

First: the market is underwater most of the time

People anchor on all-time highs and any deviation feels like a problem. Investors should look at how often the S&P 500 is actually in a drawdown of some kind. Since 1950 the index has spent more than 60% of its trading days at least 3% below its prior all-time high. On half of those days it’s off by at least 5%. Full bear markets where the decline is greater than 20% peak to though has happened 15 times since 1928 and accounted for 8,520 trading days which is roughly a quarter of this period. The calm uptrend people imagine is not actually what the market spends most of its time doing. The market spends most of its time annoying the people who own it.

The reason why I brought up these statistics is because if you try to avoid drawdowns by going to cash when things look risky you are trying to sidestep the default state of equity markets. Drawdowns aren’t anomalies but rather the mechanism that produces the equity risk premium in the first place. You cannot collect the premium without enduring the drawdowns.

What drawdowns actually look like when you zoom in

Morgan Stanley’s Counterpoint Global team ran a big study on S&P 500 drawdowns covering 1985 through 2024. The worst the index did over those four decades was a 58% peak-to-trough decline which happened during the 2007-2009 financial crisis. This peak to trough decline took about 1.4 years. The full recovery back to the prior peak took 4.2 years. The part nobody discusses is that following that trough the annual total return on the S&P 500 was 25% over the next five years.

The worst drawdown in four decades bottomed in March 2009 and if you sold somewhere during the crisis you had to make the decision to re-enter while 25% annualized returns were already running. The cohort that panicked in 2008 largely stayed in cash through 2010, 2011, 2012. The cohort that kept buying through the drawdown adding every month without flinching rode that 25% annualized recovery on shares they’d bought at generational discounts.

COVID was the same pattern just at an accelerated rate. It took 32 days to go from peak to trough which included a handful of weeks of absolute chaos. We experienced circuit breakers, bond market dysfunction, headlines calling it the end of everything then a V-shaped recovery that left the panic sellers buying back in at higher prices within months. If you kept your 401(k) contribution running through March and April 2020 you’re the one who won. If you paused it your likely still behind.

The part about your brain that you probably don’t want to hear

The data is clear as day yet investors that are smart with advanced degrees and even industry professionals still panic-sell during drawdowns. The reality is that the human brain is not built for this. Daniel Kahneman and Amos Tversky published a paper in 1979 called “Prospect Theory.” Kahneman won the Nobel Prize in Economics in 2002. The paper’s core finding is devastatingly simple. Humans feel the pain of losses roughly 2 to 2.5 times more acutely than they feel the pleasure of equivalent gains. A $1,000 loss produces emotional distress that requires something like $2,000 to $2,500 in gains to offset. The asymmetry is baked in at the wiring level and its not a preference or a choice. It’s a feature of how our brains evolved to process threats.

By this logic if you had a $500,000 portfolio and the market dropped by 20% you would be down $100,000 on paper. Per the loss aversion literature the pain of that paper loss is roughly equivalent to the pleasure you would receive from generating a gain that exceeded $200,000. Your brain is treating this as a catastrophe but nothing actually happened. Consider the fact that you still own the same shares of the same businesses but the account balance moved and your threat system is firing. The urge to do something to stop the pain becomes overwhelming. What feels like safety in that moment is selling because makes the balance stop moving. The problem is that you’ve now permanently impaired your portfolio’s ability to participate in the recovery that’s statistically almost certain to follow.

Prospect theory also tells us something about what happens next. When people are in the loss domain they become risk-seeking rather than risk-averse. This means that the investor who panic-sold at a 20% drawdown doesn’t calmly re-enter when things stabilize, they hold out for a better price because they need to feel like they won the trade. Accepting the current price means admitting they were wrong to sell and admitting mistakes is psychologically expensive. So investors wait and the market keeps climbing. Every additional rally makes re-entry more painful because now they’re buying back in at a price even higher than the one they sold at. This causes some investors to wait extended periods of time or never get back into the market.

Why getting back in is harder than anyone tells you

Re-entry is a problem because this is where most market-timing attempts actually go to die. Picture an investor sitting on cash in February 1995. They want to wait for a pullback before deploying. They would settle for a 5% pullback which may seam reasonable to some but that pullback didn’t arrive until seventeen months later in July 1996. When the pullback came the S&P 500 was still more than 25% higher than it had been in February 1995. The pullback they waited for never gave them a better entry than their starting point. It just gave them a smaller dip off a much higher base. Say you want to wait for a real drawdown before committing your cash and your looking for a 20% decline on a monthly closing basis. What are the odds that waiting will actually give you a better entry than just investing today? Going back to 1928 the answer is about 20%. The bear market you’re waiting for eventually bottoms at a level higher than where you started waiting because the market grinds upward between bear markets fast enough to swallow most drawdowns before they ever reach your target price.

Schwab ran a really clean study on this where they utilized five investors from 2005 through 2024. Each investors put $2,000 into the S&P 500 annually.

• Peter Perfect somehow timed every single contribution at the exact yearly low. Ended with $186,077.

• Ashley Action did no timing at all, just invested her $2,000 on the first trading day of every year. $170,555.

• Matthew Monthly dollar-cost averaged across 12 equal monthly payments. Third place.

• Rosie Rotten had perfect anti-timing — she invested at the yearly high every single year. Approximately $151,343.

• Larry Linger kept waiting for a better moment. Never got around to investing. $47,357 in T-bills.

These are not my numbers or my scenario, its Schwab’s. When you look at the data the investor with literally perfect hindsight that picks every annual low only beat the investor who did zero timing by $15,522 over two decades which works out to about $776 a year. Meanwhile the investor with the worst possible timing every single year still beat the investor who utilized T-bills as a supplement for cash by $103,986. The penalty for waiting is vastly larger than the penalty for bad timing and the penalty for bad timing is tiny compared to the reward for just showing up.

Fidelity did a bigger version covering 1980-2023 with $5,000 annual contributions. Perfect timing ended at roughly $5.6 million. Worst-possible timing which bought every peak ended at $4.2 million. Cash ended at $349,999. The gap between perfect and worst was about $1.4 million. The gap between worst and cash was nearly $3.9 million. The penalty for being the worst possible timer in history was real but it was nothing compared to the penalty for not investing at all. The game isn’t timing. The game is being invested versus not invested.

Why Boring Keeps Winning

The S&P 500 has never posted a negative annualized return over any 20-year rolling period in its history going back to 1928. This includes throughout periods including the Great Depression, World War II, 1970s stagflation, the Lost Decade of 2000-2009. The worst 20-year stretch ever recorded still produced a positive real return. The median delivered around 7.3% annualized after inflation. Every 30-year period has made money. That’s the base rate you’re betting against when you go to cash. Meanwhile the JPMorgan study shows that if you missed just the 10 best trading days over a 20-year window your annualized return drops from 10.5% to 6.2%. If you miss the 30 best days and your down to 1.4%. The data shows that 7 of the 10 best days over the last two decades happened within 15 days of the 10 worst days. This means that the investor who panics into cash during the ugly stretches is mathematically avoiding the best days also.

The conclusion writes itself. Pausing investments during drawdowns and waiting for pullbacks trying to outsmart the market is mathematically and psychologically improbable. Setting up an automatic contribution and letting it compound over 20 to 30 years without do anything allows you to live your life while letting your capital work for you. The strategy isn’t sophisticated and it isn’t exciting. It’s certainly not going to get you invited to talk at investment seminars. What it is likely to do is set your future self up for success. Here are the numbers I want to leave you with.

Go to any investment calculator and punch this in. If you start with $0 and add $500 per month for 30 years at an annual rate of return of 10% the ending balance is $1,031,421.66. Your total contributions are $180,000 and the appreciation is $851,421.66. If you’re lucky enough to squeeze out an extra 1% and get an annualized return of 11% the ending balance is $1,253,194.88 which generates an additional $221,773.20 in appreciation.

Last week’s rebound from 6,583 to 7,041 in ten trading days was a gift to the people who kept contributing through the Iran selloff. It was a penalty to everyone who paused. The next scary week is coming, probably sooner than most people expect, and the same thing will happen again. The buyers will win while the investors who pause will spend the following year inventing reasons why the next pullback will be their real entry.

Over an extended period boring wins because boring removes the decisions that break portfolios. If you never stop contributing, you never miss a best day and if you never try to time anything you never miss a bottom. Every fancy thing you don’t do is a behavioral mistake you don’t make. The math is the math, and the data is the data. This may not be the best strategy for generating alpha but continuously buying the S&P 500 over an extended period has been proven to work time and time again.

Disclosure: This article is for informational and educational purposes only and should not be treated as investment advice. I am not a licensed financial advisor. Every investor should conduct their own due diligence and consider their own financial situation before making any investment decisions. Past performance does not guarantee future results.

In early February 2026 roughly $285 billion in market cap disappeared from software stocks in 48 hours after Anthropic launched Claude Cowork and OpenAI followed with its own agentic AI system for enterprises. The financial press called it the “SaaSpocalypse” and it marked the single largest AI-triggered repricing event in software history.

I continue to think about how software companies will be impacted by AI and I keep coming back to the same conclusion. The fear that AI can now let companies build their own tools in-house is legitimate. When a small team can spin up functional bookkeeping software, a project management tool, or a niche solution in a weekend via vibe coding, the $50-per-seat pricing model starts looking indefensible.

In today’s market investors are currently punishing fortified platforms alongside vulnerable tools which is creating one of the more interesting opportunities I’ve seen in a while. My thesis is simple. Companies with deep network effects are not threatened by AI but rather strengthened by it. The market is mispricing them because it is conflating “AI disrupts software” with “AI disrupts everything in tech.”

You Can’t Vibe Code Enterprise Grade Software

Enterprise-grade software is being unfairly dragged down by the SaaSpocalypse narrative rather than being valued on business fundamentals. Companies like CrowdStrike in cybersecurity and Workday in enterprise resource planning and human capital management haven’t sold off because their valuations were too high. They sold off because the market bought into the idea that someone in a company’s IT department could vibe code a comparable product and save the organization money. These are not simple point solutions that a startup can replicate over a weekend. These are deeply embedded systems that run the most sensitive operations inside Fortune 500 companies. The likelihood that a C-suite executive is going to risk their career or the company’s well-being by ripping them out and replacing them with something that was vibe coded is incredibly slim.

First there are contractual obligations that many of the people screaming about how vibe coding is going to replace enterprise software doesn’t understand. Enterprise cybersecurity and ERP vendors typically lock customers into multi-year agreements, usually that are typically three to five years. CrowdStrike’s enterprise contracts commonly run on multi-year commitments with custom pricing based on modules deployed and the number of endpoints protected. Workday’s subscription contracts operate similarly, often spanning multiple years with implementation costs running anywhere from $500,000 to north of $5 million for large enterprises. These are not subscriptions you cancel on a whim. They are strategic commitments that go through procurement, legal, compliance, and board approval.

Next the SaaSpocalypse believers need to consider how difficult implementations actually are. I would be willing to bet that even if they have worked at a large company, they have never lived through the reality of onboarding a true ERP system. A Workday ERP deployment typically takes 6 to 18 months depending on organizational complexity, and that is with a dedicated outsourced consultant team running the project. The process involves migrating years of legacy payroll data, benefits records, tax configurations, org structures, and compliance rules. It requires mapping every business process, integrating with dozens of third-party systems, running parallel testing cycles, and training thousands of end users before anyone flips the switch. Gartner estimates that by 2027, 70% of ERP implementations will fail to fully meet their original business goals, and 25% of those will fail catastrophically. That is with professional implementation teams and proven enterprise software. Imagine pitching the board on replacing Workday with something a few developers built using AI in a couple of weeks and the harm it would do to your reputation when the answer is no. Years of trust would be damaged and your decision-making ability would be questioned.

What makes these SaaS products almost impossible to dislplace from vibe-coding is their data moat. An enterprise ERP isn’t just software, it’s the system of record for every employee’s compensation, benefits, tax withholdings, time tracking, PTO accruals, and compliance documentation. Years of historical data, audit trails, and regulatory configurations are baked into these platforms and they don’t transfer to a new system cleanly or quickly. CrowdStrike’s Falcon platform is just as entrenched because it sits on every endpoint across the organization, continuously ingesting telemetry data, building behavioral baselines, and correlating threat intelligence across millions of signals. The AI models powering CrowdStrike’s detection capabilities are trained on this proprietary data and they get smarter the longer they run inside a customer’s environment. Someone who vibe codes an alternative starts with zero threat intelligence, zero behavioral baselines, and zero institutional knowledge of how that specific customer’s network actually behaves. You’re not competing with CrowdStrike’s code at that point, you’re competing with years of accumulated data that you have no way of replicating.

Here is the point that I think gets overlooked in every AI disruption conversation: career risk. Put yourself in the shoes of a CFO or a CISO at a medium to large sized company. You are personally accountable for the integrity of financial reporting, payroll accuracy, regulatory compliance, and cybersecurity posture. If something goes wrong, it is your name on the line. Now imagine walking into an audit committee meeting and explaining that you replaced Workday’s enterprise payroll system, a platform used by thousands of Fortune 500 companies with decades of compliance infrastructure with a custom application built by a small team using AI tools over the course of a few weeks. No established vendor behind it, no SLA’s, no SOC 2 compliance and no proven track record. Nobody with an MBA and decades of operational experience is making that call.

The cybersecurity argument is even more straightforward. After the high-profile breaches of the past several years, no CISO is going to stake their reputation on an unproven, AI-generated security stack. The downside is existential as a single breach can cost hundreds of millions in damages, regulatory fines, and reputational destruction. Companies pay CrowdStrike because they need a proven platform with a demonstrated track record, 24/7 managed detection, and the legal and contractual protections that come with an established vendor relationship. That is not something you can vibe code.

If the career risk argument is not compelling enough, consider the legal exposure that an organization could face. Payroll is one of the most regulated areas of an enterprise’s operations. If a system error causes missed paychecks, incorrect overtime calculations, or wrong tax withholdings, the consequences are severe and in some cases lead to legal action. Wage and hour violations are the one of the largest categories of employment litigation in federal courts. These cases escalate fast as individual wage claims routinely get certified as class actions. The settlements can reach tens or even hundreds of millions of dollars. For example Kroger settled a $21 million class action last year after employees alleged the company missed paychecks and made inaccurate deductions following a payroll system switch. That was a transition between established enterprise platforms, not an experiment with unproven software. The litigation risk of deploying an untested, vibe-coded payroll system across thousands of employees is the kind of exposure that keeps general counsels and compliance officers up at night.

So when the market sells off CrowdStrike and Workday because the narrative is that vibe coding is going to decimate SaaS products I think there is a misunderstanding of what enterprise software businesses actually are. They are not selling code, what they are selling is compliance infrastructure, contractual protection, regulatory expertise, and the institutional trust that comes from running the most sensitive operations inside the world’s largest companies. AI will make these platforms better, cause them to have smarter threat detection, and produce more predictive workforce analytics but its not going to replace them. The switching costs are too high, the implementation timelines are too long, the data is too deeply embedded, and the career and legal risks of going with an unproven alternative are simply too great.

Why Network Effects Are Actually AI-Proof

A network effect occurs when each additional user makes the product more valuable for every other user on the platform. Instagram is not valuable because of its interface it’s valuable because everyone you know is already on it. Every friend who joins gives you one more reason to stay, one more reason to post, and one more reason to never leave for a competitor. While AI can replicate the functionality of a software interface, it cannot replicate the social, physical, or financial momentum of a mature network.

What I find most compelling about this dynamic is that AI does not just fail to disrupt these businesses, it creates a flywheel that makes them stronger. More users generate more proprietary data which trains better AI models that improves the product and attracts more users. The incumbents with massive datasets can fine-tune specialized models that provide better accuracy and lower inference costs than any competitor could achieve from scratch. Every interaction becomes a learning opportunity and the system gets better the more it is used.

Meta Platforms — The Purest Network Effect in Tech

I believe Meta is the clearest example of this thesis. As of their Q4 2025 report, Meta’s Daily Active People metric hit 3.58 billion across Facebook, Instagram, WhatsApp, and Messenger. This means that more than 1/3rd of the global population touches at least one Meta platform every single day. Facebook, Instagram, and WhatsApp have each individually crossed the 3 billion monthly active user threshold.

The cross-platform overlap is a structural advantage that does not get talked about enough. The vast majority of Instagram users are also active on Facebook and WhatsApp, creating an integrated user base that is almost impossible to peel apart. What makes Meta work is not vibe coded code. It’s the reality that everyone you know is already there. The billions of daily sessions, photos, shared memories, and the groups created constitute the product itself. Nobody is building a competitive social network in their garage that can replicate the social graph of 3.5 billion people.

AI is supercharging Meta rather than threatening it. Their Reels recommendation engine has gotten meaningfully better with Facebook surfacing over 25% more same-day Reels in late 2025. On Instagram 75% of recommendations are now coming from original posts. The ad targeting side is even more impressive as their Generative Ads Recommendation Model (GEM) doubled the GPU compute used for ranking which made the system significantly faster at matching ads to audiences. In January 2026 Threads officially surpassed X in daily mobile active users with 141.5 million versus 125 million. This illustrates the power of cross-promoting into an existing 3+ billion userbase.

I’ll say something that might ruffle some feathers. I think a smaller software company that offers a single specialized tool such as a niche analytics dashboard, a standalone scheduling app, or a basic invoicing product is in significantly more danger from AI disruption than software companies with large communities and network effects. One company’s entire value proposition is functionality that AI can now replicate in a weekend.

Uber — You Can’t Vibe Code a Driver on Every Block

Uber represents a different flavor of the same thesis which is a two-sided marketplace protected by physical-world network effects. Their Q4 2025 numbers were striking as there were 3.8 billion trips which increased 22% year over year. They reported $14.4 billion in revenue, 9.7 million monthly drivers and couriers, and 202 million monthly active users. For the full year, Uber generated $9.8 billion in free cash flow which was up 42% year-over-year.

The flywheel here is that more drivers lead to shorter wait times, which attracts more riders. This creates more demand which attracts more drivers. This loop took over a decade and billions of dollars to build. A startup could vibe code a ride-hailing app in a day, but they cannot vibe code a driver on every block in 10,000+ cities.

Uber is using AI to squeeze more value from every trip. Their AI-driven dynamic pricing algorithms adjust fares and driver incentives in real time to balance supply and demand. The financial leverage is showing up clearly in their margins, as their adjusted EBITDA margin hit 4.6% of gross bookings in Q4 which was up from 4.2% in 2024.

Many participants in the market believe that autonomous vehicles are Uber’s biggest threat. I am taking the opposite side of this as Uber is positioning itself as the essential demand layer for AV fleets. Their partnerships with robotaxi operators demonstrate that AV companies need Uber’s 200 million users to achieve high utilization rates. Rather than getting disrupted by autonomous vehicles, Uber is becoming the distribution platform they all need to plug into. Remember anyone can come up with a cool concept but access to vehicles and distribution are the keys to success and Uber has a stronghold on those aspects of the ride hailing industry.

My conclusion

The market right now is making a category error as it’s treating small SaaS companies, enterprise software platforms, and platforms with incredible userbases as if they face the same existential threat from AI. They don’t and selling off enterprise grade software and platforms because the market can’t decide how far AI will penetrate the software sector is crazy.

The SaaSpocalypse was a rational repricing for a specific category of software. If a company’s entire value proposition is functionality, and AI can replicate 80% of that functionality in a weekend, the stock deserved to come down. I have no argument there but the market went too far by dragging down businesses whose value has nothing to do with the elegance of their code. Network-effect platforms are not in the blast radius of AI disruption. They are the ones deploying AI as a weapon on top of moats that took decades and billions of dollars to build. Meta is using AI to improve ad targeting across 3.58 billion daily users. Uber is using AI to optimize pricing across 3.8 billion quarterly trips.

Enterprise systems like CrowdStrike and Workday occupy a different but equally defensible position. The multi-year contracts, 6 to 18 month implementation cycles, embedded data, and the career and legal risks of switching to an unproven alternative create a level of entrenchment that no amount of vibe coding can overcome. AI is more likely to make these platforms better rather than replacing them.

Artificial intelligence is becoming a commodity input and eventually every company will have access to powerful models. That means the sustainable advantage will not come from having better AI. It will come from having the distribution, the network density, and the institutional trust to deploy that AI at scale. In a world where anyone can code the only thing that matters is who is already connected and the SaaSpocalypse may have created an incredible entry point into some of the largest businesses in the market.

Disclosure: I am a shareholder in Meta Platforms. This article is for informational and educational purposes only and should not be treated as investment advice. I am not a licensed financial advisor. Every investor should conduct their own due diligence and consider their own financial situation before making any investment decisions. Past performance does not guarantee future results.

Every time I flip on a financial news network or scroll through a bearish thread on X, it’s the same recycled script, someone comparing the AI buildout to the Dot-Com bubble. It’s a comfortable narrative because it leans on nostalgia instead of math. I’ve been pushing back on these takes for months now, but the data has reached a point where I can’t just let it sit anymore.

If you’re going to call this a bubble, you need to account for the $1.66 trillion in contracted demand sitting on the books of the Big Four. You have to explain why the most profitable companies in human history are funding this entire buildout with their own cash flow, not venture debt, not equity raises, not convertible notes. You also need to reckon with the fact that Jamie Dimon just called the capital pouring into AI infrastructure an economic tailwind in his most recent shareholder letter. The man runs JPMorgan Chase. He has no incentive to pump AI stocks.

I’m going to strip away the talking-head opinions and stack today’s earnings, free cash flow, and backlogs against the actual wreckage of the Dot-Com era. If this is a bubble, the numbers will scream it. If it’s not, then the “AI Bubble” narrative might be the most expensive mistake you make this decade.

What a Real Bubble Actually Looked Like

You can’t slap the bubble label on something without understanding what a real one looked like from the inside. The internet in the late 1990s was a genuine technological revolution. Nobody serious argues otherwise. What went sideways had nothing to do with the technology itself. It was how Wall Street and Sand Hill Road threw money at it.

From 1995 to 2000, the Nasdaq ripped roughly 400–500%. Anything with a “.com” after its name suddenly commanded a billion-dollar valuation. VC firms had raised massive funds and needed to deploy capital, so due diligence got thinner, valuations ballooned, and the prevailing wisdom became “scale at all costs, figure out revenue later.” Companies with zero earnings, negative cash flow, and no realistic path to profitability were getting funded round after round after round.

By the time 2000 rolled around, most of these companies were completely dependent on the next funding round just to keep the lights on. There was no internal cash generation. There were no contracted revenue backlogs. It was a constant conveyor belt of new investor capital keeping the whole machine alive. When that conveyor belt stopped, everything collapsed.

The Fed hiked rates 175 basis points between June 1999 and May 2000. The Nasdaq peaked at 5,048 on March 10, 2000. Japan slipped back into recession. MicroStrategy had to restate earnings. The IPO window slammed shut, secondary offerings dried up, and companies that needed outside capital to survive suddenly couldn’t get any. Margin calls triggered forced selling. The telecom infrastructure companies that had borrowed heavily to build out networks including WorldCom, Global Crossing, and NorthPoint went bankrupt. The Nasdaq dropped roughly 78% from peak to trough.

That was a bubble. Cash-burning companies with no revenue, financed by endless rounds of other people’s money. Keep that picture in your head, because what I’m about to walk through looks nothing like it.

Who’s Actually Spending — and Where’s the Money Coming From?

This is where the Dot-Com comparison falls apart immediately. Ask yourself one question: who is writing the checks?

In 1999, it was venture-backed startups burning through other people’s money. In 2025 and 2026, it started with Microsoft (MSFT), Amazon (AMZN), Alphabet (GOOGL), and Meta Platforms (META), 4 of the most profitable businesses that have ever existed. Their spending record amounts on data center infrastructure and funding virtually all of it out of their own operating cash flow. In the second half of 2025 Oracle (ORCL) solidified itself in the conversation as well.

I’ve said this before on X and I’ll keep saying it: when the CEOs and boards at these companies are taking a blank-check approach to building out AI infrastructure, they probably know something the bears sitting on the sidelines don’t. Everyone has a right to their opinion, but these are not pre-revenue startups chasing a dream. Not even close.

Quick note on methodology: MSFT, NVDA, and ORCL don’t report on a calendar year, so I’m using trailing 12-month figures throughout to keep everything on equal footing.

Trailing 12-Month Revenue

The Individual Breakdowns

Microsoft — $305.5 Billion in TTM Revenue and Accelerating

MSFT’s trailing 12-month revenue came in at $305.5 billion. Operating cash flow over the same stretch hit roughly $160 billion against about $83 billion in CapEx, leaving $77.4 billion in free cash flow.

The most recent quarter tells the growth story. Q2 FY2026 delivered $81.3 billion in revenue, up 17% year-over-year. Operating income grew 21% to $38.3 billion. Microsoft Cloud blew past the $50 billion quarterly mark for the first time at $51.5 billion — a 26% jump. Azure put up 39% growth. Cash from operations was $35.8 billion in a single quarter.

Now here’s where the bears lose it. MSFT dropped $37.5 billion on CapEx in Q2 alone, 66% more than the year-ago period. H1 FY2026 CapEx totaled $72.4 billion, which puts them on pace for around $100 billion in full-year spending. Sounds reckless — until you look at the demand backing it up. Commercial remaining performance obligations surged 110% to $625 billion. Yes, about 45% of that is OpenAI. But strip OpenAI out entirely and you still have $344 billion in diversified enterprise RPO growing 28% year-over-year. Amy Hood went out of her way to note that the non-OpenAI portion is larger and more diversified than most of their peers. MSFT still has $94.6 billion in cash and short-term investments sitting on the balance sheet. Nadella said it plainly: they’ve already built an AI business bigger than some of their largest existing franchises, and they’re just getting started.

Amazon — $716.9 Billion in Revenue, AWS Reaccelerating

Full-year 2025 revenue landed at $716.9 billion, up 12%. Operating cash flow was $139.5 billion, a 20% increase. Q4 alone did $213.4 billion in revenue.

AWS is where the AI story lives for Amazon. Q4 AWS revenue came in at $35.6 billion, growing 24% — the fastest growth rate in 13 quarters. The annualized run rate is now roughly $142 billion, and AWS threw off a 35% operating margin. Backlog hit $244 billion at year-end, up 40% year-over-year. On the infrastructure side, AWS added 3.9 gigawatts of power capacity over the trailing 12 months. For context, that’s 2x the total capacity AWS had back in 2022 when it was running at an $80 billion annual rate.

Andy Jassy said on the earnings call that customers want AWS for both core and AI workloads, and that they’re monetizing capacity as fast as they can install it. Matt Garman, the AWS CEO, went further — his best estimate is that AWS will remain capacity-constrained for the next couple of years. They spent $131.8 billion on CapEx in 2025 and guided to roughly $200 billion for 2026. The demand is there. They can see it in the backlog. They’re building to fill orders, not to speculate.

Alphabet — $403 Billion in Revenue, Cloud Growing 48%

Alphabet crossed $400 billion in annual revenue for the first time in 2025, finishing at $403 billion on 15% growth. Operating cash flow was $164.7 billion, up 27%. After spending $91.4 billion in CapEx, they still generated roughly $73.3 billion in free cash flow.

Google Cloud was the standout. Q4 cloud revenue grew 48% to $17.7 billion with operating margins around 30%. The cloud backlog jumped to $240 billion — up 55% from just the prior quarter and more than double where it sat a year ago. Alphabet signed more billion-dollar cloud deals in 2025 than in the previous 3 years combined. Gemini now has over 750 million monthly active users, and Alphabet’s models are processing north of 10 billion tokens per minute through direct API use.

For 2026, Alphabet is guiding CapEx of $175 billion to $185 billion. That’s close to double what they spent in 2025 and roughly $60 billion above what the Street was expecting. Pichai pointed out that they reduced Gemini serving-unit costs by 78% during 2025 through optimization and utilization improvements, so the infrastructure is getting more productive with scale, not less. They told investors they expect to remain supply-constrained through 2026.

Meta Platforms — $201 Billion in Revenue, 41% Operating Margins

Full-year 2025 revenue hit $201 billion on 22% growth. Q4 operating income was $24.7 billion at a 41% operating margin. Full-year operating cash flow came in at $115.8 billion, and free cash flow was $43.6 billion.

Meta is the AI monetization proof point that bears conveniently ignore. The Family of Apps reaches 3.58 billion daily active people. Their Advantage+ AI ad suite is already handling over $60 billion in annualized ad spend. They put $72.2 billion into CapEx in 2025 and guided $115 to $135 billion for 2026. Every time someone tells me there’s no ROI on AI spending, I point them at Meta’s ad business. They are printing money directly because of AI-driven improvements to targeting, creative generation, and recommendation systems. You don’t stumble into 41% operating margins.

Nvidia — The Picks and Shovels, $215.9 Billion in TTM Revenue

NVDA’s fiscal 2026 just closed in January and effectively serves as the trailing 12-month figure. Revenue: $215.9 billion, up 65%. Q4 revenue alone was $68.1 billion, up 73%. Data Center revenue in Q4 hit $62.3 billion — a 75% increase. The company now generates over 91% of its revenue from Data Center.

On the cash flow side: $102.7 billion in operating cash flow, $6 billion in CapEx, and approximately $97 billion in free cash flow. Net income was $120.1 billion. They returned $41.1 billion to shareholders through buybacks and dividends, and the cash pile still grew to $62.6 billion. NVDA guided Q1 FY2027 revenue at $78 billion. Colette Kress told investors the company has visibility to $500 billion in revenue from Blackwell and Vera Rubin through the end of calendar 2026. She also said NVDA believes total AI infrastructure spending globally could reach $3 to $4 trillion per year by 2029 or 2030. These are the numbers from the company supplying the GPUs that make the whole buildout possible.

Oracle — $553 Billion in RPO, the Infrastructure Dark Horse

Oracle doesn’t get nearly enough attention in this conversation, but the numbers are wild. Q3 FY2026 revenue came in at $17.2 billion, up 22%. Cloud revenue was $8.9 billion, up 44%. Cloud infrastructure alone grew 84% to $4.9 billion.

On a trailing 12-month basis, Oracle generated about $64 billion in revenue and $23.5 billion in operating cash flow while deploying $48.3 billion in CapEx. The headline number is remaining performance obligations: $553 billion, up 325% year-over-year. They’re guiding to $50 billion in CapEx for FY2026 and $67 billion in revenue. Oracle has landed cloud deals with OpenAI, xAI, Meta, NVIDIA, and AMD. When you add Oracle’s RPO to the backlogs at MSFT, AMZN, and GOOGL, the total contracted demand across these 4 companies exceeds $1.66 trillion.

The Combined Picture

TTM Capital Expenditures vs. 2026 Guidance

MSFT ~$100B estimate based on H1 FY2026 spending of $72.4B and analyst consensus. Sources: Company filings and earnings calls.

Contracted Backlog / Remaining Performance Obligations

Combined Total: $1.662 Trillion in contracted demand

Sources: MSFT commercial RPO (Q2 FY26). AMZN AWS backlog (Q4 2025). GOOGL cloud backlog (Q4 2025). ORCL RPO (Q3 FY26).

TTM Operating Cash Flow vs. Capital Expenditures — Big Four Hyperscalers

On a trailing 12-month basis, the Big Four hyperscalers pulled in roughly $1.63 trillion in combined revenue, generated $580 billion in operating cash flow, spent approximately $379 billion on CapEx, and still walked away with about $206 billion in free cash flow. After spending $379 billion building out data centers and AI infrastructure, they still had $206 billion left over.

That CapEx number is only going up. The 2026 guidance from these 4 plus Oracle could push combined spending past $675 billion. Jamie Dimon pegged it even higher in his 2025 shareholder letter — $450 billion across 5 hyperscalers in 2025, ramping to roughly $725 billion in 2026. I’ll come back to Dimon’s comments in a moment because they matter.

During the Dot-Com era, CapEx was funded by venture debt, equity offerings, and convertible notes from companies that had no revenue. In 2025 and 2026, CapEx is being funded by $580 billion in annual operating cash flow generated by 4 companies that collectively produce $1.63 trillion in revenue. There is no version of reality where those 2 pictures look alike.

$1.66 Trillion in Contracted Demand

Forget the theoretical TAM. Forget the commentary about future potential. Look at what customers have actually signed up for and committed to in writing.

• MSFT: $625 billion in commercial RPO

• ORCL: $553 billion in RPO

• AMZN: $244 billion in AWS backlog

• GOOGL: $240 billion in cloud backlog

That’s $1.66 trillion in contracted demand sitting on the books right now. The Dot-Com companies were building capacity and hoping customers would show up. The AI hyperscalers are building capacity because customers have already shown up, signed the contracts, and committed the dollars.

I don’t know how else to say it: the demand has a receipt.

Jamie Dimon Said the Quiet Part Out Loud

I want to bring in a voice outside the tech world here, because this argument shouldn’t just come from people who are long these stocks.

Jamie Dimon released his 2025 shareholder letter on April 6, 2026. He runs the largest bank in the country — a 227-year-old institution that did $185.6 billion in revenue and $57 billion in net income last year. He has zero incentive to pump AI stocks and every reason to be measured. So when he weighs in on whether AI is a bubble, I pay attention.

His words: the investment in AI is not a speculative bubble, and it will deliver significant benefits. He went on to say that AI adoption will likely move faster than prior technology shifts like electricity or the internet, which each took decades to fully play out.

Dimon doesn’t say things like that casually. He also dropped a number that matters: hyperscaler AI CapEx hit $450 billion in 2025 and will be approximately $725 billion in 2026. That’s JPMorgan’s research team sizing the spend, and it exceeds everything I’ve documented from the individual earnings reports above.

Hyperscaler AI CapEx — Jamie Dimon’s Estimates (2025 JPMorgan Shareholder Letter)

He’s not blindly bullish, either. He flagged that the landscape will shift fast — changing assumptions around power consumption, costs, and chip technology. He said no single AI model will dominate, warned about deepfakes and cybersecurity threats, and acknowledged the possibility that AI deployment could outrun workforce retraining. All fair points.

But here’s the part that matters most: Dimon listed the hyperscaler CapEx buildout as an economic tailwind for 2026, right alongside fiscal stimulus and deregulation. He didn’t frame it as a risk factor. He framed it as a growth driver. When the guy running JPMorgan Chase and its $4 trillion balance sheet categorizes $725 billion in AI infrastructure spending as a positive force in the economy rather than a warning sign, that should tell you something about where the smart money stands on this.

The Valuations Don’t Scream Bubble Either

If AI were really in bubble territory, you’d expect the stocks at the center of it to be trading at absurd multiples — the kind that require everything to go perfectly for a decade. That’s not what I see when I pull up the numbers.

Forward EPS for Seeking Alpha earnings page, consensus EPS

When I look at the forward EPS this is not the story of a bubble. For the current year ORCL, NVDA, META, and MSFT trade below 25 times earnings. Based on next years projections ORCL, NVDA, META, and MSFT trade below 20 times earnings while AMZN and GOOGL trade at below 25 times. Looking out to the 2028 fiscal year ORCL and NVDA trade below 15 times while META, MSFT and AMZN trade below 20 times.

Think about that for a minute. NVDA is projected to grow their EPS by 60% over the next two fiscal years and trades at under 15 times 2028 earnings. On top of that The Street is expecting ORCL, META, MSFT, AMZN, and GOOGL to all grow their earnings by 30% or more over the same period and they all trade at under 21 times 2028 earnings.

When I look at Walmart (WMT) and Costco (COST), they trade at 43.58 times and 50.36 times 2026 earnings which is significantly larger multiples than the companies building and supplying the infrastructure for the biggest technology shift of our lifetimes. If these companies were trading at multiples that were 2-3x grocery stores and warehouse clubs the bears may be on to something but they’re not. The companies building out AI infrastructure are nowhere near bubble territory and if anything, the AI leaders look like relative bargains compared to the consumer staples names that everyone considers “safe.”

Here is what could be the most interesting aspect. Palantir (PLTR) which just became profitable in 2023 trades at 53.73 times 2028 earnings with 98.48% EPS growth over the next two years. This means that they are basically growing into the multiple that COST is trading at over the next several years which should be looked at as another reason that the current valuations do not support the idea that AI is a bubble.

I’m Not Going to Pretend There Aren’t Risks

AI is not risk-free, and I’m not here to tell you it is. Here is what could go wrong and it would be irresponsible if I didn’t outline this.

First, the CapEx cycle could overshoot. It has happened before. Cisco in 2000. Telecom fiber in 2001. Companies spending $100B+ per year on infrastructure are making bets on demand curves that stretch out 5 to 10 years. If enterprise AI adoption hits a speed bump, if ROI takes longer to materialize than the CFOs of these companies are projecting, or if a recession forces customers to delay cloud commitments, the market will punish these stocks viciously regardless of how strong the underlying business is. Wall Street does not care about your 5-year thesis when the next quarter misses.

Second, not everything touching AI deserves the multiple it’s trading at. There are small-cap names out there with no revenue, no product, and a pitch deck that says “AI” on every slide trading at absurd valuations. When sentiment turns — and it always turns — those names go to zero while the companies with actual cash flow survive. The gap between the real AI winners and the pretenders is going to get very ugly for people who can’t tell the difference.

Third, power infrastructure is a genuine bottleneck and I don’t think enough people are taking it seriously. You can’t build data centers without electricity. Every company I just covered is talking about being supply-constrained, and the constraint isn’t chips anymore — it’s megawatts. Grid permitting takes years. Utility capacity doesn’t scale on a software timeline. If power becomes the chokepoint, it doesn’t kill the AI thesis but it absolutely slows the revenue recognition timeline, and slower growth at these valuations means multiple compression.

Fourth, regulation is a wild card. The Anthropic situation I mentioned earlier shows you that government relationships matter in this space. Export controls on chips could tighten further. Data privacy rules could limit what AI models can train on. The political environment around AI is shifting fast and nobody has a clear picture of where the rules land in 2 years.

And fifth, concentration risk. The numbers I walked through above are dominated by 6 companies. If you’re building a portfolio around the AI thesis, you need to understand that you’re making a concentrated bet on a handful of names. Diversification within the theme matters more than people think.

I’m bullish on AI infrastructure. I’ve made that clear. But being bullish doesn’t mean being blind. The investors who do well in this cycle will be the ones who understand these risks and size their positions accordingly, not the ones who pretend the risks don’t exist.

The Bottom Line

If the best argument you’ve got is “this looks like 1999,” you haven’t opened a single quarterly filing.

The Dot-Com bubble was cash-burning startups living off VC money and equity raises, building products nobody was paying for. What I just walked through is the most profitable collection of businesses in human history — generating $1.63 trillion in trailing 12-month revenue, $580 billion in operating cash flow, spending $379 billion on infrastructure, and still banking $206 billion in free cash flow. Add NVDA doing $215.9 billion in revenue with $120 billion in net income and $97 billion in free cash flow, and you’re looking at the largest infrastructure buildout in history funded organically.

Over $1.66 trillion in contracted backlog and RPO sits on the books of MSFT, AMZN, GOOGL, and ORCL. Jamie Dimon put it in writing that AI investment is not a speculative bubble and pegged 2026 hyperscaler spending at $725 billion as a tailwind for the economy.

I don’t know what else the bears need to see.

If you’re sitting on the sidelines waiting for a 2000-style crash in mega-cap tech before you get involved, you might be waiting for something that isn’t coming. The numbers are there. The demand is there. The cash flow is there. This is a different era, and the bears are wrong.

Disclosure: I hold positions in several of the securities discussed in this article. This article is for informational and educational purposes only and should not be treated as investment advice. I am not a licensed financial advisor. Every investor should conduct their own due diligence and consider their own financial situation before making any investment decisions. Past performance does not guarantee future results.