The Trillion-Dollar AI Bet Runs on a Number Nobody Has Actually Earned Yet

Amazon spent $128.3 billion building data centers in 2025 — up from $77.7 billion the year before — and called it, plainly, an investment to support AWS growth.⁶ Then it told investors it planned to spend roughly $200 billion in 2026.⁶ Alphabet doubled too, from $91.4 billion to a guided $175–185 billion.⁴ Meta went from $72.2 billion toward $115–135 billion.⁵ Four companies, one direction, the numbers climbing so fast the year's actuals make last year's guidance look quaint. This is the most-watched capital-allocation decision of the decade, and almost everyone is reading it as a story about conviction. It is really a story about a number that has not been earned yet.

The official story is that AI demand is so overwhelming the hyperscalers are racing to keep up — a build-out validated by paying customers. The truer story is quieter and harder. The spending is real, primary-source documented, and accelerating. What is not yet documented is the demand on the other side of the ledger. The bet is enormous; the revenue justifying it is, so far, mostly a forecast.

The trillion-dollar figure that doesn't exist yet

Start with the number everyone repeats: a trillion dollars of AI capex. It is real — as a projection. Add the verified 2025 actuals for the largest spenders and you get roughly $370–380 billion, not a trillion: Amazon at $128.3 billion⁶, Alphabet at $91.4 billion⁴, Meta at $72.2 billion⁵, and Microsoft guiding to about $80 billion for its fiscal year.¹ The trillion lives in the future. Goldman Sachs's baseline model projects roughly $7.6 trillion in aggregate AI capex across compute, data centers, and power from 2026 through 2031 — a forecast, stacked years deep.⁸ The distinction matters because it is the difference between a fact and a faith.

Here is the most reliable thing about this boom: the people forecasting it keep being wrong, in the same direction. Goldman's research found consensus capex estimates undershot actuals by more than 30 percentage points in both 2024 and 2025 — analysts modeled around 20% growth, and growth came in north of 50% both years.⁷ That is a striking, two-year miss, and the bulls read it as proof the boom is even bigger than skeptics think. Maybe. But a forecasting error that only ever runs one way is also the signature of a momentum that has outrun its own models — and momentum is not the same thing as demand.

The revenue that has to show up, and mostly hasn't

Capex is only half of any infrastructure bet. The other half is the cash flow that pays it back. In June 2024, Sequoia's David Cahn ran the arithmetic and got an uncomfortable answer: to justify the projected data-center build-out, the AI industry would need to generate on the order of $600 billion in annual end-user revenue — and he argued the only way to justify the full scale of the build-out by 2030 was something close to AGI.⁹ It is worth being precise about what that number is and isn't. Cahn's $600 billion was a forward stress-test — the revenue required to make the spending rational, not a reported deficit sitting on anyone's books. Many retellings have since hardened it into a current shortfall, which is an embellishment of the original framing. But the gap it points at is real, and you can see its size in the disclosed figures.

Consider Microsoft. It planned to spend $80 billion in a single fiscal year on AI-enabled data centers.¹ Its disclosed AI business, in that same window, had just surpassed an annual revenue run rate of $13 billion — growing fast, up 175% year over year, but still an order of magnitude below the spend it is meant to justify.² That is the shape of the whole sector in one company: spending denominated in tens of billions, against AI revenue denominated in low tens of billions, against a required-revenue figure denominated in hundreds of billions. The build-out is years ahead of the income statement.

When your biggest customers are also your competitors

Now the part that turns a large bet into a fragile one. Much of the demand the hyperscalers cite to justify spending comes from each other and from the model labs they fund. They sell each other cloud capacity; they invest in the labs that rent their chips; the labs' spending appears as the cloud revenue that validates more spending. It is a loop, and a loop can look like demand right up until it stops. Epoch AI, which tracks this spending closely, makes the structural point sharply: the companies attribute their capex growth to AI infrastructure on earnings calls, but none of them separately disclose how much is actually AI versus ordinary cloud.¹² So the headline figure that anchors the entire bull case — total AI capex — is partly an act of faith in corporate phrasing. The denominator is precise. The numerator is a category nobody breaks out.

Covello has been the house skeptic, and he has been half right and half wrong, which is more instructive than being either. Citing an MIT study, he noted that 95% of enterprise generative-AI pilots produced zero measurable P&L impact on $30–40 billion of corporate spending — evidence the revenue isn't materializing where you'd most expect it.¹⁰ His diagnosis: FOMO, not ROI, is driving the build-out. But his 2024 prediction — that weak hyperscaler stocks would force capex cuts — was flatly wrong. Spending accelerated instead.¹⁰ The lesson isn't that the skeptic was a fool. It's that the boom has a logic immune to the usual discipline: when everyone fears being the one company that under-built for the future, capital stops responding to ROI and starts responding to dread.

The honest case that this is fine

The strongest counter is not that the bears are wrong — it's that being early is not the same as being mistaken. Infrastructure always precedes the revenue it enables; you cannot rent out a data center you haven't built. The two-year forecasting miss could mean exactly what the bulls say: demand outrunning the models.⁷ And the most overheated framing of all — that this is the largest infrastructure bet in corporate history — doesn't survive contact with the record. On a GDP-adjusted basis it isn't even unprecedented: Goldman puts current AI capex at about 0.8% of US GDP, well below the 1.5%-plus peak of the late-1990s telecom build-out.⁸ The dollar figures are staggering; the share of the economy is not yet historically anomalous. So the fair read is that this could be a rational early bet on a genuine platform shift, financed by companies with the cash flow to absorb being wrong.

And yet. The telecom comparison cuts both ways. That build-out also looked rational at 1.5% of GDP — right up until the fiber sat dark. A bet can be smaller than a famous past mania and still be a mania. The thing that separates this from a sure thing isn't the size of the spend. It's that the revenue justifying it remains a forecast, the demand is partly circular, and the people who manage these companies have admitted, in so many words, that the incentive driving them is fear of being left behind.¹⁰

The AI capex boom is not a fraud and not obviously a folly. It is a bet placed on a number that does not yet exist — the revenue that has to arrive to make hundreds of billions in concrete and silicon pay off. The spending is documented to the dollar. The demand is documented mostly as a forecast and partly as the companies buying from one another. That is the fork: not whether AI matters, but whether the income shows up before the depreciation does. The hyperscalers have decided that being early is survivable and being late is not. They may be right. But they are not yet being paid back — they are being believed.