The Earnings Call Tells You Which Hyperscaler Is Faking the CapEx ROI

Every hyperscaler is spending like the future depends on it, because they think it does. The four biggest US names, Amazon, Google, Meta, and Microsoft, guided to roughly $725 billion of combined capital expenditure in 2026.^[1]That is up about 77% from around $410 billion in 2025. Numbers that size stop being a line item and start being a macro event. Nvidia's revenue, national power grids, and a big slice of the S&P 500's forward earnings all hang off that spending continuing.

Here is the thing nobody says on CNBC. Spending is not the signal. Everybody is spending. The signal is what shows up on the other sideof the spend, and whether management will look you in the eye and quantify it. That is what the earnings call is for. Not the beat-or-miss theater. The moment an analyst asks “what are you getting back,” and you listen to whether the answer is a number or a dodge.

The tell: does the number exist, or doesn't it

Start with the cleanest contrast the 2026 earnings season handed us.

Microsoft, in fiscal Q3 2026, said its AI business surpassed a $37 billion annual revenue run rate, up 123% year over year.^[2] Capex that quarter was $30.88 billion, up 84%. Azure grew 40% in constant currency. Whatever you think of the valuation, that is a company willing to draw the line from dollars in to dollars out and let you do the arithmetic. Total revenue rose 18% to $82.9 billion and EPS jumped 23% to $4.27, both ahead of consensus.^[2] When a company is confident in the return, it hands you the numerator.

Now Meta. On the Q1 2026 call, an analyst pressed Mark Zuckerberg for signs of return on the AI investment. His answer: it was “a very technical question.”^[4] No run rate. No revenue-per-GPU. No cohort. And this is against a 2026 capex guide Meta just raised to $125 to $145 billion, up from an initial $115 to $135 billion, and nearly double the $72.2 billion it spent in 2025.^[5] The reason given for the raise was higher component pricing and additional data center costs. Fine. But you are asking me to fund a doubling of spend and answering the ROI question with a shrug dressed up as engineering humility.

To be fair to Meta, its AI spend partly feeds ads, and ad lift is genuinely harder to attribute than a cloud invoice. But that is exactly the point. The businesses that can cleanly attribute the return (Microsoft, Amazon's AWS, Google Cloud) are the ones showing it. The ones that can't are the ones asking for patience. As an investor, know which one you own.

Backlog is the honest version of a promise

The next place to look is backlog, because backlog is a contract, not a vibe. Google Cloud revenue accelerated 63% year over year to about $20 billion in Q1 2026, and the backlog nearly doubled sequentially to roughly $462 billion, with management saying over 50% is expected to convert to revenue within 24 months.^[3]That last clause is the whole ballgame. A backlog with a stated conversion window is a testable claim. In two years you can check it. That is the opposite of “a very technical question.”

Amazon told a similar story from a different seat. Q1 2026 capex was $44.2 billion, up 77%, but AWS grew 28%, its fastest in fifteen quarters, at a 37.7% operating margin.^[6]Margin like that under a capex ramp like that means the spend is landing on demand that already exists. And the sleeper line: Amazon's own chip business, Graviton, Trainium, and Nitro, exceeded a $20 billion annual revenue run rate, with over $225 billion in Trainium commitments.^[6] That is Amazon quietly building a moat that also lowers its own cost per unit of compute. Vertical integration paying rent.

Then there is Oracle, which is the cautionary version of the same chart. Its remaining performance obligations backlog hit about $523 billion in Q2 FY2026, up 438% year over year, and later grew toward $553 to $638 billion.^[9] Enormous. Except roughly $300 billion of it is tied to a single customer, OpenAI, through Project Stargate. And Oracle is funding the build with negative free cash flow of about $23.7 billion, with capex near 80% of revenue.^[9] A half-trillion backlog reads great in a headline. A half-trillion backlog where the largest chunk depends on one counterparty that is itself burning cash, financed by your own negative free cash flow, is a different animal. Same word, opposite risk.

The depreciation game that flatters every income statement

Now the part that should make you slow down, because it applies to all of them and it is invisible unless you go looking.

The four hyperscalers purchased $433.9 billion of property and equipment in the four quarters through March 2026. Over the same span they reported only about $149 billion of depreciation.^[5] So the income statement is expensing a fraction of the actual build-out. The cash left months ago. The cost hits earnings in slow motion, spread across the assumed useful life of the gear.

And that useful life keeps getting longer at exactly the convenient moment. Microsoft stretched server life to 6 years in FY2022. Google went to 6 years in January 2023. Amazon went to 6, then trimmed back to 5 for a subset. Meta moved to 5.5 years in January 2025.^[7] Collectively that stretching cut an estimated $18 billion in annual depreciation expense.^[5]Every year you add to the assumed life, you shrink this year's expense and lift this year's profit, without a single new dollar of revenue.

Michael Burry, the “Big Short” guy, ran the math and it is ugly. He estimates depreciation will be understated by about $176 billion between 2026 and 2028, which would cause hyperscalers to overstate profits by more than 20% if the true economic life of AI servers is closer to three years.^[8] That last condition is the crux. These useful-life schedules were built for general-purpose servers that age gracefully. GPU-dense AI training gear is a different beast. A generation-old accelerator can be economically obsolete long before it is physically broken, because the newer part does the same work for a fraction of the power. If Burry is even directionally right, a chunk of reported hyperscaler profit is a timing illusion.

I am not saying it is fraud. It is defensible under the accounting rules and every company discloses the assumption. I am saying that when you compare a hyperscaler's reported net income to its actual cash going out the door, you are comparing two very different pictures, and the gap is now measured in hundreds of billions.

The crossover nobody wants to talk about

Which brings us to the single most important chart in this whole story. For years, hyperscalers funded their build-outs out of pocket. Operating cash flow comfortably covered capex, so the spending was self-financing and basically risk-free to the balance sheet. That era is ending on a schedule.

Aggregate hyperscaler cash capex is on track to overtake operating cash flow around Q3 2026.^[10] Operating cash flow is growing about 23% a year. Cash capex is growing about 70%. Two lines at those slopes cross, and then keep diverging. Oracle has already crossed. Amazon is crossing now. Alphabet lands around Q1 2027, Meta around Q3 2027, and Microsoft, the strongest of the bunch, around Q3 2028.^[10]

Past the crossover, the spending has to be funded some other way. Debt, equity, leases, off-balance-sheet special-purpose vehicles. That is not automatically bad. Cheap debt against productive assets is how you build railroads and fiber networks. But it changes the risk profile completely. A self-funded build can slow down for free the moment demand wobbles. A debt-funded build has coupons due whether or not the GPUs are earning. The whole thesis quietly shifts from “these companies are printing cash” to “these companies are financing a bet.”

The 2001 rhyme, and why it is worse this time

If your pattern-matcher is firing, good. The divergence between AI capital spending growth and revenue growth was running near 46% in 2026.^[11] During the 2001 telecom overbuild, the cycle that gave us dark fiber and a lot of bankruptcies, that same divergence peaked around 32%. We are already past it. The estimated annual gap between AI infrastructure spend and actual AI ecosystem revenue is roughly $600 billion.^[11]

Here is the honest counterpoint, because I am not trying to sell you a crash. The telecom guys built capacity for demand that took a decade to show up, and most of them were pure infrastructure plays with no cash cow underneath. The hyperscalers are different. They have enormous, real, profitable core businesses (search, ads, Windows, Prime, AWS) throwing off cash to absorb the mistake. Microsoft's $37 billion AI run rate is not dark fiber. It is booked revenue growing 123%. The demand is partly here, not purely hoped for.

But the fiber overhang still took years to clear and vaporized a lot of equity on the way. “The demand is real” and “the spending is ahead of the demand by $600 billion a year” are both true at once. That is the uncomfortable place we are actually in.

How I actually read a hyperscaler earnings call now

Strip away the beat-or-miss noise. Four questions, in order, and the call answers all of them if you listen for it.

1. Is there a disclosed AI revenue number? A run rate, a segment, a growth rate tied to the spend. Microsoft gives one. Meta, asked directly, would not. Absence is data.
2. What is the backlog, and from whom?Google's $462 billion with a 24-month conversion window is a promise you can audit. Oracle's $523 billion with $300 billion from one cash-burning customer is a concentrated bet. Same word, different animal.
3. What is the gap between depreciation and capex? If capex is three times depreciation, current earnings are flattered and a future writedown is loading. Watch for useful-life extensions buried in the 10-K. They are free earnings and they always show up right when the spend accelerates.
4. Has capex crossed operating cash flow? Before the crossover, the build is self-funded and can stop for free. After it, the build is financed and has to keep earning to service itself. Oracle is past it. Microsoft has two more years of runway. That distance is the margin of safety.

Everyone at this table is spending three-quarters of a trillion dollars. The market is treating that spend as if it were the achievement. It isn't. The achievement is the revenue on the far side of it, the honesty about how fast the gear actually wears out, and the cash flow left to fund the next round without borrowing. Some of these companies show you all three. At least one answers the most important question by telling you it is too technical to answer. On an earnings call, that is not humility. That is the tell.