Key takeaways
- The four largest hyperscalers guided to roughly $700 billion of 2026 capital expenditure, most of it AI infrastructure, with quarterly spend roughly doubling year over year.
- About two thirds of that capex buys GPUs and CPUs that depreciate in three to five years, far faster than fiber or rail, so the revenue has to arrive inside that window.
- Spend now exceeds cash generation: hyperscalers are issuing debt, and Amazon's free cash flow is projected to turn negative in 2026.
- Reported profits were flattered by large non-cash gains on stakes in AI labs (Alphabet $36.8B, Amazon $16.8B, Microsoft $5.9B), a circular markup that looks like growth until it reverses.
- Read the next earnings call by the gap: AI revenue growth versus capex growth, cash profit versus markups, and whether the largest compute buyers are hitting their own targets.
Put the two numbers on the same page. The four largest hyperscalers told investors they will spend roughly $700 billion on capital expenditure in 2026, most of it on AI infrastructure. And about two thirds of that spend goes to short-lived assets, primarily GPUs, that depreciate over three to five years. Whether the buildout pays is not a question about the size of the spend. It is a question about that clock, and the clock is faster than the headlines let on.
This is the honest ledger, both columns.
The cost column is staggering and still climbing
The spend roughly doubled in twelve months. Quarter over quarter against early 2025, Meta went from $13 billion to $20 billion, Alphabet from $17 billion to $36 billion, Microsoft from $17 billion to $31 billion, and Amazon from about $25 billion to $44 billion. Microsoft set calendar-2026 capex at $190 billion, well above the $152 billion analysts expected, with $25 billion of it attributed to memory and component inflation. Amazon’s $200 billion is the largest single-company technology commitment in history. Goldman Sachs projects combined 2025 to 2027 capex around $1.15 trillion, more than double the prior three years, roughly three quarters of it tied directly to AI.
And the spend now exceeds what these companies generate in cash. They are turning to debt, with Morgan Stanley expecting hyperscaler debt issuance above $400 billion and Amazon’s free cash flow projected to turn negative in 2026. Historically cash-funded businesses are now leveraging up to keep building.
The return column is real but lagging
The other side of the ledger is not empty. Azure’s AI-related revenue is growing about 39 percent year over year, Microsoft’s AI run rate has passed $37 billion, and Google Cloud’s backlog jumped past $460 billion. Demand for the capacity is genuine, and underinvesting carries its own existential risk if AI becomes the primary computing interface.
But capex is outpacing cloud revenue growth, and the gap showed up on the tape. When a report suggested OpenAI had missed internal revenue and user targets, AI infrastructure stocks sold off on April 28, 2026, with Oracle down about 3 percent and Nvidia, Broadcom, and AMD following. The biggest single buyer of compute may be growing slower than the buildout assumes, and enterprise demand underneath is broad but shallow: adoption is wide, yet fewer than 40 percent of firms have scaled AI past pilots.
The wrinkle the earnings reports bury
Here is the part worth reading twice. Three of the four hyperscalers padded a recent quarter with large non-cash gains on their stakes in AI labs: Alphabet booked $36.8 billion, mostly a markup on Anthropic; Amazon booked $16.8 billion on Anthropic; Microsoft recognized $5.9 billion in OpenAI-related gains. The companies funding the AI labs are reporting profits inflated by the rising paper value of those same labs. That is a circle, and circles look like growth right up until the markup reverses.
The altitude shift
Take the depreciation schedule, an accounting footnote, and follow it to the income statement. A GPU bought today is largely written off within three to five years, which means the data center it sits in has to generate its return inside that window or the asset becomes a loss before it earns out. Stretch the spend across a quarter and it is a capex line. Stretch it across the depreciation life and it is a deadline. Every dollar of this year’s $700 billion is a bet that the revenue compounds faster than the silicon ages, and the accounting does not let you defer that bet the way a railroad could defer the payoff on track that lasts fifty years.
That is the real difference between this buildout and the infrastructure cycles it gets compared to. The fiber and the rail kept their value for decades. The GPU is a melting asset.
The rule for reading the next earnings call
Ignore the capex headline; it will be a bigger number than last quarter, every quarter, by design. Watch the gap instead. The buildout pays only if AI revenue compounds faster than the hardware depreciates, so track three things on each call: AI revenue growth against capex growth, how much of reported profit came from cash operations versus markups on AI-lab stakes, and whether the largest buyers are hitting their own targets. The day those three stop moving together is the day the supercycle and the bubble become distinguishable. Until then, the spend is not the story. The clock is.
The Counter Brief — one email, every Monday.
The week's AI-for-revenue moves in a 5-minute read: which tools are worth the budget and which to skip, plus what to do this week. Source-checked, no vendor decks.
Edited by Aditya Marin Gasga
Free. One click to unsubscribe.
Frequently asked questions
How much are tech companies spending on AI in 2026?
The four largest hyperscalers have guided toward roughly $700 billion in combined 2026 capital expenditure, with estimates ranging from about $600 billion to $725 billion, and around three quarters of it tied to AI infrastructure. Goldman Sachs projects combined 2025 to 2027 capex near $1.15 trillion.
Why does depreciation matter for AI capex?
About two thirds of hyperscaler capex goes to GPUs and CPUs that depreciate over three to five years, far faster than traditional long-lived infrastructure. That means the revenue has to arrive within a few years or the asset becomes a loss, which is the core difference from slower-depreciating cycles like fiber or rail.
Is the AI buildout a bubble?
It is contested. Demand signals like Azure AI growth and large cloud backlogs are real, but capex is outpacing cloud revenue, hyperscalers are funding the gap with debt, and reported profits have been flattered by paper gains on stakes in AI labs. The answer depends on whether revenue compounds faster than the hardware depreciates.
What triggered the April 2026 AI stock selloff?
A report indicating OpenAI had missed internal revenue and user targets prompted a sharp drop in AI infrastructure stocks on April 28, 2026, with Oracle falling about 3 percent and chipmakers following, reflecting concern that the largest compute buyer may be growing slower than the buildout assumes.