The AI Capex Story Is Really a Profit Story

The headline number is capex. The four largest hyperscalers are guiding to roughly $725 billion in combined 2026 capital expenditure, up about 77% from around $410 billion in 2025, with the overwhelming majority going to AI data centers, GPUs, and power. Goldman expects $5.3 trillion from the same four companies through 2030. Capex is reaching 54% of sales for Meta, 47% for Microsoft, 46% for Alphabet, and 86% for Oracle.

Against that, the revenue is real but small. Anthropic’s annualized run rate hit about $30 billion in April 2026; OpenAI sits around $24–25 billion. Even combined, the labs’ revenues remain a fraction of the infrastructure being built on their behalf. In 2024, Sequoia’s David Cahn calculated the industry needed about $600 billion in annual revenue to justify the buildout; two years later, five companies plan to spend more than that bar in a single year — on infrastructure, not revenue.

That’s the capex story, and it’s the one everyone tells. But underneath it, the real story is about profit — and about a pivot that’s already happened.

The Pivot Nobody Announced

For most of the AI boom, the money went to training. Roughly 80% of AI spending went to building large language models, with the remaining 20% to running them. Training is the one-time cost of building a model. Inference — running it to answer real user requests — is the ongoing cost, and it’s the only one attached to paying customers. Inference can account for 80% to 90% of the lifetime cost of a production AI system because it runs continuously.

In 2026, the balance flipped. Gartner found that of the $42 billion in AI-optimized cloud infrastructure spending, $23.3 billion goes to inference and $19 billion to training — the first year enterprises are spending more running AI models than building them. Deloitte estimates inference rose from half of all AI compute in 2025 to two-thirds in 2026. Lenovo now forecasts the old 80/20 split will reverse.

This matters more than the headline capex figure, because it changes the economics. Inference-time compute scaling shifts cost from a one-time capital expenditure to a recurring operational cost — spending that grows with revenue instead of running years ahead of it. That’s the healthiest possible move for companies that need profitability sooner rather than later. Competitive advantage is shifting from who can afford the biggest training run to who can serve most cheaply.

You can see it in the labs’ own numbers. Roughly 85% of Anthropic’s revenue comes from enterprise and developer customers, and by one estimate it burns about $1.8 billion a year against OpenAI’s $9 billion. The deployment-first lab is the one closer to profit.

Why the Story has to Keep Going

Here’s the uncomfortable part. Reaching those capex numbers means a big drop in free cash flow, with Amazon projected to turn negative this year. Trillions in market cap ride on the assumption that the spending converts into profit on a reasonable timeline. If it doesn’t, the correction isn’t a dip. So everyone — hyperscalers, labs, chipmakers — has an incentive to keep the narrative of imminent transformation alive until the revenue catches up. The market already flinches when the proof lags: Meta was punished around 6% when it raised capex without proportional revenue.

The safety narrative fits this suspiciously well. Economics is already pulling labs off the frontier race — trainings share of spend is shrinking while inference takes over. But no lab wants to announce that the arms race stopped paying. “We’re slowing down because this is dangerous” is a far better story than “we’re slowing down because the marginal training dollar no longer returns.” The doom narrative doesn’t cause the pivot to deployment — the numbers do. It just lets a balance-sheet decision look like principle. And it supports licensing and compute thresholds that only incumbents can afford, narrowing the field while everyone applauds the caution.

The Honest Counter-Case

The bulls have a real argument. Anthropic went from roughly $1 billion in annualized revenue in early 2025 to $30 billion fifteen months later — a pace no software company has matched. Microsoft’s AI revenue passed a $37 billion run rate, Google Cloud’s backlog jumped to over $460 billion, and Nvidia’s data center revenue hit $75.2 billion in a single quarter. Hyperscalers say they can’t build fast enough to meet demand. And Gartner is clear that training spending will keep growing in absolute terms — frontier development isn’t ending, only its share is shrinking.

So the question was never whether AI works. It clearly does, and companies are paying for it. The question is whether spending and valuations are running ahead of the revenue that can support them — and whether the pivot to deployment closes that gap before the market loses patience. My read: yes on the first, and the whole industry is quietly betting everything on the second.

Siyuan Guo has extensive experience developing software, launching two startups, in IT and non-profit software. As CEO of Bitta AI, he works to revolutionize technology, and at the same time he leads the Mentors.net AI Academy to help educators see what AI can do for teaching, giving them hands-on experience building something they can immediately use. This piece was originally published September 17, 2026 and is cross-posted here with permission. You can email Siyuan here.

Practical workshops building AI skills for educators created by educators:

-------------------

Got something that needs to be heard? We'll get it said and read on the Worthy Educator blog! Email it to walter@theworthyeducator.com