← All articles
AI economics

Where the money in AI actually goes

My ChatGPT subscription costs twenty dollars a month. About the price of lunch for two. Behind that lunch sits the biggest construction project I can think of, and I wanted to know who is actually paying for it.

Start with the bill. Amazon, Alphabet, Microsoft and Meta spent $228 billion on capex in 2024 and $376 billion in 2025. For 2026 they plan $720 to 740 billion. Nearly three times as much in two years. Each company counts capex a little differently, so take the older years as approximate. The third quarter of 2026 alone came to $207 billion.

3×Big Tech capex in 2026 vs. 2024Amazon, Alphabet, Microsoft, Meta$228bn2024$376bn2025$730bn2026plan $720–740bn
Capex of Amazon, Alphabet, Microsoft and Meta, USD. 2024–25: cash spending; 2026: company guidance, which for Microsoft and Meta includes finance leases. Sources: Trimline Research, Supercycle (Oct 2026).

Even for these companies it hurts. Alphabet and Meta had negative free cash flow in Q3. And it isn't mostly buildings. Microsoft says about three quarters of its capex goes into short-lived kit like GPUs and CPUs. Electricity, which everyone worries about, is a smaller line than you'd expect: Epoch AI puts energy at 2 to 6% of the cost of developing a model. The money goes into hardware and people.

Most of the compute now goes into answering us

You can burn compute in two ways. Training builds a model. Inference runs it, every time someone types a question or an agent does a task. In 2023 inference was about a third of all AI compute. This year Deloitte puts it at roughly two thirds.

2/3of all AI compute now goesinto running models (2026)2023~1/32025~1/22026~2/3Inference (running models)Training and research
Share of all AI compute. Source: Deloitte TMT Predictions 2026.

OpenAI sat right on the tipping point in 2025: $8.0 billion on inference, $8.3 billion on R&D compute. Now the part I liked best. In 2024 the final training run was only 9.6% of OpenAI's R&D compute. The rest went on experiments. If you've ever launched a product, you know that ratio. The first good serial part is cheap. The prototypes, trials and rework before it are not.

So who makes money?

In a gold rush, sell shovels. Nvidia took in $96.2 billion in one quarter (to 26 July 2026), $89 billion of it from data centres, at a 75% gross margin.

The labs live in a different world. OpenAI's leaked audited statements show $13.07 billion of revenue for 2025 and an operating loss of $20.92 billion. The FT reported it expects to burn $278 billion of cash by 2030. Sounds like a disaster. Look closer, though. In the leaked accounts, cost of revenue fell from 72% to 57% of revenue (my calculation). Internal figures reported by The Information say the opposite: gross margin slipped from 40% to 33%, because demand forced OpenAI to buy pricier compute at short notice. Either way, serving customers still brings in more than it costs. Most of the hole is dug by research, training and marketing.

The best look under the hood comes from Anthropic, because an IPO forces you to open the books. In the first nine months of 2026 it took in $33.24 billion. Serving paying customers cost $4.81 billion. All compute together, including training, research, internal use and free users, came to $18.81 billion.

43%of revenue left after all computeAnthropic, Jan–Sep 2026 (my calculation)Revenue9 months 2026$33.2bnAll computetraining, research, serving, free−$18.8bnLeft after computemy calculation$14.4bnOther costspeople, sales, admin−$15.7bnGAAP operating result−$1.26bnDarker bars are costs. USD.
Anthropic, January to September 2026, USD. Source: S-1/A (2 Nov 2026). The 43% and "other costs" are my own calculation.

After every bit of compute, about 43% of revenue is left (my calculation from the filing). Other costs turn that into a GAAP operating loss of $1.26 billion. The $4.08 billion profit you may have seen is the company's adjusted figure, without stock pay and a few other items. So to be clear: under standard accounting, Anthropic still loses money. The third quarter slipped back into an operating loss of $567 million, mainly because it had to buy $5 billion of pricier short-term compute. And the net loss for the nine months was about $104 billion, mostly a paper charge on convertible notes that grew in value with the company. Then there's the elephant in the server room: compute commitments of roughly $900 billion, about $70 billion of it due in the next twelve months. None of that is in today's numbers. Chinese labs, by the way, follow the same pattern. Z.ai already earns a 24.6% gross margin on its API and still loses money, because research eats all of it.

The one number I'd watch

How long a chip is allowed to live on paper decides what it costs per year. Most big players depreciate servers over five to six years. Alphabet stretched from four to six in 2023 and added $3 billion to net income. Amazon went the other way, cut part of its servers to five years because of the pace of AI, and booked $1.4 billion of extra depreciation.

5–6 yrshow long servers live on paperAmazon is the one going shorterAlphabet4→6 in 20236 yrsOracle6 yrsCoreWeave5→6 in 20236 yrsMicrosoftrange 2–62–6 yrsMetaextended 20255.5 yrsAmazonpart 6→5 in 20255 yrs
Accounting useful life of servers and network equipment (GPUs are usually not reported separately). Sources: 10-K filings via Exent, DECUS, CNBC.

Anyone who has run a plant knows the gap between book life and real life. I've seen fully written-off machines that were still the best in the hall, and machines still on the books that nobody wanted to touch. Chips are the same story at ten times the speed. If a server written off over six years really lasts three, its yearly depreciation doubles (simple straight-line math, mine). OpenAI and Anthropic mostly rent their chips, so the hit would land first on the cloud companies that own them, and sooner or later in the rent.

Serving paying customers already pays its way. Most of the losses come from the race for the next model.

None of the big labs is profitable yet. For those of us who use AI, that's still good news. The most expensive part, the race, is paid for by investors and Big Tech balance sheets. We get the result as one line on a monthly invoice.

Hundreds of billions are going into concrete and silicon so I can rent a genius for the price of lunch. What nobody will build for me is the process that puts that genius to work. That part is still our job. Turning plans into parts.

Sources: Epoch AI, Deloitte, company annual reports and regulatory filings (Anthropic, Nvidia, Microsoft and others), and the business and tech press (CNBC, Reuters, Ars Technica, Tom’s Hardware). My own calculations are marked in the text. Figures as of 14 November 2026.

Building AI into a real operation, not just a slide? Let's talk.

Get in touch LinkedIn