Analysis
Who the AI build pays
We built a model of the AI build-out and ran it through 22 worlds and 3,000 random ones. The chip makers get paid in almost all of them and the labs in most. The owners of the data centres earn their cost of capital in two named worlds, and both are shortages.
A model, not a forecast
The forecast page is a model of the AI build-out from 2026 to 2035, and it runs in your browser. Demand is the world’s knowledge work, counted in the tokens an AI would need to do it. Supply is the fleet of AI chips, built year by year and worn out over time, held back by memory, packaging, power and credit. Between them sit the labs, which buy compute from the fleet and sell work to everyone else.
Each year the model asks one question of every layer: who got paid, and did it cover its cost of capital? The 2026 starting point is set against the filings: top-five cloud capex of $182B in the June quarter, against $97B a year before; 97% of their operating cash spent; debt of $487B, double a year earlier. This piece is what version 1 says. It is a simulation, not advice, and you can change any input on the page.
The chip makers first, the labs later
In the soft-landing world, buyers pay $22.3T for AI work over the decade. The labs keep $10.9T after paying rent. Chip and system makers take $9.5T in revenue. The data centres’ owners — the hyperscalers, Oracle and the neoclouds, who own the chips and rent out their time — fall $4.4T short of their cost of capital.
The order matters. The chip and memory makers are paid now, as the filings show: Micron’s revenue was $54B in the September quarter against $11B a year before. The labs’ margin starts near zero and passes the chip makers’ revenue in 2032. The owners in the middle start underwater, earning $175B in 2026 against $410B needed to cover the cost of their fleet, and never catch up.
Across all 22 named worlds the pattern holds.
The two worlds where the owners win are shortages: agents that eat compute faster than anyone can build it, and a hardware supply chain that cannot keep up. A power wall, a third kind of shortage, comes within $0.7T of breaking even. Holding back on building helps; it never fixes it.
A race the owners need demand to win
Each year software gets more out of every chip: better models, smaller models, smarter serving. That is good for whoever sells the work and bad for whoever owns the chips, because the same fleet can now serve more demand than it did, and a fleet with slack cannot charge much for its time.
So the model’s crux is one rate. We took the agents world, let it build 70% a year, and changed only the software gain.
At 30% a year the owners come out $31T ahead. At 50% they make $10T. Past about 62% they fall behind, and at 100% they are $8T short while the labs keep $57T. The same test in the other two shortage worlds breaks even at 62–63%.
A 62% gain in output per chip takes about 38% a year off the cost of each task. The record suggests prices for cheap models are falling faster than that, while frontier prices hold. OpenAI cut GPT-5.6 Luna by about 80% in July, and GPT-6 Luna came in about 50% cheaper again in September. An index of average token prices fell from $2.07 per million in May to 97 cents in September. Yet OpenAI priced its top model, GPT-6 Astra, at $10 and $50 per million tokens, matching Anthropic’s best. And DeepSeek raised its prices at peak hours in August, a sign of shortage.
That is the model’s two-tier market: frontier work priced on its value, everything else priced near cost. The record is paying attention. In the TEXXR corpus, articles about falling token prices and cheaper models numbered 22 in the third quarter of 2025 and 123 in the third quarter of 2026, a 5.6-fold rise; 410 have run since the start of 2024.
The order book
The owners have one defence: contracts. Microsoft, Oracle, Google and CoreWeave report $1.9T of remaining performance obligations, three times the figure a year earlier.
Not all of that is AI, and not all of it is owed by labs; the filings do not split it by customer. OpenAI and Anthropic together earn about $105B a year, by Epoch AI’s estimate. Paid over five years, $1.9T is nearly $400B a year. If the labs spend four-fifths of their revenue on compute, they need close to $500B a year of revenue, almost five times today’s, or investors to make up the difference. A contract protects the owner only as far as the buyer can pay it.
Three thousand worlds
The named worlds are stories. To test the result without them, the model drew 17 inputs at random over stated ranges, 3,000 times.
The owners earn their cost of capital in 8% of them. The labs keep a positive margin in 99%.
What would change the answer
- GPU rents. One-year H100 contracts rose from $1.70 an hour to $2.35 between October and March. Two quarters of falling rents would mean the shortage is over.
- Frontier prices. If the price of a frontier task starts falling more than about 38% a year, software is winning the race.
- Lab revenue. About $500B a year by 2028 would carry the order book. That is the agents world, where the owners come out ahead.
- Memory prices. DRAM contract prices rose 95%, 60% and 15% in the first three quarters of 2026. A quarter of falling prices would mark the top for the makers paid first.
- Credit. Cloud capex above operating cash for consecutive quarters while debt keeps rising.
The limits
The owners’ return includes the fleet that already exists, which starts underwater. The software gain is a fleet-wide average, so price cuts on small models overstate it. The frontier price, the frontier share of work and buyers’ 2026 spending of $150B are estimates. The model has no stock market in it, so it says nothing about any share price. Every input is on the page to change.
Sources
Filed figures: SEC EDGAR 10-Q and 10-K filings (capex, operating cash flow, debt, remaining performance obligations for Microsoft, Oracle, Alphabet, Amazon, Meta and CoreWeave; Micron and Nvidia revenue). Lab revenue: Epoch AI, August 2026. GPU rents: SemiAnalysis one-year H100 contracts; Ornn H100 index. DRAM contract prices: TrendForce. The model: static/js/forecast_core.mjs, and the experiments behind this piece, scripts/forecast_thesis.mjs and scripts/forecast_figures.mjs, run on 2026-10-15.