On 8 September, OpenAI announced that roughly 10,000 autonomous agents, running an unreleased internal model, had resolved the Navier–Stokes existence and smoothness problem — one of the six Millennium Prize Problems still open. Time to a proof: 88 hours. Time for a separate model to formalise it in Lean: a further 17.
The result is disputed on provenance, and barely reviewed. But notice what it actually was. Ten thousand agents for eighty-eight hours is not a training run. It is inference — on the order of 880,000 agent-hours of frontier compute spent on a single problem. That is a demand signal as much as a capability one, and inference demand is exactly what has to fill the capacity this industry is building.
Our last report, The Rental Bill, spent fifty pages on why the capital behind the AI buildout is ahead of the revenue. Everything in it still stands. But the models got dramatically better while we were all arguing about the balance sheets — and when you redo the arithmetic on the right denominator, the answer changes.
1. “AI labs can’t make money” is a measurement error
Here is the most common calculation in AI commentary. Take the labs’ revenue, divide it by all the AI computing capacity that exists, observe that the result is small, conclude the business does not work.
The error is in the denominator. A large share of that capacity is doing training, and training generates no revenue by design — it is R&D. Including it is like dividing a drug company’s revenue by its laboratories plus its factories and calling the result a margin.
Source. Holds revenue per inference GW constant at $30B, so every movement in the lighter bars is mix, not performance. Inference share of compute: Deloitte cycle-based estimates (spend-based measures run as high as 80–90%).
Inference was about a third of AI compute in 2023, half in 2025 and roughly two-thirds in 2026. Hold revenue per inference gigawatt constant and revenue per total gigawatt still climbs from $10B to $26B — purely from mix. Which cuts both ways: reported AI economics are about to look considerably better without anything fundamental changing, and anyone extrapolating that improvement is extrapolating an accounting artifact.
The same error runs through the margin line. 2026 estimates of Anthropic’s gross margin range from 44% to 60% — sixteen points that turn entirely on how training cost is treated.
2. What a gigawatt actually earns
The cost side comes from The Rental Bill: roughly $38B to build a one-gigawatt AI datacentre plus $0.9B a year to run it. Amortised over a five-year hardware life at an 8% cost of capital, that is a rent of $8.5–10.4B per gigawatt per year.
The revenue side now comes from three independent directions. A bottom-up token model — about 7,700 NVL72-class racks per gigawatt at 30,000 tokens per second — produces a theoretical ceiling near $100B a year at full utilisation and list prices. Broadcom’s Hock Tan, at Goldman Sachs Communacopia this month, offered an illustration of $30B of annual recurring revenue against about $10B of annual cost per gigawatt. And $30B is exactly where our model sits at 30% of its ceiling.
Source. Rent: Epoch AI 1 GW datacentre cost, amortised over five years at 8% ($8.5–10.4B). Hock Tan, Goldman Sachs Communacopia, Sep 2026 (Broadcom sells silicon to these labs — an interested party). Proflex bottom-up token model. Ceiling assumes full utilisation at list prices, which no operator achieves.
A caveat we would put in bold if we could: Broadcom sells silicon to these labs, so Tan is an interested party. His figures are used because they agree with a cost model and a token model built without reference to them. And nobody publishes fleet utilisation, which is the single largest unknown here. But look at how wide the margin for error is.
Source. Coverage of the full all-in cost of capacity — depreciation, cost of capital and operating expense — against a ~$9B/GW/yr rent. Fleet utilisation is published by nobody; it is the single largest unknown in this analysis.
Inference capacity needs to realise only about 10% of its theoretical ceiling to cover its entire cost — depreciation and cost of capital included, not just electricity. You can be wrong about utilisation by a factor of three and the capacity still pays for itself.
“Ten Years in Six Months” — the complete 36-page Proflex research note. The per-gigawatt model with every assumption exposed, the open-source question with production routing data, the circle problem, the macro transmission chain, the straight-line-versus-sawtooth framework and the seven-indicator watch list. Every figure sourced, including where sources conflict.
3. The labs are already demonstrating it
Source. Published estimates and company-attributed figures; 2028 is a company projection. The 2026 figure is estimated at 44–60% depending on how training cost is treated — 60% is the upper bound. Driven by inference efficiency, not price increases.
Anthropic’s gross margin moved from −94% in 2024 to roughly 60% in 2026, driven by inference efficiency rather than price increases — the more durable of the two mechanisms. It reported more than $11.5B of revenue in Q2 2026, turned adjusted operating income positive, and is tracking toward more than $1B of operating profit in Q3. OpenAI looks worse for an explicable reason: about 900 million free users drag its gross margin down by twenty to thirty points. That is a business-model choice about consumer free tiers, not a difference in the unit economics of inference.
And if capacity were heading for a glut, the price of renting it would be falling. It is doing the opposite.
Source. IREN contract pricing per MW of IT load; company statements and reporting, Sep 2026. These appear to be total contract value per MW, not annual (a $25M/MW three-year deal is ~$8.3M/MW/yr). The level is ambiguous; the direction is not.
IREN priced a five-year Microsoft contract at $9.70M per megawatt in November 2025, NVIDIA at $11.33M in May, three-year deals above $20M in August — and about $25M is now showing up in live conversations. Rising prices on multi-year contracts are what scarcity looks like, not the early stage of an overcapacity cycle.
4. Open weights won the tokens. Frontier labs kept the dollars.
Source. Vercel AI Gateway data via OpenRouter analysis and industry reporting, 2026. Open weights run at roughly a tenth of the platform’s average token price; in June the four leading US frontier labs took 95% of the money. What customers spend is not what they earn.
Open-weight models went from 11% of tokens on Vercel’s AI Gateway in April to a record 62% by 22 August, while taking under 4% of the spending. That confirms the structure Tan described. It does not confirm his inference — a tenfold price gap fully explains that spending split without any difference in delivered value. What enterprises spend is not what they earn, and we could find no study measuring enterprise ROI by model class. For the investment case the point is simpler: a gigawatt serving open models is still a gigawatt earning its rent.
5. The contradiction in the extinction week
In the same week Dario Amodei published “We Must Pace the Frontier” — asking Washington for a narrow antitrust waiver so rival labs could coordinate on safety — The Information reported what Anthropic has been contracting for.
Source. The Information via Forkast, TechCrunch and secondary reporting, Aug–Sep 2026. $275B of named deals is the firm floor; $517B is the reported ceiling and may include options and letters of intent. At $8.5–10.4B/GW/yr, 14.8 GW implies $126–154B a year of eventual rent — Proflex arithmetic, not a disclosed figure.
Up to $517B of compute and at least 14.8 gigawatts in eleven months, with an IPO prospectus expected within weeks. Apply the rent arithmetic and that capacity implies $126–154B a year of eventual rent against a tracked run-rate of about $74B. Our reading: the risk concerns are not insincere, but a pace limit plus an industry-funded pre-release testing body is, operationally, a compliance regime — and compliance is a fixed cost that favours the incumbents who can already afford it. That is ordinary American regulatory capture, not a scandal. It does not change the capacity forecast. It does raise the political risk premium on the sector.
6. The circle problem: the heaviest AI user is AI
The deepest, best-measured penetration of these tools is in software engineering — where junior developer demand is down roughly 40% and Gartner expects 75% of developers to orchestrate rather than write code by year end. Software engineering is also, to a large degree, the activity of the technology sector itself, including building AI. Capability compounds inside the loop. The economy-wide effect does not, because the benefit has not yet left it.
Source. Deloitte survey of 3,235 leaders; KPMG pulse of 2,145 leaders across 20 countries, June 2026; enterprise AI production survey, August 2026. Seven per cent is not a verdict on the technology — it measures how far diffusion has travelled.
Three-quarters of enterprises now run AI in production. Only 7% report established, measured ROI, and only 6% can attribute 5% or more of operating profit to it. AI demonstrably compresses discovery — the search across chemical space, the hunt for a counterexample, the first draft of a system. It has not yet shown that it compresses validation — the Phase III trial, the regulated deployment, the organisational confidence to act. That is where most of the cost and nearly all of the time sits.
7. Two per cent of everything
Source. Different bases, so read each bar on its own: BEA-derived analysis (private fixed investment, Q2 2026); Deloitte (tech investment share of real GDP growth since 2023); analyst estimates via Forbes, June 2026 (share of GDP). Estimates of AI’s contribution to GDP growth range from ~20% to 65% and do not reconcile.
The US is on track to spend about 2% of GDP on AI and datacentre infrastructure this year. In Q2 2026, AI accounted for roughly 69% of the change in real private fixed investment. Hyperscaler AI capex has gone from about $235B in 2024 to over $700B projected for 2026. Whatever happens to AI capital spending now happens to the investment component of GDP more or less directly — with AI-linked debt heading toward $570B at the margin.
The mitigant is real: this buildout is financed mostly from internally generated cash flow. Amazon, Microsoft, Alphabet and Meta can absorb a disappointing few years. WorldCom could not. The exposure sits in the marginal financing — the SPVs, neocloud debt and vendor guarantees — and that is where a repricing would start.
Straight line or sawtooth
Both paths end in the same place. On the straight line, adoption broadens into ordinary industries roughly on the schedule the financing assumes, and 2026 capacity looks like one of the best capital allocations of the decade. On the sawtooth, capex expectations break before adoption matures, the unwind takes a chunk of GDP with it — and the capacity that would have produced the abundant economy stops being built. Interconnection queues run four to seven years; large transformers have 128-week lead times. A project that loses its financing in 2028 does not resume in 2029.
The technology does not need more capability to justify the spending. It needs more customers outside its own industry, sooner than the debt schedule implies. The technology is no longer the variable. The path is.
Seven indicators that tell the straight line from the sawtooth — enterprise measured ROI, inference share and utilisation, neocloud contract pricing, Anthropic’s operating profit, non-tech adoption, AI-linked credit spreads and AI’s share of private fixed investment — with the specific reading that would change our mind in each direction.
Figures are from the Proflex Finance research note “Ten Years in Six Months” (10 September 2026), which cites every source and shows its arithmetic. Neither OpenAI nor Anthropic publishes audited accounts; fleet utilisation is published by nobody. Figures marked as estimates or model outputs are Proflex arithmetic, not disclosed numbers. This field moves faster than any document — verify anything you intend to act on against current sources.