# Understanding AI in a Month 18: Price Is Not Cost Understanding AI in a Month — Day 18 · 2026-09-29 *Full transcript of the spoken edition. A quoted passage is its source read aloud — the source's own words, with numbers and initialisms spoken out — quoted for comment. Each one names its source, and the place in it, beneath it.* --- On the seventeenth of September, CoreWeave, one of the biggest companies that rents out AI chips, filed a prospectus to sell more of its shares. In it, the company said that since the end of June it had signed short contracts, three to six months long, at about forty million dollars a year for every megawatt of power needed to run those customers' chips. Five days later it finished borrowing four point two billion dollars. In August, its rival Nebius told shareholders that its second-quarter deals averaged more than twenty million dollars a year per megawatt in contract value, and that it expects them to pay back their cost in a year and ten months, in a forecast that counts capacity not yet built. These are the companies' own figures. Around them runs this year's argument about AI's spending: whether the hundreds of billions going into chips and data centres will ever earn back what they cost. Today: why the price of AI computing, per token or per megawatt, is not what it costs to provide, and the three numbers that sit in between. Two readings of that argument mislead. The first is that a price tells you the cost. It doesn't. Silicon Data publishes a daily index of what it costs to rent one of NVIDIA's H100 chips for an hour. Last December, by the firm's own account, H100 rentals were two dollars an hour. On the twenty-eighth of September its index for the specialist AI cloud providers read two dollars seventy-two. Same chip, nine months older. What the sellers describe is demand running ahead of supply. CoreWeave's chief executive told investors in August that its near-term capacity was effectively sold out, and that prices for older generations of chips were at or above where they were years ago. A price is what someone will pay today. A cost starts with what was paid, spread over the years since. The second reading is that the argument is about when the chips wear out. Last November the investor Michael Burry accused the biggest cloud companies of flattering their earnings by stretching the assumed life of their equipment. Microsoft and Alphabet, for instance, had stretched their servers' assumed life from four years to six in twenty twenty-two and twenty twenty-three. The replies to Burry pointed out that old chips still rent. In July he answered: > Depreciation is not a bet on when a chip stops working. > — *Michael Burry, 'Short Thoughts July 8, 2026 - NVDA, Neos, Hyperscalers, Jevons Paradox, and Compression', Cassandra Unchained (Substack), 9 July 2026, second paragraph; https://michaeljburry.substack.com/p/short-thoughts-july-8-2026-nvda-neos, read 29 September 2026* He added that he had never said the older chips would stop functioning. His argument is about how fast a chip's earning power falls, not how long the hardware runs. To follow it, you need one idea. It was worked out for electricity, a century and a quarter ago. In May eighteen ninety-eight, Samuel Insull, who ran the Chicago Edison Company, gave a lecture at Purdue University. His problem was a power station. It costs a fortune to build, it has to be big enough for the busiest hour of the year, and for much of the year much of it stands idle. Insull set out each kind of customer's average demand across the year as a share of its peak. The measure was called the load factor. An office building: under four per cent. An all-night restaurant: forty-eight. > This question of load factor is by all means the most important one in central-station economy. > — *Samuel Insull, 'The Development of the Central Station', lecture at Purdue University, 17 May 1898, printed in Central-Station Electric Service (Chicago: privately printed, 1915), page 27, section 'The Question of Load Factor'; Internet Archive item centralstationel0000samu* Selling at cost, he explained, meant charging enough to cover operating costs, repairs and renewals, general expense, and interest and depreciation. Depreciation is the purchase price spread across the years the machine is expected to earn. Interest is the price of the money that paid for it. On those terms, he calculated, the customer with the worst load factor would have to pay more than four times as much per unit as the customer with the best. Companies were already trying to raise the load factor. Some charged a lower rate outside certain hours of the day, to encourage use when the plant would otherwise sit idle. Now swap the power station for a room full of AI chips. Three numbers sit at the heart of what an hour of computing costs. The first is capital: the chips, the servers, the network, the building and its power and cooling, everything that has to be bought or built before any of it can earn. Accountants call that capital expenditure, or capex. Electricity, staff and repairs are operating expenditure, the running costs. The second is the useful life, which turns the capex into a yearly charge. Spread a purchase over six years instead of four, and each year's charge falls by a third. The third is utilisation, the modern counterpart of Insull's load factor: how much of the fleet's capacity is doing paid work over time. Halve it, and every hour of paid work has to carry twice as much of that fixed charge. The price is a separate decision, and scarcity moves it. In April the International Energy Agency reported a shortage of the high-bandwidth memory that AI chips need, which it expects to last until at least the end of twenty twenty-seven. Scarce parts raise the cost of building: Microsoft told investors in July that its spending included the impact of higher component pricing. Scarce capacity raises the price of renting. Nothing makes the two move together, which is how a price can sit far above its cost, or below it. Map the last two years onto those three numbers, starting with the useful life. On the first of January twenty twenty-five, two of the biggest buyers moved in opposite directions. Meta lengthened the assumed life of most of its servers and network equipment to five and a half years. Its annual report says that added about two point six billion dollars to its net income last year. Amazon shortened the life of some of its servers from six years to five, and gave its reason: > The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning. > — *Amazon.com, Inc., Annual Report on Form 10-K for the year ended 31 December 2025, filed 6 February 2026, Item 8, Note 1 'Description of Business, Accounting Policies, and Supplemental Disclosures', section 'Use of Estimates'; https://www.sec.gov/Archives/edgar/data/1018724/000101872426000004/amzn-20251231.htm* Amazon's annual report puts the cost at about a billion dollars of net income. Alphabet and CoreWeave use six years. Nebius moved from four to five this year. Then utilisation, and a rare look at price against cost. On the first of March twenty twenty-five, DeepSeek, the Chinese lab, published a day of its own serving statistics. During its busiest daytime hours, it said, it ran its service across all its nodes. At night, when traffic fell, it moved machines over to research and training. Valuing each chip at a rental of two dollars an hour, it put the day's cost at about eighty-seven thousand dollars. Had every token been billed at the price of its reasoning model, revenue would have been about five hundred and sixty-two thousand. Then it added: > However, our actual revenue is substantially lower for the following reasons: > — *DeepSeek, 'Day 6: One More Thing, DeepSeek-V3/R1 Inference System Overview', GitHub repository deepseek-ai/open-infra-index, 1 March 2025, section 'Statistics of DeepSeek's Online Service'; [https://github.com/deepseek-ai/open-infra-index/blob/main/202502OpenSourceWeek/day\_6\_one\_more\_thing\_deepseekV3R1\_inference\_system\_overview.md](https://github.com/deepseek-ai/open-infra-index/blob/main/202502OpenSourceWeek/day_6_one_more_thing_deepseekV3R1_inference_system_overview.md)* A cheaper model, free use on its website and app, and night-time discounts. And the cost was chip hours at that assumed rate, nothing more: no staff, no research, no training of the models. And the defence of the longer lives: nine days after Burry's charge, on NVIDIA's results call, its finance chief said, without naming him, that the A100 chips the company shipped six years ago were still running at full utilisation. In August CoreWeave said it had recently signed a contract for A100s running into twenty twenty-nine. Two ways to live with the idle hours. CoreWeave's is to sell them in advance. Its annual report describes its committed contracts as take-or-pay agreements, > requiring payment regardless of the level of utilization. > — *CoreWeave, Inc., Annual Report on Form 10-K for 2025, filed 2 March 2026, Item 8, Note 1 'Overview and Summary of Significant Accounting Policies', section 'Revenue Recognition', heading 'Committed Contracts'; https://www.sec.gov/Archives/edgar/data/1769628/000176962826000104/crwv-20251231.htm* So for the length of a contract, the risk of idle hours sits with the customer. CoreWeave keeps the cost of the money. Last year its net interest expense was one point two billion dollars, on revenue of about five. And when a contract ends, the question comes back: what will a chip that's three or four years old rent for? Nebius forecasts that its second-quarter deals pay back in a year and ten months, and IREN, another operator, forecasts about two years for its recent three-year contracts, each by its own definition. DeepSeek's is Insull's. It keeps the idle hours and tries to fill them. Its price list today has peak hours of nine to noon and two to six, Beijing time, on weekdays, except Chinese public holidays. Every other hour costs half as much. That's how I'd put Insull's point today: a lower price at quiet hours isn't necessarily a loss. It can be the cheapest way to make an expensive machine earn. One. The third-quarter results of Microsoft, Alphabet, Amazon and Meta. Last year all four reported on the twenty-ninth or thirtieth of October. Meta's range for this year's capital spending is a hundred and thirty to a hundred and forty-five billion dollars. Watch whether it moves, and whether anyone changes how long they assume a server lasts. Two. November is the next window for NVIDIA's results; last year's equivalent came on the nineteenth. Its commitments to buy supply, mostly memory and manufacturing, stood at two hundred and seventy-nine billion dollars in July. Three. Any day: the H100 rental index, at two dollars seventy-two on the twenty-eighth of September. As newer chips arrive in volume, whether older ones keep their rental price is, I think, the most direct public test of who is right about depreciation. The idea to keep is Insull's. For an expensive machine, cost is the purchase spread over its useful life, plus interest and running costs, divided by the hours it actually works. Price is what the market will pay today. And a price can cover the cost of running a machine without ever repaying the cost of buying it. So when you see a price for AI, per token or per megawatt, ask three things: how long is the hardware assumed to earn, how busy is it, and who carries the idle hours? To read more: DeepSeek's Day Six inference system overview from March twenty twenty-five. It's short, and it's a rare case of a provider putting its own cost estimate, price and utilisation side by side. --- ## Sources (21) - CoreWeave, prospectus supplement (Form 424B5) — 17 Sep 2026 - CoreWeave, Form 8-K (convertible notes) — 22 Sep 2026 - CoreWeave, Form 10-K for 2025 — 2 Mar 2026 - CoreWeave, second-quarter 2026 earnings call (corrected transcript) — 11 Aug 2026 - Nebius Group, Q2 2026 letter to shareholders; operating and financial review (Form 6-K) — 12 Aug 2026 - IREN, FY26 results release — 27 Aug 2026 - International Energy Agency, *Key Questions on Energy and AI*, executive summary — 16 Apr 2026 - David Cahn, "AI's $600B Question" (Sequoia Capital); "AI's $1.5T Question" — 20 Jun 2024; 8 Jul 2026 - Silicon Data, H100 Rental Price Index; "H100 Price Spike" — 28 Sep 2026; Jan 2026 - Epoch AI, "The plunging price of thought" — 22 Sep 2026 - Gundlach, Lynch, Mertens, Thompson, "The Price of Progress", arXiv 2511.23455 v2 — 23 Mar 2026 - Michael Burry, post on X; "Short Thoughts July 8, 2026" (Cassandra Unchained) — 10 Nov 2025; 9 Jul 2026 - NVIDIA, Q3 fiscal 2026 earnings call; Form 10-Q for the quarter to 26 Jul 2026 — 19 Nov 2025; 26 Aug 2026 - Samuel Insull, "The Development of the Central Station" (Purdue University), in *Central-Station Electric Service* — 17 May 1898 (printed 1915) - Amazon, Form 10-K for 2025 — 6 Feb 2026 - Meta Platforms, Form 10-K for 2025; Q2 2026 results release — 29 Jan 2026; 29 Jul 2026 - Alphabet, Form 10-K for 2025 — 5 Feb 2026 - Oracle, Form 10-K for fiscal 2026 — 22 Jun 2026 - Microsoft, Form 10-K for fiscal 2026; fourth-quarter earnings call — 29 Jul 2026 - DeepSeek, "Day 6: One More Thing, DeepSeek-V3/R1 Inference System Overview"; API pricing page — 1 Mar 2025; read 29 Sep 2026 - Wood Mackenzie, "Mind the gap" — 15 Oct 2025 Understanding Machine, an Ashita Orbis publication. The written edition of this episode, with its figures and its sources, is published beside it.