Nvidia H100 Rent Up 44% as Anthropic Token Prices Halve

Renting an Nvidia H100 from a neo-cloud provider cost $2.88 per GPU-hour on Silicon Data’s index on 24 October 2026, against $2.00 on 9 December 2025: a rise of 44% by our arithmetic, as our H100 piece reported. Over a similar stretch, the data firm YipitData wrote that “across our token-based datasets, we’ve observed a ~50% Y/Y decline in token prices,” in a note on Anthropic dated 22 to 23 October, as our token-price piece reported.

The two figures measure different things over different windows, so this piece does not combine them. It sets them side by side with the other costs and prices from our October reporting, and asks who, on the evidence we have read, is absorbing the gap.

This Week In The AI Supercycle · VIDEO

The Business Engineer’s explainer for this week’s issue, 18 to 24 October 2026: the companies selling AI’s parts say they are booked for years, while the companies paying for the build found money dearer. It does not discuss the figures in this piece.

What Got Dearer

Silicon Data’s published H100 readings went from $2.00 on 9 December 2025 to $2.20 on 6 January, $2.47 on 8 April, $2.76 on 30 May, $2.75 on 27 July and $2.88 on 24 October, as our H100 piece reported. Its August post said the H200 Neo-Cloud rate went from $2.70 at the start of May to $3.09 on 27 July, a 14.4% rise over twelve weeks.

Supply ahead is also spoken for. Amazon chief executive Andy Jassy said of AWS capacity, as Reuters reported on 23 October: “2027 is largely sold out.” Micron said its 26 strategic customer agreements cover DRAM and NAND through 2030, and Lumentum’s chief executive said the optics maker is “almost sold out for 2029,” as our sellout piece reported.

What Got Cheaper

YipitData’s note lists what it says contributed to the ~50% decline: “Price cuts from Opus 4.5, Sonnet 5’s introductory pricing that became permanent, greater cache hit rates, and Fable 5.1 cache hit pricing reductions have contributed.” Ramp’s AI Index shows the blended price per million tokens falling 25.4% between the weeks ending 13 September and 11 October, as our Ramp piece reported.

Price is also being used to win customers. At Rippling, Anthropic models handled 57% of the coding work of hundreds of engineers in May and OpenAI models 16%; now OpenAI handles 52% and Anthropic 18%, as the company sought to cut costs and OpenAI proved cheaper, chief executive Parker Conrad told The Information, as our Rippling piece reported. Techmeme’s summary of the same article says Replit moved most of its AI spend to OpenAI, which gave more than 30% off in a deal worth more than $100 million a year.

Not every price is falling. Google’s Gemini API pricing page, last updated on 22 October 2026, lists a dated step-up for seven of its 32 model listings, each of them doubling the price, as our token price paradox piece reported.

Sources: Silicon Data, as reported in our H100 piece; Reuters, as reported in our sellout piece; YipitData, as reported in our token-price piece; Ramp, as reported in our Ramp piece; The Information via Techmeme, as reported in our Rippling piece. The rows pair figures with different windows and data; they are set side by side, not combined.

Who Absorbs the Gap

None of our sources measures who is absorbing the gap. They do describe where strain could show, in three places. The model makers: The Information’s article is headlined “Anthropic’s Margin is OpenAI’s Opportunity,” and an X post by @choblin29 summarising it says OpenAI previously projected it would spend more on compute this year than it expects to generate in revenue; we have not checked that line.

The buyers: Ramp’s 14 October letter recorded “two straight months of virtually no growth in token spend” and says “80% of spend on OpenAI and Anthropic is driven by 1% of firms,” as our Ramp piece reported.

The lenders and shareholders, in a scenario rather than in losses happening now: TS Lombard’s crash scenario of 22 October puts credit losses on AI-related debt at $125 billion to $190 billion, against about $20 trillion of losses for equity holders. It says the strain is “going to appear upstream,” among neoclouds, developers and the standalone model developers, as our TS Lombard piece reported.

Connecting the Coverage

Our H100 piece of 24 October traced the rent from $2.00 to $2.88. Our sellout piece of the same day ranked how far ahead Micron, Lumentum, AWS and Sandisk say they are booked. Our token-price piece of 24 October reported YipitData’s ~50% decline, and our token price paradox piece set it against seven Gemini prices that are set to double. Our Rippling piece reported the share swing at Rippling and the Replit discount. Our Ramp piece of 23 October reported the concentration of spend, and our TS Lombard piece reported the crash scenario.

What this piece adds is the side-by-side view: on the figures we have read, the rent on a 2022 chip rose over a stretch in which YipitData measured lower token prices, and OpenAI cut prices to win coding customers, The Information reported.

The Structural Read

As we read it, the rent rise was gradual, not one jump: Silicon Data’s published readings show $2.20 in January, $2.47 in April and $2.76 at the end of May before $2.88 in October.

As we read it, price cuts are being used to win customers: Rippling’s chief executive told The Information that OpenAI proved cheaper, and Techmeme’s summary says Replit got more than 30% off.

As we read it, the price picture is not one-directional: Google lists seven Gemini prices that are set to double, against YipitData’s ~50% decline across its panels.

Three Implications

FOR RENTERS OF COMPUTE Silicon Data’s H100 index read $2.88 per GPU-hour on 24 October, up from $2.00 on 9 December 2025; 44% is our arithmetic.

FOR MODEL MAKERS The Information’s article is headlined “Anthropic’s Margin is OpenAI’s Opportunity,” and Rippling’s chief executive said OpenAI proved cheaper.

FOR LENDERS TS Lombard’s crash scenario puts credit losses on AI-related debt at $125 billion to $190 billion, against about $20 trillion for equity holders.

The Business Engineer Lens

This story maps onto the Business Engineer framework Tokenomics: The Economics of AI.

The framework puts it this way: the token “is simultaneously the unit of cognition the model produces, the unit of compute the data center serves, the unit of price the lab charges, and the unit of value the enterprise extracts.”

As we read it, this week’s figures sit on two of those roles at once: the compute behind a token got dearer to rent on Silicon Data’s index, while the price a lab charges per token fell in YipitData’s panels. The framework was published in May 2026 and does not discuss these figures.

What Is Not Established

The figures come from different providers, windows and samples, and none of them measures a lab’s margin directly. YipitData’s ~50% is an estimate from its own panels, Ramp’s figure covers four weeks, and Silicon Data’s index covers neo-cloud rates, not what labs pay under long-term contracts. We have not read The Information’s full article, and the compute-versus-revenue line comes only from an X post. We did not contact the companies named.

The Bottom Line

On Silicon Data’s index an H100 rents for 44% more than in December, by our arithmetic, while YipitData sees token prices about 50% lower year on year and OpenAI is winning coding work at Rippling on price. In its crash scenario, TS Lombard puts the strain upstream, among neoclouds and model developers.