GPU rent turned up: B200 hours rose 21% in a month, H100 hours 7%

01The line: two months up, after a summer down

Ornn’s month-end settles show the shape. The B200 fell 5.3% in August, then rose 33.3% in September and 21.5% in October to 29 October. The H200 fell 21.4%, then rose 26.2% and 6.8%. The H100 has gone sideways: up 7.6% in August, down 8.8% in September, up 7.7% in October, inside a range of US$2.49 to US$3.57 since 30 July.¹²³ Two monthly rises are a turn in the line, not yet an established trend, and Ornn’s free history starts on 30 July, so the same measure cannot be set against 2025.

Posted prices tell a longer story with a different turning point. GetDeploying’s weekly median of on-demand H100 prices across providers fell from US$2.98 at the end of October 2025 to US$2.49 in March 2026, then climbed from May to US$3.99 in late October.⁷

02What drives it, in the words of the people closest to it

The buyers of compute. On Alphabet’s results call on 28 October, Sundar Pichai said of Google Cloud: “there’s huge demand. We are working hard to bring more capacity online”. Google’s own model APIs were processing nearly 32 billion tokens a minute, 40% more than the previous quarter. Anat Ashkenazi, the chief financial officer, said that “given the supply-constraint environment, we started utilizing third-party compute capacity at the end of Q3 and will ramp usage in Q4”. Alphabet did not name the suppliers or the chips.⁹

The parts. SK hynix, which makes the high-bandwidth memory stacked on these GPUs (the B200 carries 180 GB of HBM3e), said on 28 October that it expects “the structure in which demand exceeds supply to persist through 2030”.¹⁰¹

The index publishers. Singularity Research, which republishes the posted-price series, reads it this way: “Rising rents, especially for older chips like the H100 and A100, mean demand still exceeds supply. Falling rents mean new capacity has caught up.”⁷

What does not fit. The newest chip moved the other way: the median posted price of a B300 fell 16.3% over the 30 days to 30 October.⁵ And what labs charge per token kept falling on Ornn’s blended token indices, 41.6% for Anthropic and 48.9% for OpenAI over the month. A blend moves when buyers shift between models, so that is not the same as a price cut.⁴

03What it means for a company buying GPU time or AI services

A budget built on 2025’s falling prices is two months out of date for the B200 and H200; for the H100, traded prices are roughly where they were in July. The gap between what sellers post and what buyers pay is wide. Ornn found posted on-demand H100 rates at four GPU clouds 19.5% to 111% above its 29 October settle (rates read on 6 October).² On 30 October the same H100 SXM was listed from US$1.99 to US$11.68 an hour across providers.⁵

The price of a commitment has also moved. A year ago a 12-month H100 reservation listed at US$3.06 an hour, slightly above the on-demand median of US$2.98. On 26 October it listed at US$3.26 against US$3.99 on demand, 18% lower.⁷

Buying the output instead of the hours does not escape the question. Google’s Gemini API price page, updated on 28 October, doubles Gemini 4 Argon from US$2 to US$4 per million input tokens and from US$10 to US$20 per million output tokens on 1 December, and Gemini 3.8 Flash from US$0.75 and US$3.75 to US$1.50 and US$7.50 on 1 January 2027. The page gives no reason.¹¹