A one-year rental of an Nvidia H100 now transacts at about $2.80 an hour, up from a trough of $1.70 in October 2025, SemiAnalysis founder and CEO Dylan Patel said in a talk at Modal’s Runtime 2026 conference. By our arithmetic that is a rise of about 65 per cent in a year.
Patel gave the talk on 1 October in San Francisco. Modal posted the video on YouTube on 2 November. He said SemiAnalysis brokers compute, which lets it see the prices deals actually close at.
The talk · Dylan Patel at Runtime 2026
Dylan Patel, “Compute, capital, and who actually delivers it”, Modal Runtime 2026, 1 October 2026. The price discussion starts at about 4:56.
Business Pill · GPUs
An evergreen Business Engineer explainer on why AI runs on GPUs. It teaches the general idea only and does not discuss rental prices or any company in this story.
The key insight: Dylan Patel said one-year Nvidia H100 contracts now transact around $2.80 an hour, against a $1.70 trough in October 2025: about 65 per cent higher, by our arithmetic. His own firm’s April report had the rise at almost 40 per cent by March.
At about 5:01 in the video, Patel said: “in October it was sort of the trough, right? Um you know, sort of $1.70 for H100, and you could find it as low as $1.45”.
He then said current transactions run from $2.40 to $3.20 an hour, with “the midpoint being 280”. Later he put it plainly: “Even uh H100, which is $2.80 for a year”. He added that “people are desperate to sign something in that range”.
On where SemiAnalysis gets its numbers, he said: “some of the business we do is we broker a lot of compute, and so that lets us get a lot of price um accuracy cuz we know what stuff’s actually transacting at”.
One-year Nvidia H100 rental prices as Dylan Patel, founder and CEO of SemiAnalysis, gave them at Modal’s Runtime 2026 on 1 October 2026: a $1.70 an hour trough in October 2025 and a $2.80 midpoint for deals now. The H100 $3 and B300 $7 levels are his predictions, not prices; $6 is the B300 rate he said is worth signing. The 65% rise is our arithmetic.
Which October Was the Trough
The talk itself was on 1 October 2026, so the trough he named is October 2025. SemiAnalysis’s own 2 April report, which launched its public H100 one-year rental index, dates it the same way.
That report said one-year H100 pricing “has shot up almost 40% from a low of $1.70/hr/GPU in October 2025 to $2.35/hr/GPU by March 2026”. Patel’s $2.80 is about 19 per cent above that March level, by our arithmetic.
The report says the index is built from monthly surveys of more than 100 market participants and checked against transaction data, including deals SemiAnalysis arranges itself.
What a Year of One GPU Costs Now
At $1.70 an hour, one H100 rented for all 8,760 hours of a year costs about $14,900. At $2.80 it costs about $24,500, roughly $9,600 more per GPU, by our arithmetic. The ends of his range, $2.40 and $3.20, are 41 and 88 per cent above the trough.
Patel said the term is the problem for young companies: “it’s very hard to buy any compute for a year, which is sucks for any startup because it’s like, well look, I don’t know if I’m going to be in business more than a year from now, so I should only buy compute for a year”.
B300 at $6, and Prepay
For Nvidia’s newer Blackwell chip, he said: “if you can find B300s for $6 an hour, you should sign it and you should sell tokens”. He called it “an insane price” but “a correct decision to make”. Elsewhere in the talk he tied rising prices to how profitable serving open-source models has become.
He added: “of course, all the clouds want you to prepay, so then that makes it a lot harder”. Of two providers SemiAnalysis rates highest, he said “CoreWeave will not sell you compute for the next 6 months”, and that Nebius makes buyers prepay “a full year upfront or 6 months upfront at some ridiculous price”.
SemiAnalysis’s April report described the same shift in terms: neoclouds and hyperscalers “can now negotiate for more favorable terms such as higher prepay, better pricing, longer contract lengths”.
His Prediction
Patel said the token sellers capture the value, so “all the clouds can push price”. He then said: “I would not be surprised if H100s are north of $3 and B300s are north of, you know, $7”.
That is a prediction, not a price. It sits at least 7 per cent above his $2.80 for the H100 and at least 17 per cent above $6 for the B300, by our arithmetic. He gave no date for it.
How Other Price Series Compare
Other series measure different things and are not directly comparable with a one-year contract. Our 24 October piece on Silicon Data’s neo-cloud index, Nvidia H100 Rent Hits $2.88 an Hour, put its reading at $2.88 that day, against $2.00 on 9 December 2025.
On the Ornn Compute Price Index, the H100 SXM rose from $2.65 to $3.03 between 11 and 25 October 2026, as reported in our piece Nvidia B200 Rent Jumped 30% in Two Weeks.
As we read the talk, the term Patel says matters most to startups is the one-year contract, which a young company can commit to without betting on a longer future.
As we read it, his explanation runs through inference: he ties the rise to how profitable it has become to serve open-source models, which makes renting at higher rates a sound decision for token sellers.
Dylan Patel, founder and CEO, SemiAnalysis, Modal Runtime 2026, 1 October 2026
“if you can find B300s for $6 an hour, you should sign it and you should sell tokens”
Three Implications
FOR STARTUPS Patel said it is very hard to buy compute for only a year. One H100 for a year costs about $24,500 at $2.80, against about $14,900 at the trough, by our arithmetic.
FOR BUYERS OF BLACKWELL CAPACITY He said a B300 at $6 an hour is worth signing, while the clouds want prepayment: CoreWeave, he said, has nothing to sell for six months.
FOR NEOCLOUDS SemiAnalysis’s April report said providers can now ask for higher prepay, better pricing and longer contracts. His prediction is H100 north of $3 and B300 north of $7.
The framework’s thesis is that “Inference is a revenue engine: continuous, distributed across millions of applications and agents, incurring marginal cost with every query.”
As we read it, Patel’s account of rising rents fits that frame: he says token sellers capture the value, so renting GPUs at higher prices still pays. The framework was published in March 2026 and does not discuss these prices.
Patel’s $2.80 is a midpoint of deals SemiAnalysis sees, not an audited market price, and the talk does not say how many transactions it covers. He did not give a contract term for the $6 B300 figure.
The talk is a month old. Prices quoted on 1 October may have moved by the time the video was posted on 2 November. His $3 and $7 levels are predictions with no date.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
By Patel’s numbers, a one-year H100 contract moved from $1.70 to about $2.80 an hour between October 2025 and October 2026, about 65 per cent by our arithmetic. He described a market where clouds set the terms, ask for prepayment and, in CoreWeave’s case, have no capacity to sell for six months.
A note on sourcing. Quotes are from the auto-generated captions of Modal’s video of Dylan Patel’s 1 October 2026 talk, posted 2 November; the captions write prices without dollar signs or decimal points. Index figures are from SemiAnalysis’s 2 April 2026 report. Silicon Data and Ornn readings are as reported in our earlier pieces. All percentages and annual costs are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
