Nvidia has backstopped about 6.5 gigawatts of compute, SemiAnalysis chief executive Dylan Patel said in a talk at Modal’s Runtime 2026 conference. He put the most Nvidia could ever support at 45 gigawatts. By our arithmetic, that means Nvidia has used about 14 per cent of the ceiling Patel estimates.
Patel described three ways Nvidia backs other companies’ clusters so that lenders will fund them. Nvidia’s own 10-Q, for the quarter to 26 July 2026, lists $36 billion of such cloud agreements and up to $108.5 billion of guarantees. Together that is $144.5 billion, by our arithmetic.
The source talk · Dylan Patel at Runtime 2026
Modal’s video of the talk, “Compute, capital, and who actually delivers it”. The backstop section runs from about 14:00 to 18:40.
Business Pill · Financial Conditions
An evergreen Business Engineer explainer on financial conditions: several prices, including the cost of debt, pulling at once. It teaches the general idea only and does not discuss Nvidia, Patel or any company in this story.
The key insight: Patel estimates that Nvidia has backstopped about 6.5 gigawatts of clusters, about 14 per cent of the 45 gigawatts he says it could support, by our arithmetic. Nvidia’s 10-Q lists $36 billion of cloud agreements and $108.5 billion of guarantees behind that effort.
Modal’s site dates Runtime 2026 to 1 October in San Francisco. Modal posted the video of Patel’s talk on YouTube on 2 November. All quotes below are from the video’s automatic captions, with timestamps.
Patel started with the cost of debt. A year ago it “used to be much cheaper”, he said [14:06]. For a neocloud trying to grow from 100 megawatts to a gigawatt, “The only way they can get a loan is if they go sell that gigawatt to one of the hyperscalers” [14:21]. The catch, in his words: “none of that compute gets onto the market for the general industry” [14:32].
He said a startup trying to rent a cluster will be asked “to pay half of it down today” [15:26]. Nvidia’s answer, as he described it, is to “backstop your cluster” [15:41]. The cloud can “give us a revenue share on that cluster” and “use our name as a price floor to go get the loan”.
“Nvidia’s loading up their balance sheet with all these compute commitments because it’s the only way that the open ecosystem can get compute,” he said [15:57].
Gigawatts of compute: Dylan Patel’s estimates that Nvidia has backstopped about 6.5 GW and could support at most 45 GW, against the one site Nvidia’s 10-Q sizes in gigawatts, SB Energy’s Ohio campus (about 4.25 GW of leases guaranteed, plus an option on about 3.8 GW more). The 14% share is our arithmetic. Patel did not say whether his 6.5 GW includes the Ohio site, so the bars are not additive. Sources: Patel’s Modal Runtime 2026 talk; NVIDIA Form 10-Q for the quarter ended 26 July 2026.
The Three Deal Types, Checked
1. Land, power and shell. Patel said that with SoftBank Energy Nvidia “only guaranteed the land, the power, and the shell” [16:59]. He called it a “like 9-gigawatt facility in Ohio” and said “the whole thing is earmarked for OpenAI though”.
Nvidia’s 10-Q describes guarantees “capped at a total of $105 billion” with affiliates of SB Energy Corp., on behalf of an OpenAI affiliate. They cover leases for “approximately 4.25 gigawatts of IT load” at the PORTS Technology Campus in Pike County, Ohio. Nvidia also holds an option on “approximately 3.8 additional gigawatts”. That is about 8 gigawatts in all, by our arithmetic, a little under Patel’s 9.
The filing says the first of nine phases “is expected in fiscal year 2029”. It also says “Our guarantees are limited to defined portions of lease and power payments and not the full cost of the site”.
2. A rental floor with a revenue share. Patel’s second example is the company the captions spell “Fermus”: Firmus Technologies, which Reuters describes as Australian. He said Nvidia agreed to rent its cluster at a price where it “makes no money” but does not “lose money” [17:32].
Then, he said, “every dollar per hour that they sell over that price, Nvidia gets to take half of it” [17:46]. He called that a “profiteering move”, then added: “I think I love it”.
Reuters reported on 29 June that Firmus would buy Nvidia infrastructure and sell Nvidia-powered cloud services. The deal would earn Nvidia “product revenue and a share of cloudrevenue”. It covers 170,000 GPUs in Batam, Indonesia, delivered from the first quarter of 2027.
The 10-Q confirms the model. Under it, “AI clouds procure our data center infrastructure products and we commit to cloud service agreements”. Those commitments “totaled $36 billion as of July 26, 2026”. “If certain criteria are met, we will participate in revenue share generated by the AI clouds from third-party customers.”
3. Residual value guarantees. Patel did not explain this one, saying he had 45 seconds left. He called it “a whole residual value guarantee, which is pretty interesting” [18:08]. The 10-Q’s exhibit list includes a “Form of Residual Value Guaranty”.
What SemiAnalysis Has Published
The 6.5 gigawatt figure also appears in SemiAnalysis’s 11 September newsletter, Nvidia’s Backstop Universe. Its free preview says: “By our count, Nvidia today backstops ~6.5 GW of capacity, most of which is not yet built.”
The same text puts the rental floor at “approximately $2.35/hr/GPU on average for the GB300 vs the $4.50 to $4.60 range for 5y contracts”. That floor is about 52 per cent of $4.50, by our arithmetic. SemiAnalysis also says Firmus’s 360MW deal in Batam “corresponds to a $21.1B total obligation”.
On the third type, SemiAnalysis says Nvidia’s only support in its $500 billion capital program is “a residual value guarantee for up to 25% of a deal”. By its numbers, that structureneeds $9.4B of obligations per gigawatt, against $59B for the rental-floor program. The rental floor needs 6.3 times as much per gigawatt, by our arithmetic.
The Pause
On 27 August, Reuters reported that Nvidia had paused some of the revenue-sharing deals, citing the Wall Street Journal. That report gave the same split Patel describes: “Nvidia would receive 50% of any revenuecloud providers earned through its chips beyond a certain threshold”.
An Nvidia spokesperson said then that the model “is still in place and continues to evolve due to high demand”. SemiAnalysis wrote on 11 September that “Nvidia is not advancing any further AICP deals at this time”. Patel’s talk, a few weeks later, describes the 6.5 gigawatts as already done.
As we read Patel and SemiAnalysis, each structure lends Nvidia’s credit at a different point: the landlord’s lease, the cloud’s rental income, or the value of the hardware if a deal fails.
As we read the 10-Q, the floor runs one way: the AI clouds “can unilaterally stop providing to us and sell to third-party customers at more advantageous rates”.
Dylan Patel, SemiAnalysis, at Modal Runtime 2026 [15:57]
“Nvidia’s loading up their balance sheet with all these compute commitments because it’s the only way that the open ecosystem can get compute”
Three Implications
FOR NEOCLOUDS Patel says the other route to a loan is selling the capacity to a hyperscaler, which keeps it off the open market. In his account, a Nvidia floor lets them borrow and keep selling, for half of the upside above the floor.
FOR STARTUPS Patel says a startup renting a cluster is asked “to pay half of it down today” and told the capacity will come in three, six or nine months. He says the 6.5 gigawatts is capacity Nvidia has wrapped to be sold to the open market.
FOR NVIDIA’S BALANCE SHEET By SemiAnalysis’s numbers, the rental floor uses $59B of obligations per gigawatt and a residual value guarantee $9.4B: 6.3 times as much for the floor, by our arithmetic.
The framework reads: “the capex race is no longer a corporate spending cycle. It is a sovereign-scale financing problem dressed in tech-company clothing”.
As we read it, Patel’s three deal types show one supplier lending its own credit so that smaller buyers can borrow, at a ceiling he puts at 45 gigawatts. The framework was published in May 2026 and does not discuss these deals.
The 6.5 and 45 gigawatt figures are Patel’s and SemiAnalysis’s estimates. Nvidia’s 10-Q does not give a gigawatt total for its backstops. Patel did not say how he reached 45 gigawatts, or whether the Ohio site counts in the 6.5.
The Firmus rental price and the exact split are Patel’s account. Reuters’ report on the Firmus deal mentions a revenue share but no price floor. The filing names no partner in its $36 billion. Nvidia did not comment on the talk in anything we read.
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.
Patel says Nvidia has backstopped 6.5 gigawatts of clusters and could do at most 45 gigawatts: about 14 per cent so far, by our arithmetic. Its 10-Q lists $36 billion of cloud agreements and $108.5 billion of guarantees, which come to $144.5 billion, by our arithmetic.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Patel’s quotes are from YouTube’s automatic captions of Modal’s video, checked against the audio timestamps; the 6.5 and 45 gigawatt figures are his and SemiAnalysis’s estimates. Nvidia figures are from its Form 10-Q for the quarter ended 26 July 2026. SemiAnalysis is quoted from the free preview of a paid article. The 14 per cent, $144.5 billion, 52 per cent and 6.3 times figures 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.
