Dylan Patel: Who Can Actually Deliver AI Compute, and Who Pays for It

SemiAnalysis CEO Dylan Patel explains why AI compute is so hard to buy, which clouds can actually deliver it, and how Nvidia now backstops the debt behind it. In this 19-minute talk at Modal’s Runtime 2026 (published November 2, 2026), Patel walks through global compute nearly doubling in two years, H100 rental prices jumping from $1.70 to $2.80 an hour, the striking profitability of serving open-source models, what SemiAnalysis found when it let AI models try to hack GPU clouds, and the $11 trillion of AI capex he expects by 2029. All figures are his, as of the talk.

TLDW

Patel says OpenAI and Anthropic are absorbing 25% to 40% of all new compute, which has pushed GPU prices up sharply, and that serving even open-source models is so profitable that anyone who can rent GPUs wants to sell tokens. His ClusterMAX review ranks CoreWeave and Nebius at the top and downgrades AWS, Azure, Crusoe and Together, partly on security failures. The deeper constraint is financing: neoclouds cannot borrow to build unless a hyperscaler or Nvidia stands behind them, so Nvidia has started guaranteeing clusters itself, though that covers only a small slice of the compute coming online.

Thoughts

The binding constraint on AI compute is no longer chips or even power. It is balance sheets. Patel’s account of the debt market is the sharpest part of the talk: a neocloud running 100 megawatts that wants to scale to a gigawatt cannot get a loan unless it first sells that gigawatt to Microsoft or Google, at which point none of it reaches the open market. That is why he puts 65% to 70% of next year’s signed compute in the hands of hyperscalers and AI labs, and why he expects it to get worse as rates rise. Of his $11 trillion capex estimate through 2029, $7 trillion is debt, and lenders price that debt on the creditworthiness of the tenant, not the quality of the cloud. The result is a market where access to compute tracks access to investment-grade counterparties. His closing line, that otherwise “we are all slaves to Anthropic and OpenAI,” is a joke, but the mechanism behind it is plain finance.

Nvidia’s response turns it into something closer to a bank and an insurer of its own demand. Patel describes three deal types. With SoftBank Energy it guarantees only land, power and shell for a 9-gigawatt Ohio site that is entirely earmarked for OpenAI. With Firmus in Australia it guarantees to rent the cluster at a break-even price, which makes the project financeable, and then takes half of every dollar per hour sold above that floor. Patel calls this “profiteering” and says he loves it, and both are fair: Nvidia captures upside on hardware it already sold, while carrying downside only if AI demand collapses. The catch is scale. Nvidia has wrapped about 6.5 gigawatts this way and could stretch to roughly 45, a small share of what gets deployed through 2030. It also concentrates risk in a familiar way: the chip supplier is underwriting the customers who buy its chips, which works beautifully until utilization falls everywhere at once.

The counterintuitive point is that open-source models are pushing compute prices up, not down. Patel’s arithmetic: a megawatt of B300s costs about $40 million to build and $9 million a year to run over five years, or around $14 million a year to rent if you are lucky. Serving DeepSeek V4.1 Flash on that megawatt at 60% utilization brings in about $70 million a year in tokens, using SemiAnalysis’s own open-source InferenceX containers with no fancy speculative decoding. Even at 30% utilization the margins are enormous, so he knows “dozens” of people whose entire business is renting GPUs and reselling tokens. That reframes the price spike. It is not only the labs hoarding capacity; every better open model raises what a GPU-hour is worth, and token providers capture the value. When Patel says he would not be surprised to see H100s above $3 and B300s above $7, he is describing a market where the rental price chases the token price.

The ClusterMAX security findings deserve more attention than they will get. SemiAnalysis pointed AI models, including an open-source model behind the frontier, at neoclouds, and several of them used already-published vulnerabilities to reach other customers’ storage. In one case the model gained access to Slurm across an entire cluster and could see job names belonging to a foreign country’s defense and intelligence agencies, at which point the team logged out immediately. Combined with failure-injection tests where many clouds never noticed a GPU error until told, and took days to respond after that, it suggests that a lot of the new capacity is being stood up faster than it is being operated. In a seller’s market, buyers cannot afford to be picky about who has GPUs. That is exactly when a published, free ranking of who runs a cloud properly is most useful.

Key Takeaways

SemiAnalysis, which Patel started alone six years ago and bootstrapped to 115 people, sells data on every data center, power plant, fab and cable, and runs a chip teardown lab in Oregon.
Its open InferenceX benchmark runs daily across about 2,000 chips, including seven Nvidia GPUs, three AMD GPUs, Google TPUs and OpenAI’s internal chip, with Trainium coming.
Global compute (AI and non-AI) went from about 60 gigawatts at the end of 2024 to over 100 gigawatts by the end of 2026.
OpenAI and Anthropic together had about 4 gigawatts at the end of 2025, about 10 this year, and should pass 20 next year, taking 25% to 40% of all new compute.
The labs spend far more compute on training than inference: Patel says all of Claude Code is served from about 1 gigawatt while Anthropic has 4.
H100 rental prices bottomed around $1.70 an hour (as low as $1.45) and now trade between $2.40 and $3.20, with $2.80 the midpoint for one-year deals.
A B300 megawatt costs about $40 million to build and $9 million a year to run, but can earn about $70 million a year serving an open-source model at 60% utilization.
Patel’s advice: if you can find B300s for $6 an hour, sign the contract and sell tokens, even though clouds will want you to prepay.
ClusterMAX tracks 323 GPU cloud providers, reviewed 77 and gave medallions to 19, with CoreWeave and Nebius rated best and Oracle and Google also strong.
Azure was downgraded to silver, and Crusoe, Together and AWS to bronze; AWS lost ground on how hard it is to get a cluster and on its HyperPod offering.
CoreWeave is sold out for six months, and Nebius requires six to twelve months prepaid, so startups often must pay half up front and wait months for delivery.
SemiAnalysis expects $11 trillion of global AI capex by the end of 2029, about $5 trillion from cash and equity and $7 trillion from debt.
Nvidia has backstopped about 6.5 gigawatts of compute for the open market and could reach about 45 gigawatts at most.

Chapters

0:07 What SemiAnalysis Tracks

Patel describes a team drawn from chip equipment makers, fabs, Nvidia, hyperscalers and model training groups. The business is mostly selling data on data centers, power plants, fabs and cables, plus testing work: ClusterMAX for clouds, InferenceX for chips, and an Oregon lab that delayers chips to compare Intel 18A and Samsung 2nm against TSMC.

2:35 The Compute Race and the Doubling of Global Capacity

It took about 50 years to reach roughly 50 gigawatts of compute and two years to reach 100. Patel says OpenAI and Anthropic drive much of the growth, with Anthropic buying aggressively after Claude Code revenue took off, and Meta and OpenAI doing the same. He expects the labs to add around 30 gigawatts in 2028 alone.

4:27 Training Over Inference, and the H100 Price Spike

Most lab spending goes to training, not inference. Patel then shows rental prices: H100s bottomed near $1.70 an hour in October 2025 and now clear around $2.80, using data from SemiAnalysis’s own compute brokering. Long contracts are scarce, which hurts startups that cannot commit beyond a year.

6:16 The Economics of Selling Tokens

Patel compares a $40 million build and $9 million a year in operating cost per megawatt with about $70 million a year in revenue from serving DeepSeek V4.1 Flash at 60% utilization. Running inference at small scale is not hard, he says, and dozens of people he knows simply rent GPUs and resell tokens at large profits.

8:28 Why $6 B300s Are Worth Signing

Because token providers capture the value, clouds can keep raising prices. Patel says Anthropic’s revenue per megawatt on its frontier models is above $100 million, citing investor materials from its IPO process, and he would not be surprised to see H100s above $3 and B300s above $7 as open models like GLM 5.3 and DeepSeek V4.1 improve.

9:48 ClusterMAX and Letting AI Hack the Neoclouds

ClusterMAX ranks clouds on security, networking, storage, lifecycle and managed Slurm or Kubernetes. In version 3, SemiAnalysis had models attack neoclouds; several exposed other customers’ storage through published vulnerabilities, and one exposed a foreign defense agency’s jobs. The findings became an article titled “most neoclouds suck at security.”

12:09 Downgrades and Failure Injection

Azure fell to silver, and Crusoe, Together and AWS to bronze. AWS is penalized because getting a cluster means talking to many people and dealing with awkward networking. SemiAnalysis also injected Nvidia GPU errors and found many clouds did not detect them until alerted, and some took days to respond.

13:13 $11 Trillion in Capex and the Debt Squeeze

Patel expects $11 trillion of AI capex by 2029, mostly chips but also data centers and on-site power plants. With the US 10-year yield spiking, debt is getting more expensive, and neoclouds can only borrow by leasing capacity to hyperscalers, leaving 65% to 70% of next year’s signed compute with the hyperscalers and labs.

15:00 Nvidia Steps In to Backstop Clusters

Startups renting a cluster are asked to pay half up front and wait three to nine months. Nvidia now lets clouds use its name as a price floor to secure loans, in exchange for a revenue share, loading its balance sheet with compute commitments. Inference companies and neolabs are increasingly building their own clusters this way.

16:45 Three Kinds of Nvidia Deals, and the Limits

Patel outlines the SoftBank Energy deal (land, power and shell for a 9-gigawatt Ohio site reserved for OpenAI), the Firmus deal (a break-even rental floor with Nvidia taking half the upside) and residual value guarantees. Nvidia’s 6.5 gigawatts of backstops, and a ceiling of about 45, are small next to total deployment. He closes on 100 trillion-plus tokens a day across a few providers.

Notable Quotes

“It took us, you know, 50 years to get to 50 and then, you know, just a couple years to get to 100.”
Dylan Patel, on global compute capacity in gigawatts

“It’s pretty shocking how little compute they spend on inference relative to training.”
Dylan Patel, on how AI labs split their compute

“If you can find B300s for $6 an hour, you should sign it and you should sell tokens.”
Dylan Patel, on the economics of reselling inference

“No one’s in that country at SemiAnalysis, no one wants to go to jail, but we immediately closed out.”
Dylan Patel, after an AI model exposed a foreign defense agency’s jobs on a neocloud

“The only way they can get a loan is if they go sell that gigawatt to one of the hyperscalers.”
Dylan Patel, on why neoclouds struggle to finance growth

“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.”
Dylan Patel, on Nvidia’s cluster backstops

“Hopefully, they can get compute because otherwise, we are all slaves to Anthropic and OpenAI.”
Dylan Patel, closing the talk

Watch the full talk here.