SemiAnalysis: Nvidia Is Backstopping $588 Billion of AI Infrastructure Off Its Balance Sheet, and Anthropic Has Turned Compute-Profitable
Dylan Patel and Jordan Nanos on the Big Technology Podcast, October 9, 2026
SemiAnalysis CEO Dylan Patel and technical staff member Jordan Nanos argued that the consensus is underestimating both the size of the AI buildout and the revenue it will need to support. Patel said the widely circulated chart showing AI investment at 3.6% of GDP, against 2.2% for the railroads, is too low. He expects US capex of roughly $2 trillion next year once the broader supply chain is included, against GDP of just over $30 trillion, or "5 or 6%, not 3% or 2%." He added that US spending overstates domestic demand because America is serving the world: "Most of Europe's inference for Anthropic is in America, and most of OpenAI's inference in Europe is in America."
Nvidia's $588 Billion Backstop Universe
The most striking new figure came from Nanos, who said SemiAnalysis now tracks $588 billion of off-balance-sheet backstops supported by Nvidia. He put that at roughly 2.5% of the US M2 money supply. The support takes the form of revenue floors, landlord guarantees and leases that Nvidia plans to transfer to third parties. These let neo clouds and AI labs obtain investment-grade financing without a hyperscaler involved. Nanos said the exposure extends beyond GPUs to construction costs, memory suppliers, fab builders and industrial equipment makers, which he called "the long pole in the tent for that specific project."
Nanos rejected the argument that this is circular demand propping up a bubble. In his account, Nvidia signs as off-taker, and when the chips arrive, a third party usually wants them at a higher price, so Nvidia resells them rather than consuming them in internal research. He said the compute Nvidia retains for its own use is "a very small percentage" of its revenue. He framed the logic as pragmatic: Nvidia wants a diverse customer base and does not want to depend on three or four hyperscalers. The backstops are nonetheless a large contingent exposure, and the sheer size of the number is itself a risk worth tracking.
Anthropic Is Profitable on Compute, and OpenAI Is Close
Patel said Anthropic is now profitable when revenue is measured against total compute cost, training and inference combined. He expects OpenAI to get there "within the next couple quarters, maybe two or three." He said Anthropic's inference gross margins are around 75%, slightly higher on frontier models, in line with Salesforce, which took far longer to reach that level. He also argued that Anthropic's R&D spend is growing more slowly than revenue and that its customer acquisition cost is minimal. Switching costs are meaningful, since new models from rivals or Chinese labs "don't really shift the spend share that much for existing customers." His conclusion was that "it's arguably a better business than software."
On revenue, Patel is far more bullish than published break-even analyses from Bain and others. He expects OpenAI and Anthropic together to reach $600 billion to $700 billion of revenue by December 2026, and potentially $1.5 trillion by the end of 2027. The 2026 jump follows roughly 10x growth this year, with Anthropic going from under $10 billion to above $100 billion. He said the revenue requirements cited by Bain, a Columbia professor and others are "quite low." Revenue always trails capex, he said, and a $1 trillion investment on a six-year depreciation schedule implies about $250 billion of annual revenue on a straight-line basis.
Patel also dismissed the solvency concerns around Meta as "kind of silly." He noted that Meta can halt new construction at will and that its cash flow covers what it has already signed. Nanos added that the cancellation risk applies to self-built campuses. Neo cloud contracts typically run five to six years, with full capacity off-taken at fixed rates.
The Anthropic IPO Debate
Asked about his post saying one cannot invest in Anthropic's IPO at $2 trillion because "it could be zero, or it could be 20 trillion, but at 20 trillion, we're all screwed," Patel said he is highly bullish on the company and thinks it could be the first $10 trillion company. A $10 trillion valuation, however, would require something like $1 trillion of revenue at today's multiples, which he said is plausible in 2028 or 2029. At $20 trillion the implications become societal. If AI generates multiple trillions in revenue, the gains would accrue to a tiny group of shareholders and suppliers, and he warned of unrest if distribution is poor. Nanos was more sanguine, expecting revenue to come from new verticals such as drug development rather than per-token coding fees. He said the neo clouds are "the center of the world" on the road from a $2 trillion IPO to a $20 trillion valuation.
Oracle: A Good Cloud, a Troubled Project Pipeline
Patel drew a sharp line between ClusterMax 3.0, which rates GPU cloud usability, reliability and security, and stock recommendations. He said Oracle is "one of the top neo clouds in terms of quality of service," yet it faces real execution and financial problems. SemiAnalysis told clients in April that Oracle's New Mexico data center, contracted for OpenAI, would be delayed because required pipelines would not be built. Oracle has since announced a force majeure on the project, and the stock fell. Patel said the site could be delayed by multiple years.
He listed three further problems. Oracle's OpenAI contracts carry essentially fixed margins, so it cannot capture the rise in GPU rental prices. Its debt issuance spooked markets despite its being "completely solvent." Other sites, including one in Wisconsin, also have delays. He added that Nebius, which had the shortest average contract length among the top-tier clouds, has been able to "really jack up pricing." CoreWeave and Oracle locked into long-term contracts and captured less of the upside.
Neo Cloud Security Is Poor
SemiAnalysis published an article titled "Most Neo Clouds Suck at Security," and Ilya Sutskever posted a warning about neo cloud cyber weakness about eight hours later. Patel said open-source Chinese models, GLM in particular, can hack dozens of neo clouds. He described an incident in which SemiAnalysis, without trying, could see other tenants' storage and Slurm jobs on one cluster. Those tenants included government security agencies from a country with a top-10 GDP, and the firm reported the problem immediately. Nanos said the sovereign AI push means state agencies often share infrastructure with others because there may be only one facility in a country with GB200 NVL72 systems: "It's a seller's market."
Nanos said that with 323 providers tracked, the long tail of unrated or "unavailable" clouds reflects a market in which the priority is raising capital and deploying chips quickly rather than service quality. He called the bar for recommendation "quite low."
Google's Compute Allocation and Frontier Labs Holding Back
Patel said the share of Google's compute allocated to DeepMind has fallen over the past six months relative to two to three years ago, in part because Google is supporting Anthropic with TPUs. He said DeepMind will have the least compute of the three leading labs by year-end, and quipped that "Thomas Kurian won and Demis lost." He sees the risk growing next year if OpenAI and Anthropic each add 5 to 10 gigawatts while Google adds only a few. Nanos said Gemini lacks a niche, as it trails the top two or three on coding.
Patel also said the leading labs are already withholding their best models. Anthropic had its Mythos model ready in February and has released only a reduced version. OpenAI is similarly sitting on a newer model. He expects both labs to release new models in the coming weeks, Anthropic ahead of its IPO. Nanos expects "minimum viable models" to be released to stay ahead of Chinese and open-source rivals. Patel, for his part, argued that real pacing of the frontier is limited, since neither lab can slow down given mutual distrust and China's proximity.