AI capex hits 5-6% of GDP and demand still outruns supply

1. Capex is understated — ▶ 1:47

   Next year's US capex is close to $2 trillion once the wider supply chain is counted, against $30-plus trillion of GDP. That is 5-6%, not the 3.6% or 2.2% on the railroad chart. Part of the gap is structural. Railroads and fiber served the country that paid for them. US AI capex also serves other countries: most of Europe's Anthropic and OpenAI inference runs in America, and training happens there too. The economic impact is larger than any physical infrastructure cycle before it.

2. Slowdown is arithmetic — ▶ 5:57

   Hyperscaler capex growth peaked at 116% year over year, and consensus has it falling below 100% and then toward 70%. Nanos expects limits by the end of next year: it gets hard to marshal resources for another gigawatt and to raise capital at trillion-dollar scale. Patel calls the deceleration definitional. Anthropic's 10x year, from under $10 billion to north of $100 billion, cannot repeat into a trillion. Nvidia still doubles, but off a bigger base.

3. The payback math — ▶ 9:24

   GPUs depreciate over six years and data centers last 15. A trillion dollars of capex needs about $1.5 trillion of revenue, or $250 billion a year on a straight line. Patel thinks break-even estimates like $6 trillion by 2031 will be cleared. He expects $600-700 billion from OpenAI and Anthropic by December next year and $1.5 trillion by the end of 2027. Anthropic is already profitable after training and inference compute. OpenAI follows in two or three quarters. At only $3 trillion in 2031, he concedes, "big problems".

4. Labor share keeps falling — ▶ 15:09

   Customers' R&D and cost-of-goods budgets shift to AI, and some money comes from labor. Labor's share of GDP has fallen since the 1970s, and AI tokens are "basically just capital" doing labor's work, so the trend accelerates. Patel hopes labor dollars stay flat or grow as the rest of the economy expands. His model is chip design: American design engineers have been flat for 20 years while the value at Nvidia, Broadcom and Google exploded. Same headcount, better tools, growing revenue and margins.

5. Demand outruns supply — ▶ 20:49

   Nanos separates contract types. Neocloud customers sign five- or six-year terms, take full capacity at a set hourly rate and cannot cancel monthly. The flexibility is in self-build sites like Meta's campuses, which can sit as empty shells for years. Construction is a small share of project cost, with HBM, optics and GPUs the bigger bill. SemiAnalysis's models show demand above supply, but nobody can say whether by 10%, 50% or infinitely, because supply cannot catch up to find out. No demand limit has been found.

6. The $20 trillion Anthropic problem — ▶ 24:41

   Patel thinks Anthropic could be the first $10 trillion company. At $100 billion of revenue and a 20x multiple, an IPO lands near $2 trillion. Reaching $10 trillion likely needs about $1 trillion of revenue, possibly in 2028 or 2029. At $20 trillion, revenue would be $2-4 trillion, and gains concentrating in few hands could push distressed people to "overthrow governments". Nanos is more optimistic. He sees revenue from products like cancer drugs, as with GLP-1s and Lilly, not per-token metering.

7. Better margins than SaaS — ▶ 29:21

   Salesforce, among the best SaaS companies, runs about 75% gross margins and took years to get there. Anthropic does 75% or slightly higher on inference, especially frontier models, in far less time. SaaS customer acquisition can cost up to 30% of revenue. Anthropic's is near zero. Its R&D has grown slower than revenue, and customers are sticky: a better Chinese model barely shifts existing customers' spend. Patel calls it "arguably a better business than software".

8. Nvidia's $588 billion backstop — ▶ 35:24

   SemiAnalysis tracks $588 billion of off-balance-sheet Nvidia backstops, about 2.5% of the US M2 money supply, against Nvidia revenue of roughly $400-500 billion. Nanos reads it as pragmatism: Nvidia wants customers beyond three or four hyperscalers, so it acts as investment-grade offtaker that lets neoclouds raise debt. When someone pays more, Nvidia resells the GPUs instead of using them for research. Internal research is a very small share of revenue. The backstops also cover construction, memory suppliers, fab builders and industrial equipment.

9. Neoclouds are insecure — ▶ 38:54

   SemiAnalysis published a report arguing most neoclouds are poor at security. About eight hours later, Ilya Sutskever tweeted that rogue agents might try to take over a neocloud to run more copies. Patel says open-source Chinese models can already hack dozens of them. At one cloud his team accidentally saw other tenants' storage and Slurm jobs, including non-American government and intelligence agencies from a top-10 GDP country, among roughly a dozen customers. Nanos ties it to sovereign AI: with one GB300 NVL72 facility in a country, it is a sellers' market.

10. Oracle's execution gap — ▶ 44:19

    ClusterMAX rates cluster quality, not stocks, and Oracle ranks among the top neoclouds. Yet SemiAnalysis told clients in April that Oracle's New Mexico site for OpenAI would slip on a pipeline problem, repeated it in May and June, and Oracle then declared force majeure. Its OpenAI deals carry fixed margins, so it cannot capture higher GPU prices. Nebius has the shortest average contract length among the top clouds and repriced upward. CoreWeave and Oracle locked into long terms. Patel still calls Oracle fine.

11. Labs hold models back — ▶ 55:44

    Anthropic had Mythos ready in February, and the public still has only a reduced version. OpenAI's Astra sat for a couple of months, and a newer model shown with math results is unreleased. Months of progress stay internal because releasing invites distillation. Nanos says only Anthropic publicly paces the frontier and his bet is minimum viable releases to stay ahead of Chinese open source, about 6 months behind. Patel sees little room to slow: each lab distrusts the other, and China is not far behind.

Capacity, not appetite, sets the pace of this buildout. Jordan Nanos puts it plainly: nobody can say whether demand exceeds supply by 10%, 50% or an infinite amount, because supply cannot catch up far enough to find out. Chips, data centers and power are hard constraints. Until demand shows a ceiling, SemiAnalysis's models say it outstrips what can come online.

That reframes the slowdown Wall Street expects. Hyperscaler capex growth hit 116% year over year, and consensus has it falling below 100% and then toward 70%. The deceleration is real, but it is arithmetic and logistics, not weak demand. Doubling again off a bigger base is harder. So is marshalling the resources for another gigawatt, and raising capital when the numbers run into the trillions.

The obvious worry is that big customers walk away. Nanos separates two cases. Neocloud customers sign five- or six-year terms, take full capacity at a set hourly rate and cannot pull the plug monthly. The flexibility sits in self-build campuses such as Meta's, which can be left as empty shells for years. Even there, concrete and steel are a small share of project cost. HBM, optics and GPUs are the bigger bill. Dylan Patel adds that Meta can stop buying at any point and its cash flow covers what it has signed, so solvency is not the issue.

Whether it pays back is a revenue question, and Patel's answer depends on a 10x year. GPUs depreciate over six years and data centers last 15. A trillion dollars of capex therefore needs about $1.5 trillion of revenue, or $250 billion a year on a straight line. Outside estimates that hyperscalers need $6 trillion by 2031 look low to him. Anthropic went from under $10 billion of revenue to more than $100 billion this year. Patel expects OpenAI and Anthropic together to reach $600-700 billion by December next year and $1.5 trillion by the end of 2027. If AI revenue is only $3 trillion in 2031, he concedes, there will be big problems. Revenue always trails capex, because today's revenue rests on infrastructure bought in prior years. Anthropic is already profitable once training and inference compute are subtracted. OpenAI follows in two or three quarters. Patel says Amazon's AI infrastructure is profitable today, and Bedrock lifted its gross margin last quarter.

Where does that money come from? Patel says customers' R&D and cost-of-goods budgets shift toward AI, and part of it comes out of labor. Labor's share of GDP has fallen since the 1970s. AI tokens are "basically just capital" doing the work of labor, so the trend accelerates. He hopes, and admits it may be an irrational belief, that the rest of the economy grows fast enough that labor dollars stay flat or even rise. His model is chip design. The number of American chip-design engineers has been flat for 20 years, while the value created by designers at Nvidia, Broadcom and Google exploded. Better tools improved every chip, from phones to sensors. Same headcount, more revenue, better margins. He expects the pattern to repeat in pharma, trading and coding. Nanos stretches the idea further: AI revenue will not stay a per-token line, because labs can bring a cancer drug or a self-driving vehicle to market.

The business behind that revenue is better than skeptics assume. Salesforce, among the best SaaS companies, runs about 75% gross margins and took years to get there. Anthropic is at 75% or a little higher on inference, reached faster. SaaS companies can spend as much as 30% on customer acquisition. Anthropic spends almost nothing, and its R&D has grown slower than revenue. Customers are sticky: revenue has been a straight line up, and a better Chinese model barely shifts existing customers' spend. Patel calls it "arguably a better business than software".

The weak points are in the plumbing. Nvidia is backstopping $588 billion of off-balance-sheet commitments, about 2.5% of the US M2 money supply, against annual revenue of roughly $400-500 billion. Nanos reads that as pragmatism. Nvidia wants customers beyond three or four hyperscalers, so it acts as an investment-grade offtaker that lets neoclouds raise debt. When someone wants the GPUs at a higher price, Nvidia resells them rather than using them for its own research. The guarantees reach construction, memory suppliers and fab builders too.

Security is thinner. SemiAnalysis published a report arguing most neoclouds are poor at it, and about eight hours later Ilya Sutskever warned that rogue agents might try to take over a neocloud to run more copies. Patel says open-source Chinese models can already hack dozens of them. At one cloud his team accidentally saw other tenants' storage and Slurm jobs, including a government agency from a top-10 GDP country. Oracle shows a third failure mode. Its cluster quality ranks near the top, yet its New Mexico site for OpenAI slipped on a pipeline problem SemiAnalysis flagged from April, and Oracle eventually declared force majeure. Its OpenAI deals carry fixed margins, so it cannot capture higher GPU prices. Nebius, with the shortest contracts among the top clouds, repriced upward.

The hardest objection is that the success case is the dangerous one. Patel is bullish enough to think Anthropic could be the first $10 trillion company. But at 20x revenue and $100 billion of sales, an IPO lands near $2 trillion. Reaching $10 trillion likely takes about $1 trillion of revenue, possible in 2028 or 2029. A $20 trillion company would be pulling in $2-4 trillion, which means AI is extraordinarily capable and the gains concentrate among stockholders and their supply chains. That could push distressed people to "overthrow governments". He frames this as the pessimistic picture and says he is optimistic. Nanos is more so. He sees revenue coming from drugs and products rather than metered tokens, with bumps along the way.

The labs are also not releasing everything. Anthropic had Mythos ready in February and the public still has a reduced version. OpenAI sat on Astra for a couple of months and has shown results for a newer model it has not released. Months of progress stay internal because releasing invites distillation. Nanos says only Anthropic is publicly pacing the frontier, and nobody else is slowing. Patel sees little room to slow: each lab distrusts the other, and China is not far behind.

Two questions stay open. Nobody has found the limit of demand for intelligence, so nobody can say who must slow, when or where. And the benign reading of the labor story rests on a hope Patel himself flags as possibly irrational. If the 10x revenue years continue, the plumbing is the risk. If they stall, the long contracts and the backstops are where the strain shows.