Oh boy. Elon Musk says that “debt markets have almost entirely shut off access to money” for companies without substantial revenue outside AI data centers. He expects a wipeout among the 200-plus neoclouds buying GPUs and renting them out.
Some of those companies will absolutely blow up. Borrowing billions against hardware that depreciates quickly, then waiting on power connections while the interest clock runs, is a good way to turn a great technology into a terrible stock. But three days before Musk’s post, Crusoe announced a $700 million financing facility for power infrastructure serving its AI data centers. That is different from an unsecured bond issue or a GPU loan. It also makes “almost entirely shut off” a pretty sweeping description of capital available to the sector.
But Musk’s broader neocloud obituary is almost certainly early. He is treating a financing problem as though it were a demand problem. A lender can lose money on a data center while the GPUs inside it remain valuable to someone else. The equity can disappear. The customer contract doesn’t have to.
The meter is still running
AI compute is expensive to rent because people keep renting it. Lambda’s posted price for a single H100 PCIe instance is $3.29 per GPU-hour. Run that GPU around the clock and the list-rate bill is about $28,800 a year. A hundred thousand would cost roughly $2.9 billion a year at the same rate. That’s an illustration, not a quote for a fleet-sized contract; big buyers negotiate, and configurations, storage, networking, and utilization change the bill. But this is a real market with real prices.
In September, CoreWeave said it had signed new three-to-six-month capacity contracts at roughly $40 million per megawatt per year, expressed as annualized rental revenue. Its June revenue backlog was about $104 billion, with more than $25 billion in additional customer commitments announced early in the third quarter. Backlog is no substitute for cash, and building the capacity on schedule matters. Still, customers are signing up to pay for the compute.
The stronger operators also have counterparties that can actually write the checks. Meta committed about $21 billion for CoreWeave capacity through 2032. An SEC filing says NVIDIA agreed to buy CoreWeave’s residual unsold capacity under an arrangement with an initial value of $6.3 billion, subject to delivery and availability conditions. That backstop does not extend to every neocloud. It does tell you why treating all 200-plus operators as one undifferentiated pile of debt is lazy analysis.
Musk knows compute can find another buyer. Anthropic’s new S-1 puts a number on it: agreements with SpaceXAI call for up to $120.5 billion of spending through 2029, or about $3.1 billion per month once the full capacity is delivered, which Anthropic expects in December. “Up to” matters. After initial lock-ins ending in January and March 2027, either side can cancel the respective agreements with 90 days’ notice. This is a very large customer for Musk’s compute, with renewal risk attached. Funny how that works.
Anthropic’s filing describes the market better than Musk’s post
In the first half of 2026, Anthropic recorded $16.3 billion of revenue and $8.9 billion of compute and infrastructure expense. A year earlier, those figures were $1.2 billion and $2.3 billion. This company is turning compute into revenue much faster than its compute bill is growing. It also reported an operating loss for the half, so nobody gets to declare the economics solved.
The filing says Anthropic has historically underestimated its compute needs and run short of capacity. It also warns that suppliers require commitments months or years ahead of use, so demand can arrive late and leave paid-for capacity idle. Both can be true! Buying compute here means underwriting a forecast of model progress, customer adoption, power delivery, chip supply, and the rate at which better algorithms make a GPU more useful. That’s a hard forecast even if you have a very large checkbook.
Anthropic is responding by buying flexibility across suppliers and time. It calls its shift toward directly leased chips in dedicated facilities “Anthropic Direct”. It can reserve the data center earlier, then commit to the more expensive accelerators once demand is clearer. Its five-year Broadcom equipment leases carry about $161.2 billion in aggregate rent before interest. That is serious long-term demand for infrastructure and a serious fixed obligation for Anthropic. It also means today’s AI labs are learning to procure capacity themselves instead of relying entirely on a neocloud to do it for them.
The payoff is uneven and still being discovered
The hyperscalers are already seeing benefits. Microsoft reported 43% year-over-year Azure growth in its June quarter and said customer demand exceeded available capacity. Extra GPU and CPU capacity brought online during the quarter was monetized quickly. Amazon said AWS AI revenue had passed a $25 billion annual run rate. Meta’s June-quarter revenue rose 28%, though its operating margin fell as spending ran ahead of some of the payoff. These are different businesses with different accounting. They are seeing demand and paying heavily to serve it.
Meanwhile, we are still discovering what people will do with all this intelligence. The returns are lumpy. An a16z analysis of observed U.S. consumer spending found the top 1% of AI payers generated 19.5% of spending, more than the bottom half combined. That is spending, not a measure of productivity, but it shows how concentrated the early market is. In a study of customer-support agents, AI lifted productivity 14% on average, with the biggest gains among novice workers and little benefit for the most experienced. One firm, one use case, very different payoffs.
I think AI’s benefits are power-law distributed across companies and use cases, and we have barely begun to discover them. That makes aggregate ROI hard to measure right now: a single economy-wide number will miss the early winners. A handful of uses are creating disproportionate value while others are still experiments. This is the innovation expansion phase; the world is up for grabs.
Everyone knows the game
Microsoft, Amazon, Google, and Meta would each love to keep its margins and let the others pay to discover the next platform. None can risk sitting out while a rival owns it. So they build, rent, and bid against one another. If all of them overinvest, prices for intelligence fall and customers get the surplus. Their own returns can look muted until the market shakes out and a smaller group has pricing power. That’s the prisoner’s dilemma in this buildout (and, for once, the prisoners have extremely expensive GPUs).
Anthropic describes some of its cloud suppliers as investors, customers, and competitors at the same time. That’s the game in one sentence. Every player wants access to the next platform and a share of the rent. None knows which model, chip, or application will deserve it yet.
Some lenders will be right to pull back. Some neocloud equity will be wiped out. We may get stranded buildings waiting for power. Those outcomes don’t imply that useful compute vanishes or that the hyperscalers stop buying it. The assets, contracts, and economics have to be examined company by company.
2028, 2031, 2034
Two years out, in 2028, I expect a messy mix of shortages and stranded capacity. The bottleneck will often be energized, working clusters, not a shortage of GPU purchase orders. Anthropic’s SpaceXAI contract can be canceled after its short initial lock-in, while its additional TPU capacity is scheduled for delivery from 2028 through 2031. A neocloud with power, uptime, and a durable customer contract can charge for those advantages. One with a half-built site and short rental commitments may get crushed even if AI demand keeps rising.
Five years out, in 2031, my bet is that the biggest AI labs buy more of their own chips, reserve data center capacity directly, and spread workloads across hardware vendors. Anthropic estimates another $125 billion to $150 billion in payments for TPU compute to be delivered between 2028 and 2031, dependent on deployment and pricing terms. Those estimated payments are outside the filing’s contractual-commitments table; they are not guaranteed revenue for one supplier. The plain GPU-rental middleman has to offer something beyond a financing wrapper: power, fast deployment, excellent operations, or the ability to absorb demand spikes. The most valuable operators survive. The undifferentiated ones get acquired, refinanced, or erased.
Eight years out, in 2034, compute is likely a much larger, more ordinary input to the economy, with much more intelligence produced per dollar of hardware. Anthropic has Lambda agreements extending into 2034 with approximately $41.3 billion of spending as capacity is delivered and accepted, and a minimum $110 billion Amazon commitment extending into 2036 that applies even if usage falls short, subject to the agreement’s exceptions. Those terms tell us demand is being planned on a long horizon. They do not tell us what a 2026 GPU will rent for in 2034. I expect the model and infrastructure markets to be more concentrated by then, but I would not promise oligopoly profits to today’s buyers. Efficiency may create new demand faster than it cuts prices, or it may strand expensive commitments. Either way, the consumer gets a lot of the benefit while these companies fight for the rest.
Musk is making this argument from inside the game. SpaceX acquired xAI, and its own compute now has a potentially multibillion-dollar-per-month customer in Anthropic. I read that acquisition as a bailout of a model business losing ground, not as proof that its hardware is worthless. Consumer market data puts ChatGPT, Gemini, and Claude at the center of paid adoption. Grok is playing from behind. Musk’s warning may be strategically useful to a competitor trying to shape the terms of the shakeout, but it is not disinterested market analysis.
You’d have to be an idiot to read it that way.
Grok is a loser. His compute is not.
Disclosure: This article is for education and general information, not individualized investment advice. AI infrastructure investments involve financing, execution, technology, and demand risks.