One Word, 'Excess', Sank the AI Landlords. Meta's Own Contracts Point the Other Way

On the morning of 1 July 2026, Bloomberg reported that Meta was building a cloud business, internally styled 'Meta ComputeComputeShorthand for raw computing power: the chips, machines, and electricity needed to train and run AI models.,' to resell excess AI computing power. The market finished the sentence on its own. If Meta, one of the largest renters of AI chips on earth, was about to turn around and sell computing power, then it was about to compete with the very firms it rents from.

Those firms are neoclouds: companies like CoreWeave and Nebius that buy vast fleets of Nvidia AI chips and rent them out by the hour. By the closing bell CoreWeave had fallen about 13.9% (to a reported $85.68) and Nebius about 17% (to a reported $229.18). Meta, the supposed aggressor, rose about 10%. We are deliberately not putting a single-session dollar figure on those drops: doing so honestly requires each firm's share count at the moment of the move, which we could not verify against filings in this session, and a percentage applied to a post-crash market value understates the loss. The percentages are the reported facts; we reason from them.

Strip away the noise and the market made a single, testable claim: Meta has more computing power than it needs and is about to flood the market with it. So the narrow question is whether the selloff priced a glut that Meta does not actually have.

The question has a sharp edge because it is falsifiable, it can be proven wrong by evidence. If Meta were truly awash in idle chips, we would expect to see it winding down outside contracts, cutting its building budget, or running its own data centers at low utilizationutilizationHow much of a company's computing capacity is actually being used; low utilization would mean genuine idle capacity to resell., the share of capacity actually in use. If instead Meta is signing for more outside capacity, the 'excess' story collapses and the selloff needs a different justification.

The sequence matters because the story mutated over five days. Bloomberg's 1 July report lit the fuse; in the same session, CoreWeave and Nebius fell double digits and smaller AI-infrastructure names were dragged along. By 2 July, CNBC had reframed the move as one Wall Street could like even 'despite lower margins,' shifting the bear case from 'Meta abandons its suppliers' to 'Meta squeezes their profitability.' Also on 2 July, the research boutique SemiAnalysis called the neocloudneocloudA new-style cloud company whose product is raw AI computing power, it buys fleets of Nvidia chips and rents them out by the hour; CoreWeave and Nebius are the two in this story. fears 'erroneous,' pointing to Meta's capital spending set to surge, not shrink.

By 6 July the tape showed a revealing split: Meta higher, the neocloud complex still re-rated lower. When a customer's stock rises while its suppliers' fall on the same news, the market is not voting on AI demand overall. It is re-pricing who captures the profit.

A supplier dumping a glut would drag its whole market down. Here the customer went up. That is not a demand story; it is a margin story.

This is the cleanest signal in the episode, and it uses only the reported percentage moves, no inferred dollar totals. If investors truly believed a compute glut was arriving, they would mark down the buyers of compute too, because a glut means cheap chips and weaker pricing everywhere. Instead they bid Meta up. The market's own reflex says it does not believe in a demand collapse. What it is re-pricing is the division of spoils between the landlord and the tenant.

Start with the contracts, because contracts are harder to spin than headlines. Meta's total disclosed commitment to CoreWeave runs to roughly $35 billion: an original deal worth about $14.2 billion signed in 2025, then a further roughly $21 billion expansion announced on 9 April 2026 that runs through December 2032. OpenAI's commitment to CoreWeave is about $22.4 billion. Both sit inside a CoreWeave revenue backlogbacklogThe total value of orders a company has signed but not yet delivered, a rough measure of future revenue already under contract. reported at about $55.6 billion in its third-quarter 2025 results, orders signed but not yet delivered. Nebius, for its part, holds a Meta agreement worth up to $27 billion over five years.

Look at the direction of travel, not just the size. The pivotal fact is the date: on 9 April 2026, less than three months before the selloff, Meta expanded its CoreWeave deal by about $21 billion rather than trimming it. A company sitting on surplus compute does not go out and lock in another $21 billion of future outside capacity through 2032. That is the behavior of a buyer that cannot get enough, not one with a glut to unload.

The independent research shop SemiAnalysis, which sells detailed data-center supply analysis to investors, reached the same conclusion from the supply side. It estimated that in just the first half of 2026 Meta contracted over 5 gigawatts of external capacity, cloud and colocation combined, and before counting its own accelerating self-build, and called the market's cannibalization fear 'erroneous,' projecting that Meta's 2027 capital spending would be 'shockingly high.' These are outside estimates, not Meta disclosures, so we treat them as corroboration rather than proof. But they cut the same way as the contracts: procurement is accelerating.

The single hardest number to argue with is Meta's own spending guidance. At its first-quarter 2026 earnings in late April, Meta raised its 2026 capital-expenditure forecast to $125-145 billion, up from $115-135 billion a quarter earlier, citing higher component prices and added data-center costs. Its stock fell more than 6% after hours, not because investors feared Meta had too much compute, but because they blanched at how much more it intended to buy. A firm drowning in excess chips does not guide spending up and spook its own shareholders with the size of the bill. That reaction is the fingerprint of scarcity, not surplus.

Here is where honesty demands a turn. Showing that Meta is not leaving as a customer does not make the selloff irrational. There is a second, more durable bear case, and the contracts do not refute it. Even if Meta keeps buying, a giant like Meta credibly entering the business of selling compute adds a fourth deep-pocketed supplier to a market whose sky-high valuations rested on scarcity. More suppliers means less pricing power, which means the fat margins these firms were valued on could thin over time.

The demand fear can be reaffirmed away. The margin fear cannot. That is why the two stories resolve on different clocks.

This case has a foothold in the numbers, and crucially the pressure predates the Meta headline. CoreWeave's adjusted EBITDAadjusted EBITDAA company's own version of that operating-profit measure with certain one-off or non-cash items stripped out; useful for spotting trends but not a standardized accounting figure. margin, the slice of each sales dollar left as operating profit before interest, taxes, and wear-and-tear, had already compressed to about 56% from 62% a year earlier by the first quarter of 2026. Analysts at Bernstein called Meta's cloud move 'problematic' for CoreWeave. And CNBC's framing, buy Meta 'despite lower margins', is really an admission that the market expects the whole sector's profitability to be squeezed, whoever the buyers are.

The two stories are easy to confuse but resolve differently. The demand-panic story, Meta abandons its suppliers, is the one that would round-trip: a single reassuring capex line or a reaffirmed contract can undo it, exactly as it should, because the contracts already contradict it. The margin story, a fourth big supplier permanently thins the terminal profit pool, cannot be reaffirmed away, because it is a claim about the shape of the industry three and five years out. If it is right, the re-rating sticks.

We have seen the pattern's first half before. In February 2025, TD Cowen analysts reported that Microsoft had walked away from data-center leases; AI-infrastructure stocks sold off, with CoreWeave, then heavily dependent on Microsoft, caught in the downdraft on the eve of its IPO. The scare reversed within weeks: Microsoft reaffirmed its plan to spend more than $80 billion on AI capacity that year, and CoreWeave said no contracts had been cancelled. The demand fear round-tripped.

But the rhyme is not a repeat, and the difference is the whole point. In 2025 the trigger was a rumor about a customer leaving, the kind of thing a company can simply deny or reverse. In 2026 the trigger is a customer building the product it currently buys. Meta cannot 'reaffirm' its way out of having entered the compute-selling business. So the part of this selloff that mirrors 2025, the fear that Meta walks, should fade the same way. The part that does not have a 2025 analog, a new structural competitor, may not.

This is where the tape stops being useful and the filings take over. Our falsifiable claim, that there is no glut, resolves on Meta's next quarterly report. If Meta begins winding down its neocloud contracts, cuts its capex guidance, or reveals it is running its own fleet at low utilization, the 'excess' story earns its keep and we were wrong. Nothing in the evidence available today points that way; everything points the other.

The margin question resolves on a different document: the neoclouds' own next 10-Qs. Watch CoreWeave's and Nebius's gross and EBITDA margins and any change in pricing. If margins keep compressing as new supply arrives, the re-ratingre-ratingWhen investors change the multiple they will pay for a company's earnings, moving the stock even if this quarter's numbers do not change. was foresight, not panic, and it will not round-tripround-tripWhen a price moves sharply on news and then returns to roughly where it started once the news is understood, a reaction that reverses.. If they hold, the market over-read a headline. Either way, the answer will be written in black and white in a filing, not in a single day's price move. The one thing this week's tape has already settled is the narrow question we started with: the selloff priced a supply glut that Meta, on every disclosed number, does not have.