On Wednesday July 1, 2026, Bloomberg and CNBC reported, within hours of each other, that Meta Platforms is quietly assembling a cloud infrastructure business — internally called Meta Compute — to rent excess AI capacity from its own data centres to outside customers. Meta has not confirmed the plan. Zuckerberg's public position, per Silicon Republic, is that "we haven't done that yet because we think that we have a use for the compute. Obviously if we get to a point where we feel that we have overbuilt, then that is an option."

Meta shares closed up 8.8% on the day, per CloudComputing News, with Silicon Republic marking the peak at 9.3%. CoreWeave dropped between 10.8% and 14%. Nebius Group fell 17%, per SiliconANGLE. The two neoclouds lost roughly a quarter of their combined market capitalisation on a story about a business Meta will not publicly say exists.

That is what a rumour is worth when the rumour is that the biggest single GPU customer in the world is about to become a GPU landlord.

Two service models on the whiteboard

Per Bloomberg's reporting, summarised by CloudComputing News and Silicon Republic, Meta is weighing two shapes for Meta Compute. The first is a hosted-model API: outside developers get access to Meta's own Muse Spark closed-weight model on Meta infrastructure, in an arrangement structurally identical to AWS Bedrock or Google Vertex. The second is raw GPU rental — the CoreWeave and Nebius playbook — where you buy compute by the hour and bring your own model.

The reason the market moved so hard is that Meta could reasonably do both. It has the fleet for the raw-compute play and the model for the hosted-API play, and the fleet is not a subscale one. Meta's own Hyperion campus in Richland Parish, Louisiana — a 2,250-acre, 5-gigawatt site being built to train the next Llama generation, per IEEE Spectrum — is on the order of the largest data-centre complexes in the United States. The Ohio site, per Silicon Republic, is expected to come online this year. If any of that capacity flips from internal-only to rent-out-what-we-don't-need, the neocloud market ships from a shortage story to a glut story overnight.

Leadership is not being run out of a corporate-development back office. Per CloudComputing News, the initiative is being led by three people whose combined remit covers most of what a cloud business needs to have working on day one: Santosh Janardhan, Meta's head of infrastructure; Daniel Gross, head of Meta Superintelligence Labs; and Dina Powell McCormick, Meta's president. That is an infra chief, a model chief, and a policy-and-partnerships chief. It is not a skunkworks. It is the roster of an actual product.

The capex number that made the rumour credible

The reason the market believed the story is the balance sheet. Meta has told investors it will spend up to $145 billion on capital expenditures in the current fiscal year, per SiliconANGLE — a step up from $70 billion in 2025. As of March 31, 2026, Meta's future lease obligations for data-centre capacity stood at $182.9 billion, per CloudComputing News. Those are hyperscaler numbers on a company whose only paying external customer for compute, at the moment, is a shell that pays itself.

That kind of buildout has to earn something back. Meta's advertising business is a large but not infinite sink for internal AI inference. Every additional gigawatt beyond what ads and Reels and Threads can consume is either revenue or write-down. Meta Compute is the revenue answer, and the analyst class has been waiting for it to be publicly on the table for the last two quarters.

That is the read behind the 8.8% pop. The market is not pricing in a new business line. It is pricing in the end of the question of whether the AI capex was going to pencil out.

The hedge in Zuckerberg's quote

The Silicon Republic version of Zuckerberg's on-record statement is a masterclass in optionality. "We haven't done that yet because we think that we have a use for the compute. Obviously if we get to a point where we feel that we have overbuilt, then that is an option." Every clause in that sentence buys Meta a way out. It hasn't launched. It thinks it needs the compute. If it turns out it doesn't, it might launch. Three prepositional phrases, no commitment, no timeline.

Read it a different way and it is the on-record version of the leak. The rumour says Meta is building a cloud. The CEO says if we overbuild, we might sell it. Those are the same sentence spoken with different vocabularies. The one that ran in Bloomberg is the version the market priced. The one Zuckerberg said out loud is the version the SEC will accept.

Meta's other on-record framing, quoted by CloudComputing News, is "it's definitely on the table." That is the phrase corporate-affairs uses when the answer is yes, but we would like six more months of denial.

What actually crashed on the neocloud desks

The 14% CoreWeave drop and 17% Nebius drop are the most interesting part of the day's tape, and they are worth reading carefully. CoreWeave and Nebius are not built to be diversified compute businesses. Their thesis is that hyperscalers need more capacity than they can build, so they lease from the neoclouds, at margins the neoclouds can grow into. That thesis dies the day one of the hyperscalers becomes a fifth neocloud with a bigger fleet and a first-party model to bundle.

Meta Compute does not need to be very good to hurt CoreWeave. It needs to exist. The reason is that CoreWeave's forward multiples are priced on the assumption that hyperscaler demand outstrips hyperscaler build. If that assumption stops holding — if Meta has genuinely overbuilt, and if AWS or Google ever quietly follow — the neocloud comparables reset from scarce infrastructure premium to commodity GPU rental. That is the trade the market started pricing on July 1.

The interesting question is whether CoreWeave's own long-term contracts survive that reset. Per the CoreWeave 8-K filed in June, the company has an expanded agreement with Meta itself, worth approximately $21 billion through December 2032. Meta is a customer of the neocloud it is about to compete against. That is the kind of overlap that ends in a footnote in a 10-K explaining why one line went to zero.

What this means

1. The customer became the landlord. Meta spent five years buying more GPUs than any single company in the world, and the market has spent the last eighteen months asking when that capex would monetise beyond ads. Meta Compute is the answer, or rather the answer's rumour. The nine-percent pop is what monetisation-clarity buys you when the base rate has been "trust the capex."

2. The neocloud thesis is now conditional on hyperscaler discipline. CoreWeave and Nebius are priced on the bet that AWS, Azure, Google, and Meta will keep buying more compute than they can absorb. Meta just told the tape that hyperscalers can absorb their own compute after all. That is a category re-rating, not a Meta-versus-CoreWeave trade.

3. The two-service-models question is the real product decision. Hosted models (a Bedrock clone with Muse Spark inside) and raw GPU rental (a CoreWeave clone) have very different competitive shapes. The first fights AWS on model catalogue. The second fights CoreWeave on price per H100-hour. Meta has the balance sheet to do both, and the Janardhan/Gross/McCormick trio suggests it plans to. Which one launches first will tell you whether Meta thinks the money is in developer mindshare or in wholesale capacity.

4. The Zuckerberg quote is the leak. "If we get to a point where we feel that we have overbuilt, then that is an option" is not a denial. It is a public confirmation with a delay dial. The launch date is whenever Meta's investor-relations desk decides overbuilt is defensible to say in an earnings call. That call is on July 30.

The story the market told on Wednesday is that four hyperscalers, not three, will sell AI compute by the token. The story the CEO told is that maybe, someday, if. Both stories are compatible with the same product being on Santosh Janardhan's whiteboard right now. The one that gets confirmed first is the one whose price the market already paid.