For three years, Mark Zuckerberg asked investors to take his AI spending on faith. On July 1, he offered them something closer to a receipt.
Bloomberg reported that morning that Meta Platforms is standing up a cloud infrastructure business — internally dubbed Meta Compute — to sell access to its AI training and inference capacity, and to models hosted on its own silicon, to outside customers. The market's verdict arrived within the hour, and it was violent in both directions. Meta shares jumped roughly 9%, closing near $613. CoreWeave fell about 14%. Nebius fell about 17%, shedding roughly $12 billion of market value in a session. IREN dropped too.
The reason for the split is simple arithmetic. Meta has spent the AI boom as one of the largest buyers of GPU capacity on earth. If it becomes a seller, the supply curve moves.
What Meta is actually building
The reporting describes two layers. The first is a hosted model service — enterprises reaching AI models through an API, in the manner of Amazon's Bedrock. Meta has reportedly considered offering not just its open-weight Llama family but Muse Spark, the closed-weight model it launched earlier this year and has so far deployed mostly inside its own products. The second layer is rawer and more disruptive: bare-metal GPU capacity sold on price and availability, which is precisely the neocloud business that CoreWeave and Nebius occupy.
The effort is led by Santosh Janardhan, Meta's head of infrastructure, alongside Meta Superintelligence Labs leader Daniel Gross and company president Dina Powell McCormick — a trio that reads as infrastructure, research and enterprise sales, which is roughly the shape of a cloud org chart.
Zuckerberg has been circling this publicly for months. At Meta's May shareholder meeting he said a cloud business was on the table, telling investors that companies were approaching Meta almost every week to buy model access or spare compute at a premium. On July 9, per Bloomberg, he went further and said selling compute “makes sense” — while pointedly adding that Meta has not overbuilt, is currently using all of its available capacity, and still sees substantial internal demand.
That last caveat matters, and it is where the reporting gets thin. As of mid-July there is no launched product, no published pricing, no named customer and no announced revenue target. Meta has not issued a formal product announcement; the story rests on Bloomberg's sourcing plus the CEO's confirmation that the idea is under consideration. Several outlets have described Meta Compute as “unveiled” or “launched.” On the available evidence, “under construction” is the more defensible verb.
The financial figures are similarly contested. Meta's 2026 capital expenditure has been reported at $115-135 billion, raised to $125-145 billion, and estimated by some analysts at $145 billion outright; Morgan Stanley's Brian Nowak pegs $145 billion as the 2025 figure and models $175 billion for 2027. Meta's own Q1 filing disclosed $182.9 billion in committed future AI infrastructure spend. The numbers point the same direction — up, enormously — but they are not interchangeable, and anyone quoting a single one should say whose it is.
Analysis: a supply-side shock with an asterisk
The bear case for the neoclouds is that Meta changes the competitive math overnight. A company with Meta's balance sheet, custom MTIA silicon and a Broadcom chip agreement running to 2029 does not need the business to be profitable in year one. It only needs idle clusters to stop being idle.
The bull case, argued most sharply by Rosenblatt's John McPeake — who reiterated a street-high $250 target on CoreWeave after the news — rests on three planks. Meta's contract with CoreWeave, which he sizes at $35.2 billion, reportedly bars Meta from reselling or subleasing rented GPU capacity, meaning Meta Compute cannot be built on borrowed inventory. Global GPU scarcity still lets specialists price. And Meta Compute looks like utility optimisation — monetising spare cycles and quieting return-on-capital critics — rather than a war of extermination.
Morgan Stanley lands in a similar place from the other side of the trade. Nowak argues Meta is likelier to build a neocloud than a full hyperscaler, and estimates that leasing 250 megawatts at $40 per watt could add roughly 8% to Meta's 2028 EPS. He kept a Buy and a $775 target. Wedbush is cooler, holding Neutral at $671 on doubts about AI returns.
There is a structural wrinkle nobody has resolved. Cloud is a service business — contracts, uptime guarantees, support engineers, enterprise procurement cycles. Meta has never run one. It still draws about 98% of revenue from advertising, a business with margins no infrastructure rental will match. Wall Street cheered the diversification on July 1 and spent July 2 recalculating blended margins.
And there is the irony: Meta is one of the neoclouds' largest customers, with commitments reported at $21 billion to CoreWeave through 2032 and roughly $27 billion to Nebius. The company that would compete with them is currently underwriting them.
What to watch next
Three things. First, Meta's Q2 earnings call — whether management uses the words “Meta Compute” on the record, and whether it attaches any revenue framing or capex reallocation to them. Second, the first named external customer, which converts a Bloomberg story into a business. Third, contract language: if the no-sublease firewall in the CoreWeave agreement is real and holds, Meta Compute can only sell capacity Meta owns outright, which caps the near-term supply shock considerably.
Until then, the cleanest reading is the one Zuckerberg himself offered. Meta is not sitting on surplus GPUs today. It is betting that at tens of gigawatts of scale, it eventually will be — and that when that day comes, it would rather be a landlord than a tenant.