Meta’s Cloud Option Collides With the GPUs It Rents

Meta is hiring an Amazon cloud veteran and taking calls to rent spare GPUs, yet it still pays tens of billions for other people’s chips. Mark Zuckerberg told shareholders on May 27 that a public cloud was “definitely on the table” if the company overbuilt.

The hedge is being staffed. The irony is that Meta remains a giant renter of the same class of capacity, which is why a Meta cloud business would hit its suppliers long before it looked like Microsoft Azure.

Zuckerberg Put a Public Cloud on the Table in May

The line came at Meta’s virtual annual meeting, after the board had already pointed 2026 capital spending at AI data centers. Outside firms, he said, were not waiting for a product page.

Almost every week there are different companies that come to us from outside, asking us to both stand up an API service or asking if we have compute that they could buy from us, at some premium to what we bought it at. We haven’t done that yet because we think that we have a use for the compute, but obviously if we get to a point where we feel that we have overbuilt, then that is an option that we have, and that is partially what gives us confidence in investing in building this out.

Mark Zuckerberg, CEO, Meta 2026 annual meeting

That is an overflow valve, not a pledge to copy Azure’s catalog. An earlier read on Meta’s cloud interest treated the May comments as the opening bid. The summer then added a hire, a second earnings-call hint, and a stack of GPU contracts that still run the other way.

On the July 29 call Zuckerberg said Meta was “getting a lot of offers for compute at a significant premium over what we paid for it.” He also said the company expects “to grow a large business serving large customers,” with an API and business agents already in motion. He did not describe a regions-and-zones cloud with a price list.

The Same Company Just Signed Tens of Billions in GPU Rentals

The option only works if Meta has spare silicon. Through the spring it was still writing checks to the firms that exist to rent that silicon out.

META AS A GPU CUSTOMER

CoreWeave’s April 9 statement put the expanded $21 billion CoreWeave capacity deal on Nvidia’s Vera Rubin platform, spread across sites, and aimed at inference. Add the September 2025 contract and the disclosed CoreWeave relationship is $35.2 billion. Nebius described a five-year structure that starts at $12 billion of dedicated kit and lets Meta take up to $15 billion more if other customers do not.

This is another example that leading companies are choosing CoreWeave’s AI cloud to run their most demanding workloads.

Michael Intrator, CEO, CoreWeave, April 9 statement

Those lines do not age well if Meta turns around and sells the same class of hours. Gil Luria, a managing director at D.A. Davidson, put the first-order hit on the neoclouds rather than on AWS or Azure, because CoreWeave and Nebius still need Meta as a growth account. The contracts also undercut the clean story that Meta is simply flushing extra GPUs. It is buying capacity through 2032 while talking about renting its own.

Dave Brown Was Hired to Run the Buildout

In July, people familiar with the matter said Dave Brown would leave Amazon Web Services for Meta. Brown spent 19 years at AWS, sat on the S-team that advises Andy Jassy, and ran compute and machine-learning services, a line of work that traces back to early EC2. Amazon confirmed the departure in a note from AWS chief Matt Garman and named Dave Treadwell, a longtime Amazon executive, to take the compute and ML unit on August 1.

Brown was expected to report to Santosh Janardhan, Meta’s head of infrastructure, and to work the data-center build, not a named sales org. That is the tell. Meta created Meta Compute in January as a top-level push Zuckerberg described as tens of gigawatts this decade. The AWS veteran Dave Brown to Meta hire puts an operator who has shipped rented compute at industrial scale onto that program.

It does not, by itself, create the rest of a cloud company. Billing, quotas, identity, support queues, and multi-year enterprise contracts are a different muscle from standing up clusters for Facebook, Instagram, WhatsApp, and Meta’s own models. Brown can shorten the path from rack to rented hour. He cannot mint a fifteen-year Azure field organization in a quarter.

What Azure Has That a GPU Rental Shop Does Not

Microsoft Azure already sells the boring layer Meta would have to invent: identity tied to Office and security tooling, region choice, isolation options, and the procurement path governments and banks actually use. Synergy Research Group’s second-quarter tally is the scale of that market, not a forecast of Meta’s share of it.

CLOUD INFRASTRUCTURE, Q2 2026

• The pool: Q2 cloud spend hit $143.4 billion, up more than $43 billion or 43% from a year earlier, the 11th straight quarter the year-over-year rate accelerated.
• The split: AWS held 28%, Azure 20%, and Google Cloud 15%, or 63% together; in public IaaS and PaaS the top three held 67%.
• The AI slice: Generative-AI-specific cloud services grew 165% year over year, while public IaaS and PaaS as a whole grew 47%.
• The rest: Trailing twelve-month cloud infrastructure revenue reached $500 billion, and Synergy’s faster “tier two” names include CoreWeave, Oracle, Crusoe, Nebius, Anthropic, and Nscale.

John Dinsdale, Synergy’s chief analyst, said AI had “lit a fire under the cloud market.” That fire is exactly the demand Zuckerberg keeps citing. It is also the reason Azure’s 20% is not a sitting duck. The product customers pay Microsoft for is not a pile of GPUs. It is a contract they can defend to a risk committee.

WHAT META STILL HAS TO BUILD

• A sales motion: Long-cycle enterprise deals, credits, and named-account coverage, none of which Meta runs as a hyperscaler today.
• Isolation and residency: Single-tenant patterns and data-sovereignty designs that consumer pipelines are not built to sell.
• The paper: The audit reports and certifications financial and government buyers treat as table stakes.

A shop that sells leftover training hours with a clawback clause can skip most of that list. A company that wants Azure’s customers cannot. Meta can put GPUs on the market faster than it can become Azure, which is why the first collision is with the neoclouds already on its vendor sheet, not with Microsoft’s install base.

Meta Raised Capex Guidance Twice in 2026

The cloud talk is the political answer to a spending line that keeps moving up. In the first quarter Meta spent $19.8 billion on capital projects, including principal payments on finance leases, against $12.4 billion of free cash flow. Susan Li, the CFO, also flagged a $107 billion step-up in contractual commitments from multi-year cloud deals and infrastructure purchase agreements. Zuckerberg said Meta was rolling out more than one gigawatt of custom silicon with Broadcom, plus AMD parts beside new Nvidia systems.

THE 2026 CAPEX RAISES

1. Start of 2026: Guides full-year capital spending, including finance-lease principal, to $115 billion to $135 billion.
2. April 29, 2026: Raised 2026 capex to $125-145 billion, pointing to higher component pricing, especially memory, and more data-center cost for later years.
3. July 29, 2026: Lifts the floor to $130 billion and keeps the ceiling at $145 billion after free cash flow for the quarter fell to $784 million from $8.5 billion a year earlier.

Second-quarter revenue was $60.8 billion, up 28%. Net income was $15.8 billion, down 14%. Shares still dropped about 10% on the cash-flow print. “I get that this is a big investment and it’s a big bet,” Zuckerberg said on the call. The cloud option is how that bet is supposed to look less one-sided if the clusters outrun Meta AI, ads ranking, and the agent products.

The campuses behind the guide are not a brochure. Hyperion in Louisiana is now described as a $50 billion, 5-gigawatt project. Prometheus in Ohio is a 1-gigawatt cluster planned for 2026. A separate 1-gigawatt site in El Paso, Texas, is a venture with BlackRock that Meta put at more than $10 billion. Those are training factories. Turning any slice of them into a public cloud means metering, tenancy, and a promise that a tenant’s job will not vanish when Meta’s own models need the rack.

Meta One Starts Charging for Compute at $7.99

The near-term cash product is not a VM. On May 27, the same day as the shareholder meeting, Meta said it would test paid Meta One AI subscription tiers for the Meta AI app and site, with a free tier kept in place and usage caps on the heavy jobs.

THE FIRST PAID AI BRACKETS

• Meta One Plus: $7.99 a month for more image and video generation and more room on harder queries.
• Meta One Premium: $19.99 a month for the same features at higher volume, including more of the “thinking mode” used for longer reasoning.
• Test markets: Singapore, Guatemala, and Bolivia, with the trials described as starting in June.

Naomi Gleit, Meta’s head of product, said the plans “give people who use Meta AI more to work with, more capacity, bigger, more complex requests, and more room to create for businesses and creators.” That is compute sold by the month to people who already open Instagram and WhatsApp. It is also a much smaller dollar figure than a $21 billion GPU contract. The consumer meter can start without FedRAMP. The wholesale meter cannot.

Short-Term Rentals Are the Product on Offer

Put the May option, the July hire, and the supplier contracts on one page and the shape is not a fourth hyperscaler. It is a landlord clause. Meta keeps the clusters for Muse Spark, ranking, and agents, then sells a slice at a premium when someone else will pay more than internal use, with the right to take the slice back. That is closer to a spot market than to Azure.

WHAT WE KNOW

• The public line: Zuckerberg has now said a cloud is on the table at the annual meeting and that premium offers keep arriving, including on the July 29 call.
• The hire: Brown left AWS at the end of July to work Meta’s data-center build under Janardhan, per people familiar with the matter.
• The other direction: CoreWeave and Nebius still hold disclosed Meta capacity deals that stretch into the 2030s.

WHAT IS UNCONFIRMED

• Anthropic: Early talks about Meta renting compute to the Claude lab, including a figure around $10 billion, have not been turned into a signed disclosure.
• A SKU list: No public regions, SLAs, or enterprise price card has been posted for Meta Compute.
• A confirmed remit: Meta has not described Brown’s job in its own press room the way AWS described his departure.

Until those items move, the honest product is overflow compute on a clawback, sold beside a consumer AI sub, while Azure keeps the compliance catalog and CoreWeave keeps a contract through December 2032. The GPUs Meta might one day rent are, for now, still on someone else’s invoice.

Disclaimer: This article is news reporting and analysis of Meta’s disclosed spending, supplier contracts, and public comments about a possible cloud offering. It is informational only and is not investment advice, a recommendation to buy or sell any security, or guidance on how to purchase cloud capacity. Readers who may act on capex ranges, market-share figures, or deal values should consult a licensed financial adviser, and teams weighing vendors should consult their own cloud architects and counsel, before making decisions. Figures and statuses reflect the company remarks, filings, and research notes cited here and can change with later guidance, contract amendments, or product launches.