The company is moving from buyer to rival of smaller AI cloud vendors like CoreWeave and Nebius, the specialized GPU rental firms that rented it capacity while Meta built out its own fleet.
Meta spent roughly $145 billion on AI infrastructure in 18 months. Now the company that was CoreWeave's biggest customer wants to become its competitor.
That is the real shape of the Forbes column framing the move and CNBC's analysis of the strategic logic: a supplier-customer inversion at the heart of the AI build-out. Meta is exploring deals to sell spare "AI compute," the rentable data-center processing power used to train and run AI models, to outside buyers. If it follows through, it would compete head-on with the smaller neocloud vendors, including CoreWeave and Nebius, that rented it GPU capacity while Meta was still building out its own fleet.
The timing is the point. Meta's 2026 AI and compute spending, cited by analysts and repeated in columnist coverage, runs to about $145 billion. Most of that is still going into training its Llama models and the recommender systems that power Instagram and Facebook feeds. But the build has moved faster than the workloads that justify it, and the company is sitting on GPU clusters large enough to lease out. Selling that surplus turns a sunk-cost pile into a revenue line, and the most obvious counterparties are exactly the companies that helped Meta get here.
The neocloud model is built on a thin base. CoreWeave and Nebius are not general-purpose cloud providers like Amazon Web Services or Microsoft Azure; they specialize in renting large pools of Nvidia GPUs to AI labs, startups, and hyperscalers that need extra capacity fast. That model only works as long as the customer list keeps growing, and Meta is one of the largest individual customers for that kind of capacity. The expanded CoreWeave-Meta agreement announced last September runs to December 2031 and ties CoreWeave revenue to Nvidia's GB300 systems; the headline number has appeared as $14.2 billion in some coverage and $21 billion in CoreWeave's own announcement, a discrepancy worth tracking. Either way, CoreWeave has bet a meaningful slice of its forward revenue on a customer that may now be trying to displace it.
The strategic framing is mostly synthesis. TechCrunch's read compared Meta's play to SpaceX turning stranded launch capacity into Starlink revenue. It is a useful analogy and a reasonable analyst read, but it is not Meta's stated rationale. The company has not framed the cloud push as a capacity-monetization strategy in its primary disclosures. The "excess AI compute" line is columnist and analyst framing, lifted into the broader narrative as Meta's strategic intent.
What Meta does bring is structural. It owns the data centers, controls the power contracts, and pays the capital cost with the balance sheet of a company that still pulls roughly 98% of its revenue from digital advertising. CNBC's angle is that Wall Street finds the move strategically exciting precisely because the AI build is already happening and monetization is the open question. The same piece warned that cloud services will run at materially lower margins than advertising. Investors are pricing a new growth story, not a margin expansion.
The competitive math for neoclouds is harder to dismiss. A hyperscaler with its own fleet, its own power, and a captive ad business to subsidize the build has structural cost advantages over a vendor whose largest customer is now trying to undercut it. CoreWeave's forward revenue is more exposed to Meta than Meta's capex is exposed to CoreWeave. If Meta moves even a modest share of its surplus to outside customers, the neoclouds either match on price, which they cannot easily do, or they cede the segment.
The open questions are the ones that matter. Has Meta signed any external cloud customers yet, or is it still in talks? Will it undercut CoreWeave and Nebius on price, or focus on workloads those vendors cannot serve? And how long before Wall Street starts asking whether the same $145 billion AI build that produced the surplus should have been spent more carefully in the first place?