Meta is preparing to sell the thing it has spent two years frantically buying. According to a Bloomberg report this week, the company is forming a cloud infrastructure business — reportedly organized under a new unit called Meta Compute — to sell outside customers access to the AI computing power inside its data centers. The plans are early and could change, but per the report, what's being debated internally is the shape of the offering, not whether to build it.
Two shapes are on the table. One is hosted AI models: developers pay to run inference on models served from Meta's infrastructure, including its closed-weight Muse Spark model — an arrangement in the mold of Amazon's Bedrock. The other is raw capacity: renting out GPU clusters directly, the way neocloud providers like CoreWeave do. The effort is reportedly led by Meta's head of infrastructure Santosh Janardhan, with Meta Superintelligence Labs' Daniel Gross and company president Dina Powell McCormick involved.
Either version ends the same way: Meta — the last hyperscale-size AI spender with no cloud revenue line — walks into direct competition with AWS, Microsoft Azure, and Google Cloud, plus the entire venture-backed neocloud tier beneath them.
The overbuild becomes a market
Every hyperscaler justifies its AI capex the same way: the compute pays for itself through products. Meta's version of that story has always been internal — better ad targeting, better recommendations, frontier model training. Renting out slack changes the math from "cost center we defend on earnings calls" to "inventory we sell." It's the most direct answer yet to the question hanging over every AI budget: when the power bills are growing 37% a year, where does the payback actually come from?
It also lands squarely on the neocloud tier. Companies like CoreWeave exist because AI compute was scarce and the hyperscalers hoarded their own. That scarcity story is now being attacked from two directions at once: Nvidia is financing and backstopping new GPU clouds from below, and now the biggest private GPU buyer on earth may start reselling capacity from above. When the scarce asset becomes rentable from four different counters, pricing power goes to whoever has the cheapest balance sheet — and Meta's is one of the cheapest in the game.
The catch: clouds are a service business
Selling compute is not the same as operating a cloud. Enterprise customers expect SLAs, support, security certifications, and a sales force — the unglamorous machinery AWS spent two decades building. Meta has world-class infrastructure and no enterprise motion, which is why the hosted-model path looks like the likelier opening move: it's productized, self-serve, and closer to what Meta already runs internally. Microsoft's answer to the same problem was paying $2.5 billion for forward-deployed engineers. Meta would be starting from zero.
Our take: The headline isn't diversification — it's that AI compute is officially becoming a commodity market. When the industry's biggest private buyer starts reselling slack, compute stops being a moat and becomes inventory, and inventory gets priced. Note the quieter tell, too: you only sell excess if you expect to have some. Either Meta's buildout ran ahead of its internal demand, or it's now buying capacity explicitly to arbitrage it. Both readings say the era of "we need every GPU we can get" is ending — and margin, not access, becomes the battleground.
What to watch
- Whether Muse Spark ships behind a paid API. A closed-weight hosted model would complete Meta's pivot away from its open-weights era — strategy news bigger than the cloud unit itself.
- Pricing versus the neoclouds. If Meta undercuts CoreWeave-class providers, the vendor-financed economics propping up that tier get stress-tested fast.
- Meta's next earnings call. Listen for capex reframed as revenue-generating rather than defensive — the moment this stops being a leak and becomes guidance.
- The incumbents' response. AWS, Azure, and Google Cloud have price levers and exclusive model deals; how quickly they use them tells you how seriously they take it.