Meta plans Meta Compute cloud unit to sell spare AI power

For years, tech giants have been buying GPUs like doomsday preppers hoarding canned goods. Now, Meta is opening the pantry doors to see who wants to buy the surplus.

Bloomberg reported on Wednesday that Meta is developing a new cloud infrastructure business dubbed Meta Compute. The goal is simple: rent out its excess AI computing power and model infrastructure to outside customers, putting it on a direct collision course with AWS, Microsoft Azure, Google Cloud, and neoclouds like CoreWeave.

Wall Street loved the pivot. Meta shares surged over 9% on the news, while CoreWeave tumbled between 12% and 14%, and fellow cloud newcomer Nebius Group dropped roughly 12%.

From internal money pit to external revenue

Building an AI cloud wasn't entirely unexpected. In April, Meta told investors that its capital expenditures could hit a staggering $145 billion this year, fueled by massive data center expansions and Nvidia chips.

Mark Zuckerberg dropped hints as early as Meta's Q3 2025 earnings call, doubling down at this past May's shareholder meeting. He noted that if the company ever overbuilt its AI capacity, selling it off was definitely on the table.

According to reports, the project is being led by Santosh Janardhan, Daniel Gross, and Dina Powell McCormick.

Meta is currently weighing two main paths. It might offer raw compute capacity similar to CoreWeave, or roll out an AWS Bedrock-style managed platform where outside teams can access models like Muse Spark, which Meta launched back in April under former Scale AI chief Alexandr Wang.

Meta isn't the first deep-pocketed player to monetize extra chips. Elon Musk's SpaceX has already tapped this playbook this year, securing eye-popping compute contracts with Anthropic at $1.25 billion per month and Google at $920 million per month.

Why it matters

We think this move shows just how urgent it has become for Big Tech to justify astronomical AI spending.

Despite shelling out $14 billion to recruit Alexandr Wang last year, Meta's proprietary AI models have yet to deliver a standalone consumer hit. With $145 billion going out the door this year, depreciation alone will sting. Turning idle GPUs into billable hours is a smart hedge against overbuilding.

For enterprise AI developers, another massive supplier of top-tier compute entering the ring can only be a good thing. More supply should eventually relieve pricing pressure and loosen availability bottlenecks.

For dedicated GPU cloud providers like CoreWeave and Nebius, however, this is a massive warning shot. Their business models rely heavily on scarcity and hyperscaler spillover. When hyperscalers start selling that spillover back to the public, the squeeze is immediate.

The catch for Meta will be enterprise execution. Running an ad network is very different from managing enterprise-grade cloud SLAs, customer support, and developer tools. Selling raw compute hours is relatively straightforward, but building a true cloud platform will test Meta's operational patience.