Meta Platforms Inc. reportedly plans to rent its artificial intelligence infrastructure to other companies.

Sources told Bloomberg and CNBC today that the Facebook parent will sell excess capacity not used by its internal workloads. The news sent Meta’s shares 8.8% higher, while AI cloud provider CoreWeave Holdings Inc. closed 13.9% lower. Nebius Group NV, a competing provider of machine learning infrastructure, dropped 17%.

Meta could potentially make a significant amount of new computing capacity available for the AI industry. The company expects to spend up to $145 billion on capital expenses, a line item that covers data centers, in its current fiscal year. That’s on top of the $70 billion it invested in 2025.

Meta’s flagship infrastructure project is a data center campus called Hyperion that is currently undergoing construction in Louisiana. The site is expected to consume 5 gigawatts of electricity, which is enough to power more than 4 million homes. It will host 11 buildings equipped with millions of graphics processing units.

AI infrastructure providers such as CoreWeave offer GPUs from multiple chipmakers. It’s possible that Meta will take a similar approach. Earlier this year, the company agreed to purchase millions of AI chips from Nvidia Corp., Advanced Micro Devices Inc. and Google LLC. It’s unclear how Meta’s entry into the cloud market might affect its partnership with the Alphabet Inc. unit.

The social media giant’s planned cloud service may also provide access to its internally developed chips. In March, Meta debuted an inference accelerator called the MTIA 300 that offers 1.2 petaflops of MX8 performance. Next year, the company will start using a more advanced chip that is expected to be about 8 times faster.

Bloomberg reported that Meta is still weighing how to make money from its infrastructure. According to the publication, the company could sell computing capacity or offer access to hosted AI models. That raises the possibility Meta will make its internally developed large language models available to developers.

The company’s flagship LLM, Muse Spark, launched in April. Meta uses the model to power a consumer-focused chatbot service, but it demonstrated competitive performance across several knowledge work benchmarks. As a result, Muse Spark may be suitable for the enterprise use case Meta will presumably prioritize with its cloud service.

If the company launches a series of AI infrastructure offerings and hosted models, it may also decide to enter the training tooling market. Meta overhauled its internal tooling stack to speed up the development of Muse Spark. According to the company, the new stack can achieve the same performance as its previous one using an order of magnitude less infrastructure.

SpaceX Corp.’s xAI unit is also making its excess data center capacity available to other companies. It recently inked infrastructure contracts with Google and Anthropic PBC that are together worth $2.35 billion per month.

Photo: Meta