The data centers that train Meta’s models spend part of every day running near empty. Meta Platforms is now trying to sell that spare capacity, announcing plans to offer AI computing power and model access to outside customers, according to Bloomberg and the Los Angeles Times. The move puts Meta in direct competition with Amazon Web Services, Microsoft’s Azure and Google Cloud.
The offering has two parts. Meta will sell raw computing capacity, letting customers rent time on the GPU clusters that power its own AI work, and it will sell access to its models, including the Llama family and the agent tools built on top of them. Customers will be able to run inference, fine-tune models and deploy AI applications without building the infrastructure themselves.
The financial logic is simple and powerful. Meta has assembled one of the largest AI infrastructure fleets in the world, with hundreds of thousands of Nvidia H100 accelerators, and those chips cost the same whether they are idle or busy. Selling the unused hours to outside customers is, as one analyst put it, nearly pure margin: the hardware is already paid for, and every hour sold adds revenue without adding much cost.
Deutsche Bank analysts estimate the cloud business could generate up to $30 billion in incremental revenue by 2027, a figure that would make Meta’s cloud operation one of the largest in the industry within two years if realized. Meta has not confirmed any target, and the estimate assumes the company can convert its infrastructure advantage into enterprise sales, a skill it has rarely demonstrated.
The competitive hurdles are real. AWS, Azure and Google Cloud each spent more than a decade building sales forces, compliance certifications and partner ecosystems, and each counts the world’s largest companies among its customers. Meta has none of that machinery. Its enterprise history is thin, marked by products like Workplace, the corporate collaboration tool the company shut down, and its culture has been built around consumer apps, not procurement departments.
What Meta has is hardware and models. Its infrastructure fleet rivals the big three in scale, and its Llama models are the most widely used open-weight AI systems in the world. The company’s pitch to customers, according to people familiar with the plans, is built on that combination: the same infrastructure that runs Meta’s own products, and models that customers can inspect, modify and move.
The timing reflects a broader squeeze. Meta’s capital spending on AI has climbed to tens of billions of dollars a year, and investors have been asking when that spending starts paying for itself. A cloud business is one answer: it turns the biggest cost center in the company into a revenue line. The near-zero marginal cost of selling idle capacity makes the arithmetic attractive even at aggressive discounts to AWS prices.
The model side of the offering has its own history. Meta has distributed Llama openly since 2023, giving the weights away in the hope of building an ecosystem, and it has made money from the models indirectly, through the products that run on them. The cloud business changes that calculation: for the first time, Meta will sell access to its models directly, at prices customers pay in cash rather than in ecosystem goodwill.
Regulators may have something to say about that. Meta’s size has made it a frequent target of antitrust scrutiny, and a cloud business built on the company’s AI infrastructure could draw questions about whether Meta is using its scale in one market to enter another. The company has said it will keep the offering separate from its advertising business, but critics note that the same data, models and infrastructure that power Meta’s apps would now also serve paying customers.
Analysts are skeptical about the size of the opportunity. The $30 billion estimate assumes Meta can capture meaningful share in a market dominated by three entrenched providers with enormous advantages in enterprise software, government contracts and developer mindshare. “Selling compute is a commodity business with a razor-thin margin if you are not the scale leader,” one cloud analyst said. “Meta’s advantage is that its marginal cost is already sunk.”
There is also the question of what Meta is willing to build. Running a cloud requires more than GPUs: it requires billing systems, security certifications, compliance teams and a global sales organization. People familiar with the plans said Meta is starting small, with a limited set of customers in a limited set of regions, before deciding how far to go.
The strategic rationale goes beyond revenue. A cloud business gives Meta direct relationships with the companies that are building AI products, relationships that could shape which models those companies use. It also gives Meta a hedge: if advertising growth slows, the infrastructure built for AI can still earn its keep by serving someone else.
For the big three cloud providers, the entry is another sign that AI has made infrastructure the center of the technology industry. Amazon, Microsoft and Google have each built their cloud businesses on the same logic Meta is now applying, and each is spending record amounts on AI capacity. Meta’s arrival adds a fourth player with an unusual advantage: it does not need the cloud business to be profitable on its own, only to justify the hardware it already owns.
