Meta spent $182.9 billion building GPU infrastructure to train Llama models. Now it wants to rent you the leftovers. The company is reportedly launching Meta Compute — a cloud service that puts Meta in direct competition with AWS, Azure, and Google Cloud. Bloomberg reported the plans on July 1. CoreWeave dropped 14% the same day. Nebius lost $12 billion by close.
Two Products, Not One #
Meta Compute is not a single service. It is two businesses bundled under one brand, and understanding the difference matters for how you would use it.
The first is a hosted model API: Llama endpoints with per-token pricing, auto-scaling, and load balancing. Think Amazon Bedrock, but Meta controls the weights, the pricing, and the roadmap. There is also a fine-tuning API that would let teams upload their own data and ship a customized Llama model without managing infrastructure. This matters because Meta currently distributes Llama as downloadable weights only. Every developer running Llama in production today routes through third parties — Together AI, DeepInfra, Fireworks. Meta Compute would cut them out.
The second is raw GPU rental: bare-metal instances, managed Kubernetes clusters pre-loaded with AI toolchains, billed by the hour. This is the CoreWeave model — which is exactly why CoreWeave stock reacted the way it did.
The Pricing Argument #
Meta is targeting 20-30% below comparable AWS p4d and Azure ND H100 v5 pricing. The cost advantage comes from two places: the MTIA custom silicon (four chip generations in roughly two years, 300 series in production now, 400/450/500 series targeting GenAI inference through 2027) and the leverage of a $182.9B infrastructure spend that is already paid for. You do not need to price for ROI when the hardware debt is already on the books.
For context: third-party Llama 3.3 70B inference currently runs $0.23 per million input tokens on DeepInfra — already cheap. If Meta undercuts that at first-party quality, the economics of the current Llama hosting market get complicated fast.
The Neocloud Problem #
The market reaction was immediate. CoreWeave fell 14% on July 1 and is now down 35% from its 2026 high. Nebius dropped 17%, losing $12 billion in a single session. IREN also fell. The concern is not abstract: Nebius holds a $27 billion Meta contract. CoreWeave holds $21 billion. Both companies built significant portions of their business on Meta as a customer. Now that customer wants to become their competitor.
The counter-argument — and it is worth taking seriously — is that Meta building so much GPU capacity that it has surplus to sell is evidence of a growing market. If demand were softening, Meta would be quietly decommissioning racks, not launching a cloud business. The selloff may be a short-term overreaction.
Developer Tooling #
A Meta Compute SDK is expected on GitHub with Go, Python, and C# libraries. There is also a reported Visual Studio and VS Code extension for submitting training jobs and monitoring GPU utilization directly from the IDE. None of this is live yet. Meta has not made an official product announcement as of July 16, 2026 — all current details come from Bloomberg reporting and follow-on coverage. There is no pricing page, no sign-up, no public documentation.
What to Actually Do #
Do not migrate workloads to Meta Compute yet. Wait for an official launch with public pricing. When it arrives, benchmark the Llama API against your current DeepInfra or Together AI spend — the fine-tuning pipeline is where the real value could land if the pricing is aggressive. Also think about the privacy trade-off: self-hosting Llama means your training data stays on your infrastructure. Sending it to Meta cloud does not. That calculation is different for different teams.
The infrastructure play here is real. The question is whether Meta executes — and whether developers trust them enough to route sensitive workloads through the same company behind Facebook.