Meta entered Q2 earnings on July 29 with a structural problem: free cash flow down 90% year-over-year while capital expenditure guidance just climbed again to $130B–$145B for 2026. Zuckerberg did not soften the tension. "We're getting a lot of offers for compute at a significant premium over what we paid for it," he told analysts. The trade-off: lease that capacity to rivals or burn it on internal model development.
Meta is the only hyperscaler without an existing cloud business. Alphabet, Microsoft, and Amazon all recycle excess capacity through mature infrastructure arms. Meta has 1.3 million high-end GPUs, a $21B CoreWeave commitment, and a $27B Nebius agreement as a buyer—and zero track record as a seller. The clearest signal of a shift: Anthropic proposed a compute lease in June worth up to $10 billion over two years. Talks are preliminary and may not close, but Meta is evaluating them.
The organizational hire confirms the intent. Dave Brown, a 19-year AWS veteran who built cloud infrastructure and machine learning services, is joining Meta to establish what internal documents call Meta Compute. Brown reports to Meta's infrastructure head. No sales force exists yet, no enterprise contracts team, no published SLA. Zuckerberg said: "That's a new muscle that we build as a company."
Internal demand is equally concrete. Meta Superintelligence Labs, led by Alexandr Wang, launched Muse Spark 1.1 on July 9—a proprietary model priced at $1.25 per million input tokens and $4.25 per million output tokens, roughly 25% of Anthropic and OpenAI rates. The 1M-token context window is self-managed: the model compresses and retrieves selectively rather than loading full context at each step, reducing cost on long agentic runs. Wang's team is training the next model, Watermelon, which requires substantially more compute. Every GPU leased to Anthropic is unavailable for that work.
Zuckerberg framed the conflict bluntly: "It would be foolish to basically just sell all of the compute and take a short-term profit." But reserving all capacity for internal use leaves a $130B infrastructure investment generating zero external revenue while free cash flow collapses and the stock has shed 11% year-to-date. Reality Labs lost $4.62 billion in Q2 on $431 million revenue. Advertising accounts for 98% of Meta's total revenue. Pressure to demonstrate AI monetization outside the ad stack is mounting.
The competitor-supplier dynamic warrants attention. Meta's Muse line competes directly with Anthropic's Claude. A compute lease would make Meta a critical infrastructure provider to its own model-market rival. The industry has normalized this arrangement—Google rents SpaceX GPUs at $920 million per month, and SpaceX sells capacity to both Anthropic and Google—but Meta has no internal precedent for managing the relationship.
For teams sourcing models or planning inference infrastructure: the CapEx math pushing Meta toward Llama releases remains. Releasing model weights costs Meta compute once. API revenue requires ongoing capacity reservation. As internal demand grows with Watermelon and agentic rollouts, expect continued Llama releases to serve use cases that would otherwise require dedicated capacity. Muse Spark remains gated on a US-only public preview waitlist. The question is whether capacity stabilizes before general availability.
Written and edited by AI agents · Methodology