Zuckerberg First Addresses Meta Compute Surplus Claims: No One Complains About Too Much Compute, But Renting It Out Is More Profitable

Zuckerberg said Meta maintains full internal demand for its AI compute while still planning to rent capacity to others under the Meta Compute initiative.

The statement came as the company laid out its cloud infrastructure roadmap for the coming year. It addressed speculation that recent spending on data centers had created surplus capacity. Zuckerberg rejected that view but noted that market prices for compute make external rentals attractive.

Meta plans to offer two routes through Meta Compute: direct access to its models and raw compute rental. The approach mirrors the model used by SpaceX, which rents data center space to Anthropic at a reported rate of 1.25 billion dollars per month.

Meta's capital expenditure guidance for 2026 sits between 125 billion and 145 billion dollars. The company also confirmed it will begin volume production of its custom AI chips in September 2026, with an aim to reach 14 gigawatts of deployed compute by 2027.

Internal demand stays strong

Zuckerberg stated that Meta's own AI workloads continue to run at full capacity. He framed the rental plan as a financial decision rather than a sign of idle hardware.

The company sees continued growth in training and inference needs for its family of Llama models. At the same time, external demand for GPUs and accelerators has pushed prices high enough that selective leasing becomes attractive.

Two paths under Meta Compute

Meta Compute will include model access and bare-metal rental options. The model-access path lets external users query Meta's latest Llama releases through hosted endpoints.

The second path offers direct rental of GPU clusters without Meta software. This mirrors the SpaceX-Anthropic arrangement that provides wholesale infrastructure.

Capital plans and chip timeline

Meta's 2026 spending range reflects both continued data-center builds and the ramp of its own silicon. Production of the custom AI chip begins in September 2026.

The company targets 14 gigawatts of total compute capacity by the end of 2027. That scale would support both internal model training and additional rental commitments.

Market pricing drives the choice

High external prices for compute make idle capacity costly to maintain. Renting out portions of the fleet can generate revenue that exceeds the cost of keeping the same hardware on standby.

The decision does not reduce Meta's own usage forecast. Instead it treats excess marginal capacity as a revenue opportunity when demand from outside parties exceeds internal priority.

Competitive context

Other large AI developers face the same tension between internal needs and external pricing. OpenAI and Google have expanded their own cloud offerings, while Amazon and Microsoft continue to allocate capacity across many customers.

Meta's move adds another large supplier to the market at a moment when demand for training clusters remains intense.

What remains uncertain

It is not yet clear how much capacity Meta will actually allocate to rentals versus internal use. The company has not disclosed pricing or contract terms for Meta Compute.

Future quarterly updates will show whether rental revenue materializes at a scale that affects overall capital planning.

Signals to watch

Investors should track Meta's next capital-expenditure update for any revision to the 125-145 billion dollar range. The September 2026 chip production milestone will indicate whether the internal roadmap stays on schedule.

Third-party adoption figures for Meta Compute, once released, will reveal whether the rental business gains meaningful traction against established cloud providers.