SemiAnalysis on Meta's Project Prometheus: 1GW Ohio cluster set to surpass Stargate, Llama 4 called a failure

Summary

Meta's AI infrastructure ambitions are now backed by concrete, committed capital rather than aspirational announcements. According to SemiAnalysis analyst Jeremy, Meta is on track to operate 1 gigawatt of compute power — measured to the servers, not gross utility draw — at its Ohio cluster by end of 2026, followed by close to 2 gigawatts at a Louisiana site by end of 2027. That puts Meta's planned training capacity ahead of OpenAI's Stargate, which SemiAnalysis pegs at roughly 880 megawatts. The 1GW figure carries headline value more than it signals a qualitative leap, but the scale matters for recruiting and signals genuine commitment at a level that was previously confined to vague multi-year targets.

Capex Allocation and the Llama Lag

Meta's 2025 capex is guided at approximately $70 billion, but the bulk of that has historically funded what the company calls "core AI" — recommendation models, ad inferencing, and feed ranking — not generative AI. The Llama infrastructure buildout is a separate, more recent layer of spending, estimated at roughly $30 billion this year on Llama-specific infrastructure. That context explains why Meta has been late on training compute for frontier models relative to peers: the GPUs existed, but were allocated elsewhere.

The architectural shift driving the new buildout is speed. XAI's 122-day cluster deployment reset industry expectations for what's possible, and Meta responded by moving away from its legacy H-shaped, three-story, free-air-cooled data center design — which took roughly two years to build and achieved an industry-leading PUE below 1.1 — toward simpler tent-structure facilities that trade efficiency for construction velocity. The Ohio cluster is already partially built and photographed; GPU installation is expected to make it training-ready somewhere in Q3 2025.

Llama 4 Called a Failure

Llama 4 Behemoth, Meta's flagship model, is described as poorly suited to the current AI paradigm. The core issue is architectural: Meta made design choices around its attention mechanism and expert routing that optimized for pre-training scale at a moment when the frontier has shifted toward test-time compute and reinforcement learning. Those trade-offs make Behemoth weak precisely where it matters most right now — agentic tasks, tool use, long-context reasoning, and multi-step problem solving.

The contrast with Chinese labs is pointed. Firms like DeepSeek, Alibaba, and Moonshot — constrained from accessing state-of-the-art chips — had less incentive to chase pre-training scale and instead invested earlier in post-training and RL techniques that align with the current paradigm. The result is that Chinese labs are now shipping open-source models faster and at higher quality than Meta, and the West broadly is behind on open-source LLM output.

Open Source Dynamics

The strategic logic of open-sourcing Llama is increasingly defensive rather than differentiating. The argument that Llama would be forked and weaponized by Chinese labs largely hasn't materialized — DeepSeek distilled primarily from GPT-4 via API, not from Llama. Meanwhile, Meta's open-source position is eroding: it no longer holds a clear quality lead, and the labs publishing competitive open-weight models are doing so because they are not frontier leaders and have every incentive to build ecosystem rather than protect IP. The same logic applies to Meta itself — open-sourcing makes more sense the further you are from the top of the capability rankings.

Product Timeline and Market Position

In consumer AI, ChatGPT holds approximately 71% query share versus Meta AI at roughly 12%. ChatGPT's user growth shows a clear correlation with model quality improvements and price reductions, reaching 500 million weekly active users by late 2024. Meta's distribution advantage — 2 billion daily active users across its app ecosystem — is real but currently underutilized given the model quality gap.

On timing, a meaningful product release from the newly assembled super intelligence team is more likely late Q4 2025 or early Q1 2026 than mid-year. The new research hires need time to integrate, and the Ohio cluster won't be fully operational until Q3 at the earliest. The compensation Meta is paying — framed as potentially $3 to $5 billion on top researchers — is justifiable if those researchers can prevent another Llama 4-style architectural misjudgment, since each correct architectural call at this scale is worth far more than the salary cost.

Data Center Market Conditions

The broader power market context underscores the stakes. Aggregated US grid interconnection requests from prospective data centers now exceed 500 gigawatts — approaching total US peak load — though the vast majority are speculative. Realistic data center power growth by 2030 is estimated at around 100 gigawatts, meaning only roughly 20% of current pipeline requests will be built. States are competing aggressively through tax breaks, permitting acceleration, and openness to on-site natural gas generation. For smaller or less sophisticated operators, the window to flip land to a hyperscaler at easy margins has closed; a 1 gigawatt site represents $30 to $40 billion in total capex commitment, and hyperscalers now have enough options to be selective.