Amazon Web Services closed a multi-year infrastructure agreement with Meta to run agentic AI workloads on AWS Graviton chips, while separately expanding its Anthropic partnership to a total $33 billion commitment. The Meta deal marks the first time a hyperscale social platform has routed production AI inference to a competitor's custom ARM silicon at this scale. The Anthropic expansion increases Amazon's total stake from $8 billion announced in March 2024 to the largest single AI infrastructure commitment on record.
Meta will deploy agentic AI systems — autonomous reasoning workflows that chain multiple model calls — on AWS Graviton4 processors, which Amazon claims deliver 30 percent better price-performance than x86 alternatives for inference workloads. The agreement includes co-engineering on future Graviton releases, suggesting Meta sees ARM inference as structurally cheaper than continued expansion of its own NVIDIA H100 clusters. Amazon did not disclose contract value, but comparable hyperscale agreements typically run $2 billion to $5 billion annually. The Anthropic extension reserves capacity on AWS Trainium2 chips through 2027, with Amazon taking additional equity at an undisclosed valuation. Anthropic will migrate Claude training runs to Trainium clusters by mid-2025, replacing a portion of its NVIDIA A100 footprint.
The Meta routing decision is the cleanest signal yet that custom silicon has crossed the deployment threshold for production inference. Meta already runs Llama models on internal infrastructure, but agentic workflows — which require sequential reasoning steps and extended context windows — generate 4x to 8x the compute cost of single-pass generation. Offloading that load to Graviton lets Meta preserve its internal GPU capacity for training and research while avoiding the capital expense of expanding proprietary data centers. For Amazon, the deal validates a $12 billion Graviton R&D bet and opens the door to similar agreements with Alibaba, ByteDance, or other platforms facing the same inference cost curve. The Anthropic stake increase positions Amazon as the dominant infrastructure partner for frontier labs unwilling to build their own chip roadmaps. OpenAI relies on Microsoft Azure. Google uses TPUs internally. Anthropic is now structurally tied to AWS silicon for the next three years, which gives Amazon a recurring revenue stream indexed to Claude usage growth. If Claude scales to 100 million weekly active users by 2026 — analysts estimate it sits near 20 million today — the compute draw would rival Meta's entire inference load.
Allocators should track three follow-on events. First, whether Meta extends the agreement to training workloads by Q3 2025, which would signal ARM viability for full-stack AI operations. Second, Anthropic's Trainium2 migration timeline; delays past mid-2025 would suggest AWS silicon still lags NVIDIA for frontier model training. Third, pricing disclosure from Amazon's Q1 2025 earnings call in late April, particularly any breakout of AI infrastructure revenue as a standalone segment. If Amazon begins reporting Graviton and Trainium revenue separately, the market will reprice AWS growth assumptions by 200 to 300 basis points.
The agreements arrive as NVIDIA trades at 38x forward earnings and ARM at 68x, with both valuations assuming continued dominance in AI compute. Amazon just opened a third bidding surface.