Meta placed a multi-billion-dollar order for Amazon's Graviton chips, marking the largest known CPU procurement in the company's history and the clearest signal yet that inference architecture is bifurcating. The deal, reported by supply chain sources close to AWS, commits Meta to ARM-based processors across multiple generations, with initial deployments targeting agentic reasoning workloads that coordinate hundreds of GPU calls per user query. The contract value exceeds $3 billion over three years, according to two people familiar with the terms.
The shift is structural. Agentic inference—where AI models plan, iterate, and self-correct across multiple reasoning steps—requires a different compute profile than single-shot text generation. Each agent orchestrates dozens of smaller model calls, memory lookups, and tool invocations. That workload is CPU-bound, not GPU-bound. Meta's Llama agents, now in production across WhatsApp and Instagram search, send an average of 47 CPU-mediated requests per user interaction, compared to 3-5 for traditional LLM serving. The Graviton deal positions Meta to handle that orchestration layer without waiting on Intel's Xeon roadmap or paying Nvidia's margin on Grace-Hopper bundles.
The timing is deliberate. Industry-wide CPU allocation is tightening faster than GPU supply improved. Lead times for high-core-count server chips stretched to 26 weeks in Q4 2024, up from 14 weeks a year prior, per semiconductor distribution data. Hyperscalers are locking long-term volume commitments to guarantee supply, and Meta's move forces AWS to reserve Graviton wafer capacity that otherwise flows to smaller cloud customers. The deal also hedges Meta against Intel's stumbles—Xeon 6 shipments slipped twice in 2024, and yield issues on Intel 3 process nodes persist. Meta's infrastructure team no longer tolerates single-vendor dependency for workloads scaling to billions of daily inferences.
Operators should watch three developments. First, AWS pricing for Graviton instances will likely firm by mid-2025 as Meta's reserved capacity tightens spot availability. Second, competing hyperscalers—Google with Axion, Microsoft with Cobalt—will accelerate their own ARM roadmaps to retain AI customers who cannot source x86 chips at scale. Third, inference-focused AI labs without comparable procurement leverage will face a 15-20% cost penalty on CPU-bound agent workloads, creating a wedge between frontier model builders and those who can't secure chip supply. The gap between model capability and deployment economics is widening, and it now runs through the CPU socket, not the GPU.
Meta's Graviton commitment is not a hedge. It is the new baseline for any AI infrastructure running agentic workloads at population scale, and the order confirms what the spot market already knew: CPU scarcity is the next constraint, and the players who locked supply six quarters ago are the ones still deploying.