Meta signed a multiyear, multibillion-dollar agreement with AWS to deploy hundreds of thousands of Graviton5 processors across its 32 data centers. The deal gives Meta access to tens of millions of ARM-based cores—each chip packs 192—for the CPU-heavy workloads powering agentic AI systems. This is a deliberate architectural bet: inference for autonomous agents, not GPU-bound training.
Graviton5 offers five times the cache of its predecessor, 33% faster core-to-core communication, and 25% overall performance improvement. Santosh Janardhan, Meta’s head of infrastructure, framed the move as a diversification play: “Expanding to Graviton allows us to run the CPU-intensive workloads behind agentic AI with the performance and efficiency we need at our scale.”
The decision reflects a broader compute bifurcation in AI stacks. Training remains GPU-dependent—Nvidia and AMD dominate. But inference for agentic systems—code generation, multi-step planning, real-time task orchestration—leans heavily on CPU throughput and memory bandwidth. Graviton5’s power efficiency and price-performance (40% better than x86 per AWS claims) make it a practical fit for Meta’s billions of users.
Meta’s infrastructure spending is scaling fast: $115–135 billion in capex projected for 2026, up from $72 billion last year. The company has parallel deals with Nvidia, AMD, Google TPUs, and its own MTIA chips via Broadcom. Yet it cut 14,000 jobs. The pattern is clear: invest in silicon diversity while compressing operational costs.
AWS CEO Andy Jassy noted that Graviton and Trainium together now exceed a $10 billion annual run rate. Two unnamed customers wanted all of 2026’s Graviton supply—AWS couldn’t deliver. Meta now joins Uber, Anthropic, and others betting on custom ARM silicon. AWS claims 98% of its top 1,000 EC2 customers use Graviton extensively.
Market reaction was immediate: Amazon closed up 3.45%, Meta up 2.39%. Intel surged 26% on its own CPU narrative around inference. Industry chatter on X highlighted that inference costs now account for an estimated 80% of AI spending, spilling beyond Nvidia’s ecosystem.
This is not an isolated bet. OpenAI has ported code to ARM for Graviton. Uber expanded on Graviton4 for ride-hailing inference. The hyperscaler chip race—Google TPUs, Microsoft Maia, Amazon Graviton—is about reducing Nvidia dependency while optimizing for specific workload profiles. For agentic AI, CPUs are no longer legacy infrastructure; they are the compute layer that handles real-time reasoning at scale.
Meta’s 32 data centers get the upgrade. AWS locks in recurring revenue from custom silicon. The engineering trade-off is clear: GPU for training, CPU for agents. Whoever optimizes inference first wins the next wave.
Source: Webpronews