Meta’s Agent Scale on Graviton

Amazon says Meta is running agentic artificial intelligence workloads on millions of AWS Graviton cores, a sign that not every AI job is staying on graphics chips. (x.com) The claim came from an Amazon post that said Meta scaled “millions” of Graviton5 cores for agents. Amazon’s Graviton line is its in-house Arm-based central processor family for general cloud computing. (x.com) (aws.amazon.com) On Amazon’s current Graviton generation, the biggest C8g, M8g and R8g instances offer up to 192 virtual central processing units, which Amazon markets as its top general-purpose and compute-optimized Graviton capacity. Amazon says those Graviton4 systems deliver better price performance than comparable previous-generation instances. (aws.amazon.com 1) (aws.amazon.com 2) (aws.amazon.com 3) (aws.amazon.com 4) Agentic AI usually means software that breaks a task into steps, calls tools, and makes repeated model requests while it works. That pattern creates a lot of orchestration, routing and memory work that can run on central processors even when model training still depends heavily on graphics processing units. (aws.amazon.com 1) (aws.amazon.com 2) That split has become more important as companies move from training a single large model to operating fleets of assistants, coders and customer-service bots. The cost question shifts from buying the fastest chip to running huge numbers of small, always-on software workers at cloud scale. (aws.amazon.com) (x.com) Amazon has spent years pitching Graviton as a cheaper and more efficient alternative to traditional x86 server chips for many web, data and application workloads. Meta has also been one of the biggest backers of open-weight AI models through its Llama family, which has pushed more companies to experiment with custom deployments rather than relying only on closed model providers. (aws.amazon.com) (about.fb.com) Amazon’s post does not spell out which Meta products are using the cores, how long the systems run at that level, or how much of the stack still uses graphics chips. Meta had not published matching technical detail in the materials reviewed here. (x.com) Even with those gaps, the disclosure puts a number on the scale cloud providers want to sell: AI is no longer only about giant training clusters, but also about the millions of central-processor cores needed to keep agent software running. (x.com) (aws.amazon.com)