Meta Adds Tens of Millions of AWS Graviton Cores for Agentic AI Workloads

A CPU deal, in a moment dominated by GPU talk

Meta's announcement is notable for what it is buying: processing cores inside CPUs, not accelerators. The company says it is bringing "tens of millions of AWS Graviton cores" into its compute portfolio, describing this as making Meta "one of the largest Graviton customers in the world."

The stated reason is a shift in where the work sits. Meta says that as it deepens its work with agentic AI, "compute requirements are evolving to demand more CPU." The company defines agentic AI plainly — "autonomous systems that reason, plan, and execute complex tasks" — and ties the Graviton5 cores to systems "that need to continuously reason through and execute tasks at scale."

That framing is the most concrete claim in the release: the loop of planning and executing tasks is CPU-intensive, and Meta is provisioning for it separately from its training and inference accelerators.

What the automation workload actually demands

The release positions Graviton5 as "purpose-built" for these demands, citing "faster data processing and greater bandwidth." For teams building agentic systems, that lines up with a practical reality: an agent that loops through tool calls, retrieves data, coordinates steps, and executes actions spends a lot of time on general-purpose orchestration and data movement rather than pure matrix math.

Santosh Janardhan, Meta's Head of Infrastructure, was specific about the target: "expanding to Graviton allows us to run the CPU-intensive workloads behind agentic AI with the performance and efficiency we need at our scale."

As we scale the infrastructure behind Meta's AI ambitions, diversifying our compute sources is a strategic imperative. AWS has been a trusted cloud partner for years, and expanding to Graviton allows us to run the CPU-intensive workloads behind agentic AI with the performance and efficiency we need at our scale. — Santosh Janardhan, Head of Infrastructure, MetaMontana Labs

The portfolio principle, stated in the company's own terms

Meta describes its infrastructure strategy as a deliberate mix: "We invest in our own data centers and custom hardware. We partner with cloud providers who bring differentiated capabilities. And we continuously evaluate which architectures are best suited to which workloads."

The organizing principle it names is "that no single chip architecture can efficiently serve every workload." This deal is presented as evidence of that view rather than a departure from it — Meta building its own silicon and data centers while renting AWS cores for a specific class of work.

The commitment is also structured to grow. The first deployment "will start with tens of millions of Graviton cores, with the flexibility to expand as our AI capabilities grow" — an open-ended arrangement tied to how fast Meta's agentic systems scale.

The implication: agentic AI is a hardware-diversification problem, not only a model problem

The clearest takeaway from this announcement is that Meta is treating the rise of autonomous agents as a reason to broaden its compute base rather than concentrate it. The task loop of reasoning, planning, and execution is being provisioned on general-purpose Arm cores from a cloud partner, alongside Meta's own hardware.

For anyone building agentic systems, the signal is that infrastructure planning for agents should account for CPU-bound orchestration and data movement as a first-class cost — not an afterthought to accelerator spend. Meta, at its scale, is willing to sign a multi-million-core agreement to secure that capacity.