Meta and AWS announced a strategic agreement on April 24 to deploy tens of millions of AWS Graviton5 processors for agentic AI workloads. Both companies issued simultaneous first-party statements: Amazon’s newsroom confirmed the deployment starts with tens of millions of Graviton cores “with the flexibility to expand as Meta’s AI capabilities grow,” and Meta’s official blog described Graviton5 as purpose-built for the CPU-intensive demands of agentic AI. The agreement has been reported as a multibillion-dollar deal, according to Data Center Dynamics, though that figure does not appear in either first-party announcement.
Graviton5 is ARM-based. Its positioning here matters. GPU clusters excel at the parallel computation that model training demands. Agentic workflows are different: they require real-time reasoning, sequential decision-making, and multi-step task orchestration. Those workloads run differently, and Meta and AWS frame this agreement as a direct response to that difference, stating that Graviton5 is purpose-built to handle CPU-intensive agentic demands.
That framing is worth taking seriously, not because Meta and AWS say so, but because the pattern behind the announcement is real. This is the fourth major agentic infrastructure signal in roughly 30 days. Google committed $750M to partner-led agentic development. IEA data showed data center electricity demand surging 17%. Core Automation made its own infrastructure commitment. Now Meta and AWS. Four signals. One direction.
What should infrastructure decision-makers take from this? First, the scale. Tens of millions of cores is a floor, not a ceiling, both companies signaled room to expand. Second, the simultaneity: dual first-party announcements on the same day reflect a level of coordination that goes beyond a routine vendor partnership. Third, the architecture: if Graviton5 is the preferred substrate for agentic inference, that has implications for how enterprise teams design AI pipelines, not just what models they run, but what hardware assumptions sit beneath those models.
The practical question for enterprise architects isn’t whether to take the GPU-vs-CPU debate seriously. It’s whether their current infrastructure roadmaps account for the possibility that agentic workloads scale differently than training workloads. That question is now live.
What to watch: Meta’s Llama ecosystem deployment patterns over the next two quarters. If Graviton5 becomes the default inference substrate for Llama-based agentic agents, third-party developers building on that stack will face infrastructure alignment decisions. Watch also for AWS to publish Graviton5 performance benchmarks for agentic reasoning tasks, vendor-stated capability needs independent evaluation before it becomes planning assumption.
The TJS read: when two companies with competing infrastructure interests issue simultaneous first-party announcements framing the same architecture choice as the right one for agentic AI, that’s a coordinated signal worth tracking. The CPU vs. GPU debate for agentic workloads is no longer theoretical. It’s a strategic bet, and Meta and AWS just made theirs public.