The AI infrastructure industry faces a new bottleneck, and it has nothing to do with GPUs. Meta’s multi-billion-dollar partnership with Amazon Web Services to deploy tens of millions of Graviton5 CPU cores exposes how agentic inference workloads are reshaping compute priorities at hyperscale. For years, the narrative centered on training—massive GPUs grinding through data. Now, the challenge is execution: real-time reasoning, multi-step task orchestration, and code generation at the scale required to power billions of interactions.
Key Takeaways
- Meta signed a 3-5 year, multi-billion-dollar agreement with AWS to deploy tens of millions of Graviton5 cores for AI infrastructure
- Graviton5 processors contain 192 cores optimized for parallel workloads, delivering 25% better performance than the previous generation
- Agentic inference workloads—systems that reason through and execute tasks autonomously—are driving CPU demand, marking a shift from GPU-centric training models
- Meta becomes one of the largest global Graviton customers, signaling broader industry adoption of CPU-optimized architectures for reasoning and execution
- Deployment spans multiple phases with flexibility to expand, reflecting uncertainty about the true scale of agentic AI demand
Why agentic inference workloads are creating a CPU crisis
Training large language models still demands GPUs. But once a model exists, deploying it at scale requires something different: sustained, efficient compute for reasoning and execution. Agentic inference workloads perform real-time reasoning, generate code, search knowledge bases, and orchestrate multi-step tasks without constant human intervention. These operations are CPU-intensive rather than GPU-intensive, and they run continuously across millions of concurrent users. The math is brutal. A single GPU can train faster, but thousands of CPUs running in parallel can execute more efficiently at inference scale.
Meta’s deal reflects this reality. AWS Vice President Nafea Bshara stated that the agreement will run between three and five years and is worth billions of dollars over the contract term. The company will deploy tens of millions of Graviton cores across AWS data centers, with rollout happening in phases. This is not a one-time purchase—it is a commitment to building out CPU-intensive infrastructure as agentic systems mature and demand grows.
Graviton5: Purpose-built for agentic demands
Each Graviton5 processor includes 192 cores designed for parallel AI workloads, higher-efficiency compute, faster data processing, greater bandwidth, and sustained performance. The chip delivers up to 25% better performance than Graviton4, AWS’s previous-generation processor. For agentic inference, this matters. A 25% performance improvement translates to lower latency for reasoning tasks, faster code generation, and more efficient search operations—all critical for systems that need to execute at scale without introducing user-facing delays.
Meta’s announcement emphasizes that no single chip architecture can efficiently serve every workload. GPUs remain essential for training. But for inference—particularly agentic inference—CPUs optimized for bandwidth and parallel execution offer better efficiency and cost economics. Graviton5 is purpose-built for exactly this use case: systems that need to continuously reason through and execute tasks at scale.
The broader industry shift from training to execution
Meta’s deal is not happening in isolation. It signals a fundamental reorientation of AI infrastructure priorities across the industry. For the past two years, the bottleneck was training capacity—who had enough GPUs to build and fine-tune models. Now, the constraint is execution capacity—who can deploy reasoning systems efficiently enough to serve billions of interactions. This shift exposes a critical gap: CPU supply chains were never designed for hyperscale agentic workloads.
Meta’s portfolio approach—combining GPUs, custom silicon, and now tens of millions of Graviton cores—reflects this reality. No single architecture wins. Training demands different hardware than inference. Real-time reasoning demands different optimization than batch processing. By committing to Graviton at scale, Meta is signaling that the industry has moved beyond the GPU-centric narrative and into a more nuanced, architecture-diverse phase of AI infrastructure development.
What this means for CPU supply and AI infrastructure costs
The announcement comes as agentic AI systems move from research projects to production workloads. OpenAI’s o1, Claude’s extended thinking, and Meta’s own reasoning systems represent a category of AI that requires sustained CPU execution, not just GPU training bursts. If these systems scale to billions of users—which the companies clearly expect—CPU demand will dwarf current capacity.
Meta’s multi-year commitment to tens of millions of Graviton cores is AWS’s way of signaling it is taking this seriously. But the phased rollout and flexibility to expand suggest uncertainty about the true scale of demand. If agentic systems become as ubiquitous as search, CPU shortages could persist for years. If adoption plateaus, the infrastructure sits underutilized. Meta is betting on the former. The market will test that bet starting now.
Does this mean GPU demand is declining?
No. Training still requires GPUs, and training is not going away. What is changing is the proportion of infrastructure dedicated to training versus inference. As models mature and agentic systems scale, the ratio tilts toward inference—and toward CPUs. This does not displace GPU demand; it adds CPU demand on top of it.
Why is Meta announcing this now?
The timing is deliberate. Agentic AI is moving from hype to production. Systems like o1 and extended thinking modes are live and generating real traffic. Meta’s own reasoning systems are in development. The company needs infrastructure to support these workloads at scale, and it needs it now. The announcement also comes as the industry reassesses AI infrastructure priorities—away from pure training capacity and toward execution capacity that can support real-time reasoning.
Will other hyperscalers follow Meta’s lead with CPU partnerships?
Likely. If agentic inference becomes as central to AI operations as training, every hyperscaler will need equivalent CPU capacity. Meta’s deal with AWS is the first public signal of this shift, but it will not be the last. Expect announcements from Google, Microsoft, and others as they assess their own agentic AI roadmaps and infrastructure gaps.
Meta’s commitment to tens of millions of Graviton cores is not about choosing CPUs over GPUs—it is about recognizing that agentic inference workloads require a different architecture optimized for sustained reasoning and execution at scale. The AI infrastructure industry has entered a new phase, and CPU supply is now the constraint that matters most.
This article was written with AI assistance and editorially reviewed.
Source: Tom's Hardware
