Written by David McMahon
Meta’s new commitment to AWS Graviton chips suggests that the AI race is broadening from a scramble for Nvidia GPUs into a more complex contest over CPUs, cloud leverage, energy efficiency, and the full operating stack behind agentic systems.
For the last two years, the public story of the AI boom has been simple enough to fit inside a single stock chart: whoever controls the most Nvidia capacity controls the future. That story is not wrong, but it is now incomplete. Meta’s new agreement to use hundreds of thousands of AWS Graviton chips, in a deal that CNBC says will last at least three years, is a reminder that the next chapter of AI competition is likely to be defined less by a single component shortage than by control over the entire compute architecture that sits beneath large-scale AI services.
The headline number is arresting on its own. According to CNBC, Meta will deploy hundreds of thousands of Graviton chips from Amazon Web Services, while About Amazon says the rollout starts with tens of millions of Graviton cores and could expand further. AWS also frames the deal as support for the “agentic workloads” behind Meta’s AI efforts, not as a side experiment. In other words, this is not a procurement footnote. It is infrastructure doctrine.
That doctrine matters because Meta is not replacing GPU spending; it is layering on top of it. CNBC notes that the AWS arrangement follows a combined $48 billion in recent AI infrastructure commitments by Meta to CoreWeave and Nebius. Those earlier deals leaned into the high-profile side of the AI arms race: access to Nvidia-powered capacity. The Graviton agreement points to the quieter but arguably more important side of the contest: once frontier models exist, companies still need massive amounts of general-purpose and CPU-oriented compute for orchestration, post-training, inference support, data movement, storage-heavy workflows, and the operational plumbing required by agentic products.
This is the part of the AI buildout that investors and policymakers often underestimate. Training a large model is expensive and glamorous. Running an AI platform at planetary scale is repetitive, distributed, and brutally operational. CNBC reports that roughly 3.6 billion people use Meta’s apps every day and that the company expects to be operating 32 data centers once a new Oklahoma facility is complete. At that level, the economics of the “unsexy” layers matter enormously. A modest efficiency gain at hyperscale quickly becomes a strategic advantage.
That is why Graviton is more than a cost-saving chip. AWS has spent years positioning Graviton as a performance-per-dollar and energy-efficiency play, and CNBC says Amazon claims it uses 60% less energy than competing compute options in certain contexts. Meta’s head of infrastructure, Santosh Janardhan, said the expansion into Graviton will help Meta run the CPU-intensive workloads behind agentic AI with the performance and efficiency the company needs at its scale, according to both CNBC and About Amazon. Strip away the marketing and the message is clear: the AI business is becoming a systems engineering business.
This also helps explain why Amazon’s role in AI is evolving in plain sight. For much of the current cycle, AWS has looked comparatively weaker than Microsoft and Google in the model war. But the Meta deal reinforces Amazon’s alternative thesis: a company can still gain power in AI by owning the picks, shovels, and traffic lanes, even if it is not seen as the most culturally dominant model builder. Graviton gives AWS a way to insert itself deeper into AI operating stacks that do not need premium GPUs for every task. The strategic value lies not merely in selling chips, but in binding customers more tightly to Amazon’s infrastructure logic.
For Meta, the move is equally revealing. Mark Zuckerberg’s company appears to be building a deliberately diversified compute portfolio rather than making a binary bet on any one supplier or processor family. That is a rational response to an environment in which AI demand is volatile, capacity planning is hard, and the workload mix is changing rapidly. Agentic systems are especially important here. They do not just generate outputs; they execute chains of tasks, monitor state, call tools, and run continuously in enterprise and consumer environments. Those systems create a more varied compute profile than a narrow benchmark race for raw training performance.
There is a market implication here that extends beyond Meta and Amazon. If hyperscalers and frontier AI firms increasingly split workloads across GPUs, custom accelerators, and high-efficiency CPUs, the competitive map of AI infrastructure becomes broader and more political. Chip power will still matter, but so will cloud contracts, electricity intensity, regional data center buildouts, and the ability to stitch together a reliable software-and-hardware stack under real operating pressure. That makes the AI race look less like a sprint for the fastest chip and more like a campaign for resilient industrial depth.
The bigger lesson, then, is not that GPUs are losing importance. It is that GPUs are no longer the whole story. Meta’s Graviton deal suggests the companies most likely to endure this capex cycle are those that can treat AI infrastructure as a diversified portfolio problem. The era of monolithic AI compute narratives may already be ending. In its place is a harsher reality: at scale, AI leadership belongs to the companies that can optimize every layer of the machine, not just the most famous one.
Offer Your Reading of What Comes Next. Submit your KOL post today
