Meta and Amazon Web Services just rewrote the math of the AI buildout. On April 24, 2026, the two companies confirmed a multibillion-dollar, multi-year agreement that will see Meta deploy hundreds of thousands of AWS Graviton5 processors — translating into tens of millions of CPU cores — to run the agentic AI workloads now consuming a growing share of its compute budget. It is one of the largest Arm-based CPU commitments ever announced by a hyperscaler customer, and it instantly makes Meta one of the top five Graviton customers in the world.
For an industry that spent the last three years obsessing over GPU shortages, the framing matters as much as the dollar figure. Meta is not buying training accelerators. It is buying the kind of general-purpose silicon that orchestrates, retrieves, reasons in short bursts, and stitches model calls together — the unglamorous backbone work that agentic AI requires at planetary scale.
What Happened
According to the joint announcement from Amazon and Meta, the deal will run for at least three years and begin scaling immediately. Meta has used Graviton chips on a small scale before, mainly for adjacent cloud workloads, but this commitment is in a different league. Tens of millions of cores will be deployed, with room to expand further as Meta's agentic stack matures.
Santosh Janardhan, Head of Infrastructure at Meta, framed the choice in plain terms when he said AWS has been a long-term partner and that expanding to Graviton lets Meta run the CPU-intensive workloads behind agentic AI with the performance and efficiency it needs at its scale. AWS, on its side, leaned into the engineering story: Graviton5 is built on a 3-nanometer process, ships with 192 cores per chip, and packs a cache roughly five times larger than its predecessor. AWS engineers say that translates into about a 25 percent performance uplift and roughly a third lower inter-core latency, even with double the core count.
The financial backdrop is worth naming. Meta's capital expenditure plan for the year is in the range of 135 billion dollars, much of it earmarked for AI data centers, custom silicon programs, and power. Adding millions of cores from a competing hyperscaler does not replace that internal effort — it complements it. As CFO conversations on Wall Street have made clear in recent weeks, no single supply chain is wide enough to satisfy a top-tier AI lab's appetite for both training-class and inference-class compute right now.
Why It Matters
The shift here is conceptual as much as commercial. For most of the current AI cycle, the unspoken assumption was that the bottleneck was GPUs and the prize was Nvidia allocation. That is still partly true for training. But agentic systems — the next layer of products that plan, call tools, and operate over long horizons — turn out to spend most of their wall-clock time on tasks that look more like classical web services: parsing intermediate states, querying caches, marshaling structured data, dispatching to small specialized models, then waiting for a step to finish before moving on. Those tasks are CPU-bound, latency-sensitive, and deeply parallel, which is precisely the workload profile Arm-based server cores like Graviton5 were designed for.
That is why several analysts now describe agentic AI as nearly as much a CPU story as a GPU story. A model that takes ten meaningful actions per user request can issue dozens of internal RPCs, each of which has to be scheduled, secured, and routed somewhere fast and cheap. Doing that on premium accelerator silicon is wasteful. Doing it on dense, energy-efficient CPUs at hyperscaler economics is the obvious play, and Meta is making that bet at industrial scale.
There is also a strategic message inside the deal. Meta runs its own data centers and is investing aggressively in its MTIA accelerator family. Choosing AWS for a slice of its compute portfolio signals that even the largest builders no longer believe a single-vendor, single-architecture stack can serve every workload at the price-performance their product roadmaps require. Diversification, not vertical integration, is the new operating principle.
Reaction
Initial reactions across the industry were quick and largely positive on both sides. Amazon shares moved up on the news as investors interpreted the deal as a validation of the Graviton roadmap and of AWS's ability to win a marquee customer that has historically been associated with Meta's own infrastructure stack. Meta investors focused on a different angle: that paying for external CPU capacity, instead of building it all in-house, may relieve some of the pressure on a capex line item that has visibly stretched the company's free cash flow profile.
Among technical commentators, the consensus framing was that this is a meaningful endorsement of the Arm server ecosystem. Graviton5's 192-core, 3nm design is positioned squarely against the latest x86 server parts on both performance per watt and total cost of ownership, and a public commitment of this size from Meta gives ecosystem partners — operating system vendors, runtimes, observability tools — additional reason to keep optimizing for Arm. Supporters of open hardware noted that diversification of CPU architecture inside the largest AI estates is, on balance, healthy for the broader market.
Skeptics raised reasonable questions. Some asked whether the multibillion-dollar headline figure represents incremental spending or a reshuffling of compute Meta would have purchased somewhere anyway. Others pointed out that a portion of agentic workloads will inevitably migrate to specialized accelerators as inference-time reasoning matures, which could blunt the long-run growth of CPU-only deployments. Both critiques are fair, but neither dents the near-term significance of the announcement.
What's Next
The first deployments are expected to come online during the coming quarters, with Meta gradually shifting agent-orchestration, retrieval, and search workloads onto the new Graviton5 capacity as availability grows. AWS, meanwhile, has indicated it expects Graviton5 to become its fastest-ramping server platform yet, with multiple unnamed customers in the financial services, retail, and AI sectors already in onboarding.
Three things are worth watching from here. The first is whether other frontier AI labs follow Meta's lead and publicly diversify into hyperscaler CPUs for agentic stacks; if they do, the narrative around AI infrastructure spending will broaden well beyond the GPU supply story. The second is how quickly Meta can prove operational savings — measurable improvements in cost per agent action, or in latency under load — that justify the multibillion-dollar commitment to investors. The third is the chip roadmap itself: AWS has hinted that future Graviton generations will further close the gap with custom AI accelerators on certain inference workloads, which could change the buy-versus-build calculus again within a year or two.
Closing Thoughts
It is tempting to read this deal narrowly, as one large customer signing one large contract. The more interesting reading is that the silicon center of gravity for AI is quietly shifting. Training still rules the headlines, but the products users actually touch — the agents that book travel, summarize meetings, debug code, and plan multi-step tasks on someone's behalf — live or die on dense, efficient, general-purpose compute. Meta and AWS just put a very public price tag on that reality.
Whether one views agentic AI as a near-term productivity revolution or a slower-than-promised buildout, the infrastructure debate is becoming more nuanced and, in some ways, more grown-up. The next phase will be less about who has the most GPUs and more about who can stitch together the right mix of accelerators, CPUs, networking, and data center power at a sustainable cost. On that scoreboard, this week's announcement is a meaningful early data point.
한글 요약
2026년 4월 24일, 메타와 아마존 웹 서비스(AWS)는 수십억 달러 규모의 다년 계약을 공식 발표했다. 메타는 AWS의 차세대 Arm 기반 서버 칩인 Graviton5를 수십만 개 도입해 사실상 수천만 개의 CPU 코어를 새로 운용하게 된다. 이번 합의로 메타는 단숨에 전 세계 Graviton 상위 5대 고객 중 하나로 올라섰으며, 계약 기간은 최소 3년이다. AI 인프라의 무게 중심이 GPU 일변도에서 다변화되는 흐름을 보여주는 상징적 거래로 평가된다.
이번 도입의 핵심은 '에이전트형 AI(Agentic AI)' 워크로드다. 모델이 한 번의 사용자 요청에 대해 여러 단계를 계획하고, 도구를 호출하고, 검색·추론·코드 생성을 반복하는 과정에서는 GPU보다 오히려 CPU 사용량이 폭발적으로 늘어난다. Graviton5는 3나노 공정, 192코어, 전 세대 대비 약 5배의 캐시를 갖춘 칩으로, 약 25% 성능 향상과 33% 낮은 코어 간 지연이 강조된다. 메타 인프라 책임자 산토시 자나르단은 "대규모 운영에서 필요한 성능과 효율을 갖춘 선택"이라고 설명했다.
전략적 함의도 분명하다. 메타는 자체 데이터센터와 MTIA 가속기 라인을 공격적으로 확장하면서도 외부 하이퍼스케일러의 CPU 대규모 차입을 동시에 단행했다. 이는 단일 공급선·단일 아키텍처로 전체 AI 워크로드를 감당하기 어렵다는 현실 인식을 의미한다. 약 1,350억 달러 규모로 거론되는 메타의 연간 자본 지출 부담을 일부 분산시키는 효과도 기대된다. 향후 다른 프런티어 AI 기업이 비슷한 다변화 행보를 따를지, 그리고 Graviton 후속 세대가 추론 가속기와의 경계를 어떻게 좁힐지가 다음 관전 포인트다.