Meta expands AWS partnership with multiyear Graviton chip deal for AI workloads

Meta is broadening its AI infrastructure strategy with a multiyear agreement to deploy Amazon Web Services' Graviton processors at much larger scale, a move that signals the current compute race is no longer only about securing scarce graphics processors for model training. The deal, announced by Amazon and described by CNBC and Yahoo Finance as a multiyear arrangement, comes as Meta pushes more capital into the systems that support AI products used across Facebook, Instagram, WhatsApp and its broader ad and recommendation stack.

The immediate headline is straightforward: Meta will use Amazon-designed CPUs rather than relying exclusively on outside merchant processors or on GPU-heavy rented capacity for every layer of its AI work. Amazon says the deployment begins with tens of millions of Graviton cores and can expand as Meta's AI requirements grow, while CNBC reported that the social-media group will tap hundreds of thousands of chips and become one of the top five Graviton customers. That matters because it suggests Meta is not making a symbolic procurement decision but a meaningful architecture choice about what kind of compute belongs where.

For the past two years, most public discussion of AI infrastructure has centered on Nvidia accelerators, blockbuster cloud leases and the extraordinary capital budgets attached to frontier-model training. This agreement does not overturn that hierarchy, but it complicates it. Graviton is a CPU family, not a GPU line, and AWS, Meta and outside coverage all frame the partnership around workloads that sit beside training: post-training refinement, inference support, real-time reasoning, search, code generation, retrieval and the coordination of multi-step agent systems. In other words, the companies are betting that the next bottleneck in AI may be less glamorous than model training but just as commercially decisive.Meta will adopt hundreds of thousands of AWS Graviton chips in latest AI infrastructure grabcnbc.com·SecondaryAround 3.6 billion people use Meta's applications every day, and the social networking company will be operating 32 data centers to handle the load with the completion of a new one in Oklahoma. But that's not enough. Amazon's cloud unit said Friday that Meta has agreed to use Amazon's general-purpose Graviton chips in a deal that will run for at least three years.

Meta has been telegraphing that it intends to spend heavily to stay competitive. CNBC reported that the company recently signed roughly $48 billion in commitments with CoreWeave and Nebius, both providers of Nvidia GPU capacity, and that it is simultaneously expanding its data-center footprint to 32 facilities with a new site in Oklahoma nearing completion. Amazon, for its part, has been promoting its custom-silicon strategy more aggressively, arguing that enterprises care not only about absolute performance but about price, energy use and the ability to keep large always-on systems running economically. The Graviton announcement therefore serves both sides: Meta diversifies supply and cost structure, while AWS gets a marquee validation that its in-house chips can handle mainstream AI demand at hyperscale.

Amazon's official language is ambitious, and investors will naturally discount some of it because the company is describing its own product. Still, the technical case is not frivolous. AWS says Graviton5 is designed for CPU-intensive AI work, offers 192 cores, improves communication speed between cores, supports high-bandwidth, low-latency networking through Elastic Fabric Adapter, and delivers better performance per watt than earlier generations. CNBC separately reported AWS's claim that Graviton can offer its best price-performance among EC2 compute options while using substantially less energy, a point likely to matter as power availability becomes a practical constraint on AI deployment. Even critics of corporate press releases would concede that cost-per-query and power efficiency are becoming central, not secondary, questions in the AI buildout.

There is also a strategic subtext here about bargaining power. Meta has already worked with Nvidia, and Yahoo Finance noted that Nvidia announced a Meta deployment of Grace CPU-only servers earlier this year, while AMD also disclosed a similar Meta deal in February. Rather than choosing one supplier and one architecture, Meta appears to be building a portfolio of compute options. That approach gives the company leverage in pricing negotiations, supply resilience if one vendor tightens, and more freedom to match different workloads to different chips. From a conservative business perspective, that diversification looks less like hype and more like standard strategic prudence for a company whose AI bill is now large enough to affect margins, staffing decisions and investor patience.Meta will adopt hundreds of thousands of AWS Graviton chips in latest AI infrastructure grabcnbc.com·SecondaryAround 3.6 billion people use Meta's applications every day, and the social networking company will be operating 32 data centers to handle the load with the completion of a new one in Oklahoma. But that's not enough. Amazon's cloud unit said Friday that Meta has agreed to use Amazon's general-purpose Graviton chips in a deal that will run for at least three years.

That investor pressure is real. CNBC reported that Meta announced plans to cut about 8,000 jobs, or roughly 10% of its workforce, even as it continues to escalate infrastructure commitments. Supporters of Zuckerberg's strategy argue that this is exactly the discipline public markets have demanded: fewer people in lower-priority functions, more capital directed into durable infrastructure that can support products for billions of users. Skeptics make the opposite case, namely that giant platforms are overbuilding around a still-evolving vision of agentic AI and could wind up locking in high fixed costs before revenues fully justify them. Both views deserve airtime. Official company statements talk about efficiency and scale, but the unanswered question is whether consumers and advertisers will value these new AI services enough to cover the extraordinary spending now being normalized across the sector.

The agreement also carries implications beyond Meta. Intel chief executive Lip-Bu Tan said this week that demand for the company's Xeon server chips exceeds supply, and he argued that the CPU is reasserting itself as a foundational layer of the AI era. Yahoo Finance likewise pointed to renewed enthusiasm for CPUs as AI agents gain traction, because many tasks associated with those systems are better suited to central processors than to graphics chips. If that thesis holds, AWS is not merely selling one big customer more capacity; it is trying to redefine the market narrative so that the winners in AI infrastructure are the companies that can orchestrate complete stacks of chips, networking and cloud services rather than just the firms with the hottest accelerators.

Government officials are not central actors in this particular announcement, but the policy backdrop matters all the same. Washington and European regulators have grown more attentive to semiconductor supply chains, energy intensity, competition in cloud infrastructure and the concentration of AI capability inside a few US giants. A deal like this will likely be read in policy circles as further evidence that the largest platforms are vertically integrating more of the AI stack through proprietary chips, exclusive cloud capacity and long-term supply agreements. Supporters will say that is how the United States preserves its lead in a strategic industry. Critics will say it raises barriers to entry and pushes smaller firms farther from the frontier. Both interpretations have some force.

For now, the most important takeaway is narrower and more practical. The AI arms race is maturing from a simple scramble for GPUs into a broader contest over the full economics of deployment: which chips are used for training, which for inference, which for orchestration, and which cloud can turn all of that into reliable services at acceptable cost. Meta's Graviton expansion is a strong sign that hyperscalers think the next phase of competition will be won not just by raw model prowess, but by whoever can make AI systems cheaper, steadier and easier to run at global scale.