Meta has entered a multiyear infrastructure agreement with Amazon Web Services worth billions of dollars, according to the report. The deal calls for Meta to deploy hundreds of thousands of AWS-designed Graviton processors across its global data centers and AI workloads, with the latest Arm-based Graviton5 CPU at the center of the rollout. These chips are expected to support general computing, AI inference, and a portion of training tasks.
The scale of the purchase lines up with Meta’s rising AI spending. The report says Meta’s 2026 capital expenditure is projected at $115 billion to $135 billion, making efficiency a core concern as the company expands infrastructure. Graviton chips are being positioned as a way to lower the cost of inference and large-scale data center growth. Compared with traditional x86 designs, the Graviton line is described as offering stronger cost efficiency and better power performance, easing both cooling and electricity demands. The point is simple: Meta is looking for cheaper compute where it can get it.
Graviton5 rollout moves from cloud ties to hardware commitment
AWS executives did not disclose the exact contract value, but they confirmed that the agreement is substantial. The report frames this as more than an extension of existing cloud cooperation. Meta and AWS have already worked together on cloud services and deployment related to the open-source Llama models, yet this deal shifts the relationship into large-scale hardware adoption, with Meta explicitly buying into AWS’s in-house silicon strategy.
For Meta, that matters because compute sourcing has become a strategic issue. The report points to Nvidia’s dominance in AI GPUs and the high price of those systems as a major pressure point. AWS chips are expected to complement Meta’s in-house MTIA chips as well as solutions tied to Broadcom and AMD. That gives Meta a broader hardware mix for different workloads and reduces exposure to a single supplier.
CapEx pressure pushes Meta to diversify compute sources
Meta’s AI buildout has accelerated, and so has the budget behind it. In that setting, inference economics carry more weight. Bringing Graviton5 into the stack suggests Meta is not assigning every AI task to the most expensive accelerators available. Instead, the company appears to be separating general compute, inference, and some training tasks across different processor types, matching workloads to hardware with closer attention to cost, supply, and energy use.
The report also places the deal in a wider industry shift. As the AI race intensifies in 2026, major technology companies are trying to reduce dependence on any one chip vendor. Meta’s order from AWS fits that pattern. The emphasis on CPUs also highlights a practical point: not every AI workload needs to run on top-tier GPU capacity.
AWS gains another major validation for its custom chip business
For AWS, the agreement is a significant endorsement of its chip effort. The report says AWS’s custom silicon business, including Graviton and the AI-focused Trainium line, has reached an annualized revenue run rate of more than $20 billion. Securing a hyperscale customer such as Meta gives AWS stronger proof that its in-house hardware can compete not only inside its own cloud platform, but also in the highest tier of AI infrastructure demand.
The report ties this move to a broader procurement trend already visible across the sector. Meta has worked with AMD and Broadcom, Anthropic has made major commitments to AWS, and now Meta is set to buy Graviton5 at scale. Taken together, those moves show large technology companies reshaping their AI hardware supply chains around cost, performance, and supply resilience at the same time.