Meta Buys Tens of Millions of AWS Graviton Arm Cores in a CPU Land Grab

This one should give folks a sense of how much the CPU market is heating up. Just a few weeks after Meta was out saying it would be a launch customer of the Arm AGI CPU, the company is making another big deal for Arm compute and infrastructure. Meta is buying tens of millions of AWS Graviton Arm cores, but also a lot of AWS infrastructure.

Meta Buys Tens of Millions of AWS Graviton Arm Cores in a CPU Land Grab

A few key details from the press release, starting with the size:

“The deployment starts with tens of millions of Graviton cores, with the flexibility to expand as Meta’s AI
capabilities grow.” (Source: AWS e-mail.)

AWS told us that Meta is one of the largest customers of AWS Bedrock, which is the AWS solution for building generative AI applications. The release we previewed also said that Graviton is built on the AWS Nitro system. That, of course, made me wonder about a simple detail: did Meta buy the Graviton5 chips, or will these be hosted by AWS?

I simply asked, and I was told that AWS would host the parts, but I did not get an answer on whether those Graviton cores would run Bedrock or other applications. Fair enough. Still, that means that Meta is buying not just the chips, but also the power, data center infrastructure, networking, and so forth around those cores because they are being hosted by AWS.

Just for some key specs on the Graviton5:

• 192 Arm Neoverse V3 Cores (these are P-cores, not E-cores)
• Armv9.2-A
• 2MB L2 cache per core (up from 1MB on Graviton3)
• 192MB L3 cache
• For a total of 600MB of cache
• DDR5-8800 support
• PCIe Gen6 support
• Single socket with two NUMA regions
• 3nm

So this makes it more of a next-generation chip versus something like an AmpereOne, AMD EPYC 9005 “Turin”, or Intel Xeon 6. The release just said the deal was for Graviton cores and did not specify the generation.

Final Words

The key takeaway here is simple. CPUs are in such demand due to agentic AI computing that Meta is investing in AWS infrastructure, despite having experience operating at scale itself. Instead of Meta buying chips and having a company like Wiwynn integrate them into systems, the company is buying compute attached to AWS infrastructure. For AWS, this is great news. For Arm, this is also very good news. Some may immediately think that it is bad news for AMD and Intel, but that might not be the best way to frame it, especially after Intel just delivered solid guidance. This is more of a case where big companies with big budgets are locking in resources.