Meta and Arm team up on custom AI CPUs

While the entire tech world is obsessed with hoarding GPUs, Meta just made a massive pivot into CPUs—and pulled Arm straight onto the hardware battlefield.

Meta and Arm announced a multi-generational partnership to co-develop custom data center processors built for massive AI workloads. The debut silicon, dubbed the Arm AGI CPU, marks the first time Arm has ever manufactured its own data center processor rather than just licensing designs.

Meta is the lead partner and will start plugging the chips into its infrastructure later this year.

Arm stops licensing and starts building

For decades, Arm made billions doing one thing: selling blueprinted chip architectures to giants like Apple, Qualcomm, and Amazon to fab themselves.

That era is over. The new Arm AGI CPU is built on Arm's Neoverse platform, packing up to 136 cores per CPU and squeezing 64 processors into a standard air-cooled rack. Arm claims it delivers double the performance per watt compared to traditional x86 server chips while breaking through nasty memory bottlenecks.

While Meta gets first dibs, Arm is opening this silicon up to the wider ecosystem. A long list of big players is already queuing up, including OpenAI, Cloudflare, Cerebras, SAP, and SK Telecom.

Silicon heavyweights like AWS, Google, Microsoft, and Nvidia chimed in with congratulatory notes. Notably missing? Qualcomm, which scored a massive legal victory against Arm last fall.

Meta's open-source play

For Meta, the move solves a desperate infrastructure bottleneck. Feeding gigawatt-scale data centers for agentic AI requires compute density that legacy x86 CPUs simply burn too hot to handle.

Meta confirmed the new Arm processors will run side-by-side with its proprietary MTIA chips, alongside existing GPUs from Nvidia and AMD.

True to form, Meta isn't hoarding the hardware layout. The company announced it will release its complete board and rack designs for the Arm AGI CPU under the Open Compute Project later this year, handing the industry a ready-to-use blueprint.

Why it matters

We think this is the clearest sign yet that the AI bottleneck has shifted upstream from GPUs to server orchestration.

When running multi-step agentic workflows, CPUs do the heavy lifting of scheduling, memory handling, and spinning up parallel tasks. Legacy server chips choke on power limits inside dense AI clusters. By doubling efficiency, this silicon allows data center operators to pack vastly more compute into the exact same floor space.

It is also a masterstroke for Arm. Cloud AI head Mohamed Awad made no secret of their strategy: they want to be the default silicon vendor for tech companies that want custom-grade efficiency without the multi-billion-dollar price tag of designing proprietary chips.

For Intel and AMD, it is a nightmare scenario. Meta just validated Arm's production silicon at the highest level and is open-sourcing the blueprints, making it remarkably easy for rivals to ditch x86 in the AI era.