Arm is no longer just a licensing company.
On Tuesday, Arm announced the Arm AGI CPU, its first in-house data center chip and the result of a development partnership with Meta. The chip is built for the agentic AI era, specifically for inference workloads and the distributed task management that comes with AI systems that plan, reason, and act with minimal human oversight.
Meta is the lead partner. It’ll run the new chip alongside its own custom MTIA silicon as it continues to build out infrastructure for a portfolio of AI products serving billions of people.
What Arm actually built
The AGI CPU packs up to 136 Neoverse V3 cores per chip, running DDR5-8800 memory with six gigabytes per second of bandwidth per core at sub-100-nanosecond latency. A single air-cooled rack can hold 64 CPUs delivering roughly 8,160 cores. Push to liquid cooling, Arm has a Supermicro partnership for this, and that climbs past 45,000 cores per rack.
Arm’s core claim: more than twice the performance per rack versus x86 platforms. Translated to infrastructure economics: up to $10 billion in capital expenditure savings per gigawatt of AI data center capacity. That’s a real number for companies watching power budgets become the binding constraint on AI scale-up.
Volume production is targeted for the second half of 2026, manufactured on TSMC’s 3-nanometer process.
Why Meta being first matters
Meta has been working toward a diversified silicon stack for years. Its in-house MTIA chips handle training and inference tasks, but reportedly struggled to hit the performance roadmap Meta needed. The Arm partnership fills that gap, and gives Meta a partner with 35 years of CPU architecture expertise.
This isn’t a one-off deal. Arm and Meta are committed to multiple generations of the AGI CPU roadmap. Meta is also open-sourcing its board and rack designs for the chip under the Open Compute Project later this year, a signal it’s fully bought in rather than treating this as a short-term workaround.
For context: Meta’s 2026 capex guidance sits at $115–135 billion. If Arm captures even a mid-single-digit percentage of that, the revenue impact is material. Analyst Patrick Moorhead estimated to CNBC that just 5% of Meta’s capex could be “a game changer on the top line” for Arm.
Beyond the chip: what this signals about the AI infrastructure race
The GPU gets the headlines. But the CPU has become the bottleneck.
Nvidia recently told CNBC that CPUs are “becoming the bottleneck” as agentic AI changes how workloads are distributed across hardware. CPUs manage memory and storage, schedule workloads, orchestrate accelerators, and coordinate thousands of distributed tasks in real time. As AI systems scale, and as agents spawn more sub-tasks, the demands on the CPU layer are compounding fast.
Arm’s CEO Rene Haas told Reuters the chip business could add roughly $15 billion in annual revenue within five years, contributing to a broader Arm outlook of $25 billion in total annual revenue within the same window.
The broader ecosystem is moving in the same direction. OpenAI, Cloudflare, SAP, SK Telecom, Cerebras, and others have made commercial commitments to the AGI CPU. Arm has additional chip designs in the pipeline planned at 12-to-18-month intervals.
There’s also a supply context worth noting. Intel and AMD have told China customers of lengthening delivery wait times, per Reuters reporting. Computer prices are beginning to rise. The CPU shortage adds another pressure layer on companies racing to build out AI infrastructure.
The takeaway
Arm spent three and a half decades building a business on licensing designs to others. Now it’s shipping its own silicon, and one of the most aggressive AI infrastructure spenders on the planet is the first customer.
That’s not a small shift. It’s a signal that the AI infrastructure race is expanding beyond GPUs, that power efficiency at scale is the next battleground, and that the companies willing to commit to custom silicon partnerships early are trying to lock in competitive advantage before the supply chain catches up.
Whether you’re a chip nerd or someone who just wants to know if this matters for the broader AI landscape — it does.
