Arm Unveils AGI CPU: Its First In-House Silicon in 35 Years Targets AI Data Centers

A Historic Pivot for Arm

For 35 years, Arm has operated as the semiconductor industry's most successful IP licensor, designing processor architectures that power virtually every smartphone on the planet while leaving chip manufacturing to partners like Qualcomm, Apple, and Samsung. On March 25, 2026, that model changed. Arm announced the AGI CPU, a production-ready data center processor that the company designed, validated, and will sell as finished silicon. It marks the first time Arm has shipped its own chip product, a move that repositions the company from IP supplier to direct competitor in the lucrative data center market.

The timing is deliberate. As AI workloads explode in data centers worldwide, the demand for CPU-side orchestration power, the silicon that coordinates accelerators, manages data movement, and runs inference pre- and post-processing, has outstripped what existing x86 platforms can efficiently deliver. Arm is betting that its power-efficient architecture, proven in mobile, can capture this emerging "agentic AI infrastructure" layer.

Technical Specifications

The AGI CPU is built on TSMC's 3nm process node and features up to 136 Neoverse V3 cores arranged across two chiplets. The cores run at up to 3.2 GHz all-core frequency with a 3.7 GHz single-core boost, all within a 300-watt thermal design power (TDP) envelope.

Memory bandwidth is substantial: the chip supports 12 channels of DDR5 memory at speeds up to 8,800 MT/s, delivering aggregate bandwidth exceeding 800 GB/s. This is critical for AI workloads where data movement between CPU and accelerator is often the bottleneck.

Rack-Scale Density

Arm is not just selling chips; it is selling a rack-scale story. A standard air-cooled 36kW rack holds 30 AGI CPU blades, totaling 8,160 cores. For hyperscalers willing to invest in liquid cooling, Arm has partnered with Supermicro on a 200kW configuration that packs 336 chips and more than 45,000 cores into a single rack. Arm claims this delivers more than twice the performance per rack compared to the latest x86 platforms.

Partners and Early Commitments

Meta is the lead development partner, having co-designed the chip with Arm. The partnership reflects Meta's massive AI infrastructure buildout and its desire to diversify beyond x86-only server fleets. Additional commercial commitments come from Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom, a list that spans cloud infrastructure, AI accelerator companies, and enterprise software.

The breadth of this partner list signals that Arm's move is not speculative. These are binding commitments from companies already operating at hyperscale.

Competitive Implications

The AGI CPU puts Arm in direct competition with Intel and AMD in the data center CPU market. While Amazon's Graviton and Ampere's Altra have already demonstrated that Arm-based servers can compete on performance-per-watt, neither chip was designed or sold by Arm itself. The AGI CPU changes this dynamic: Arm now competes with its own licensees as well as the x86 incumbents.

For Intel and AMD, the threat is real but nuanced. Arm is specifically targeting the AI orchestration workload, not general-purpose server computing. However, workload boundaries in data centers are fluid, and a successful AGI CPU could expand into adjacent markets over time.

Outlook

Arm's entry into direct silicon sales is the most significant strategic shift in the company's history. If the AGI CPU delivers on its performance and efficiency claims at scale, it could reshape how AI data centers are architected, shifting the CPU layer from a commodity afterthought to a purpose-built component of the AI stack. The partnership with Meta and commitments from OpenAI and Cerebras lend credibility, but production ramp and real-world benchmarks will be the ultimate test.

Conclusion

The Arm AGI CPU represents a calculated bet that the AI data center market is large enough and differentiated enough to justify Arm building its own silicon. With a 136-core, 3nm design delivering over 800 GB/s memory bandwidth at 300W, the specs are competitive. The real question is execution: can Arm manufacture, sell, and support a production chip at data center scale? The answer will determine whether this historic pivot becomes a template for the next era of semiconductor competition.