[[[SUMMARY_START]]]

Arm has introduced a data center processor it designed and plans to sell as finished silicon, a shift from its long-running licensing model.
The new “AGI CPU” targets large AI infrastructure, where CPUs coordinate accelerators and manage data movement.
Arm says the processor family scales up to 136 cores and is built on a 3-nanometer manufacturing process.
The move adds a new route for customers alongside Arm’s traditional IP licensing and its pre-integrated Compute Subsystems program.

[[[SUMMARY_END]]]

Arm, best known for designing processor architectures and licensing them to other chipmakers, has taken a major step into selling chips under its own name.

This week the company introduced a data center CPU family called the “AGI CPU,” saying it will be offered as production silicon rather than only as licensable intellectual property.

Arm’s announcement marks a shift for a company that has spent decades supplying the building blocks used by partners, including many of the world’s largest smartphone and server-chip vendors. Instead of only providing CPU cores and related system IP, Arm is now positioning itself to sell a complete processor that customers can deploy directly.

## What Arm announced
Arm described the AGI CPU as a processor designed for what it calls “agentic AI infrastructure.” In practical terms, this refers to CPU-heavy work inside AI data centers, such as scheduling and orchestration around accelerators, moving data between memory and compute, and running supporting services that keep large AI systems operating efficiently.

The company said the processor family scales up to 136 cores based on its Neoverse V3 server-class CPU core design. Arm also said the chips are built using a 3-nanometer manufacturing process, with production handled by a contract manufacturer rather than an Arm-owned factory.

In the specifications it shared, Arm positioned the AGI CPU as a high-core-count, high-bandwidth part intended for modern data centers. It highlighted a 300-watt thermal design point for the top configuration and support for multi-channel DDR5 memory to feed large-scale workloads.

## Why this is different for Arm
Arm’s core business has historically been licensing. Chip companies pay for the right to use Arm’s instruction set and core designs, then build and sell their own products.

In recent years, Arm has also pushed toward more complete “platform” offerings. Its Compute Subsystems (CSS) package together CPU cores with other system components, aiming to shorten the time it takes partners to design a full system-on-chip.

Selling an Arm-designed, ready-to-deploy processor goes further. It creates a third option for customers:

1) license core IP and build a custom chip,
2) use a more pre-integrated compute subsystem as a starting point, or
3) buy a finished Arm CPU.

That expansion matters because the competitive landscape in data centers is changing quickly. AI-driven demand is pulling more companies into building their own silicon, including cloud providers and large technology firms that want tighter control over performance, power use, and supply.

## Partners and the customer relationship question
Arm said the AGI CPU was co-developed with Meta as a lead partner. While Arm did not present the move as a replacement for its partner ecosystem, the announcement raises a familiar tension: many of Arm’s customers also design chips, and some may view first-party silicon as new competition.

Arm has argued in recent briefings that the market is growing fast enough to support multiple approaches. Data center operators are also increasingly mixing architectures and suppliers, using different CPUs and accelerators for different layers of AI training and inference.

## What to watch next
The key near-term questions are about adoption and scope.

One is whether Arm keeps the AGI CPU focused on a narrow set of large data center customers, or expands into broader server markets over time.

Another is how Arm manages its relationships with existing chip partners that build Arm-based server processors. If customers can buy a complete Arm CPU, some partners could push harder to differentiate with custom features, chiplets, packaging, or specialized system designs.

Finally, the launch adds another signal that the AI infrastructure buildout is not only about GPUs and custom accelerators. CPUs still handle essential control-plane work at scale, and vendors are increasingly designing CPUs specifically for AI-era data centers.

AI Perspective

Arm’s move into selling finished processors shows how AI infrastructure demand is reshaping long-standing roles in the chip industry. It also suggests that “platform” offerings are becoming a competitive middle ground between pure IP licensing and fully custom silicon. For buyers, the practical test will be whether these new options deliver clear gains in cost, performance, and time-to-deploy at data center scale.