At Hot Chips 2026, Arm disclosed its first in-house mass-produced AI chip in the company’s 36-year history, the AGI CPU. Built on TSMC’s 3nm N3P process with a dual-chiplet design, the chip supports up to 136 cores, CXL 3.0, PCIe Gen6, and up to 6TB of DDR5 memory. Targeting low-latency inference for agentic AI and multi-agent clusters, it marks Arm’s formal entry from IP licensor into the data center AI chip race.

At Hot Chips 2026, Arm unveiled the AGI CPU for agentic AI, the first in-house mass-produced chip in its 36-year history. The processor is based on TSMC’s 3nm N3P process and a dual-chiplet design, integrates 100 billion transistors, supports up to 136 cores, 96 PCIe Gen6 lanes, CXL 3.0, and up to 6TB of DDR5 memory, and primarily targets low-latency AI inference and data center computing.

For years, Arm has served the chip industry through an IP licensing model, with its architecture behind nearly every mainstream mobile chip. By building its own chip, Arm signals that its view of the data center AI market has shifted from “licensing is enough to make money” to “we need to directly define hardware capabilities.” That strategic signal is more noteworthy than the chip’s specifications.

From a product positioning perspective, the AGI CPU is not simply about stacking general-purpose compute. It directly targets low-latency inference workloads for agentic AI and