A few weeks ago, we headed to Austin, Texas, to the lab where the Arm AGI CPU is coming to life. To me, this was a really interesting one because when I think of Arm, I think of a company that has been around for decades, powering devices from large to small. Still, the Arm AGI CPU is the company’s first foray into a modern data center CPU, so an entirely new layer of work is needed to turn IP into a physical product. In Austin, we toured the lab where that was happening and we have a video that is certainly worthwhile watching.

Thank you to Arm for sponsoring this one so we could travel to Austin and bring you this.

ARM ATE Lab-1

Welcome to the Arm ATE Lab.

How We Got Here

Arm powers everything from smartphones to servers. When we say this, however, we usually mean Arm provides the IP behind those devices. For example, my Apple iPhone and MacBook Pro are the result of an Arm architecture license that Apple uses to design its great CPU cores.

Apple Mac Mini M4 And M1 1

The next step is licensing a core from Arm, and since we are focused on the server space, that often means a Neoverse V-series or N-series core. A great example is the NVIDIA Grace CPU, like this one in the Supermicro ARS-511GD-NB-LCC liquid-cooled AI workstation.

Supermicro ARS-511GD-NB-LCC Super AI Station-NVIDIA Grace CPU and SOCAMM LPDDR5X Memory

Previously, server efforts focused mainly on those with architecture licenses or the Neoverse line. The Neoverse line powers everything from AWS Graviton CPUs to the Ampere Altra servers we have seen. Over time, however, Arm’s customers asked for more than just CPU cores. Instead, organizations building their own server CPUs and DPUs wanted a more drop-in solution.

NVIDIA Grace CPU lscpu output cropped

Over time, Arm went from providing Neoverse cores to providing the Neoverse Compute Subsystem, or CSS. The idea behind the Neoverse CSS is to provide a collection of cores, fabric, and other IP that is already validated and can be put directly into a design, greatly increasing the time to market. Many examples of Neoverse CSS in use exist, including Microsoft Azure Cobalt 100 and Cobalt 200, but one I often use today is the XSight Labs E1 DPU. XSight Labs is a startup we have covered that has a 12.8T switching platform gaining popularity. It also has an 800G-capable DPU with 64 Arm Neoverse CSS N2 cores that boots vanilla Ubuntu Linux 26.04 LTS out of the box.

Xsight Labs OCP 2025 E1 PCIe DPU Card

The major benefit of using Neoverse CSS is the time to market. Companies that wanted to build their own CPUs could take a broader set of Arm IP and build a chip program with an enormous head start. I use XSight Labs as a great example because they used Neoverse CSS to get a high-end DPU to market early, and that, plus its X2 switch, led to a recent $2.8B post-money valuation funding round. Just a few years ago, it would have been very difficult for a startup to build a DPU like the E1 and have it as one of, if not the first, 800Gbps-capable DPUs on the market. It would have needed to build so much from scratch, and larger companies have more resources to parallelize work. Arm Neoverse CSS was a major enabler for a startup to quickly build that type of product.

Xsight Labs E1 DPU Lscpu Output

Given that a major benefit of the Arm Neoverse CSS is speed and time to market, customers had an additional ask. Instead of just providing the Neoverse CSS, Arm is now selling an entire CPU, the Arm AGI CPU. I asked, and Arm is not switching to only selling CPUs overnight. Instead, the company is committed to selling along the continuum of licensing IP, to Neoverse CSS, to production silicon, which is the AGI CPU.

Arm AGI Hot Chips 2026 Neoverse IP Blocks

With that, let us get to what the Arm AGI CPU is, and then what we saw in the lab.