Arm just announced its first arm AGI CPU and has started selling to customers for revenue. Meta, Microsoft, and SAP are among the early adopters. Rene Haas, the company CEO, and Mohamed Awad, Executive Vice President of arm’s Cloud AI Business Unit, did a great job explaining what we can call the next chapter for arm.
As of now, arm is manufacturing hardware and silicon chips to sell to customers for profit. This is the new business strategy that is opening doors for a significantly higher TAM and taking the company to the next level.
Some of the first reactions from many were that arm is now competing with its own customers. This is not untrue if the customer has its own arm-based solution, like AWS or Microsoft; however, selling hardware CPUs provides an additional option for companies that do not want to spend the money or lack the know-how to develop a CPU.
Arm has an open-door policy and continues to offer IP, Compute Subsystems (CSS), or hardware. It is up to the customer to pick and choose.
Developing a chip based on Neoverse V3 cores costs in the hundreds of millions. The price ranges from the low hundreds of millions to close to a billion, depending on how much a customer wants to differentiate and develop from scratch without the help of EDA partners such as Cadence or Synopsys. Based on the keynote walk-on from Meta, the company wants to port its code to arm but obviously doesn’t want to develop its own CPU for AI data centers.
In a sense, the arm AGI CPU is an alternative to Vera, but it won’t hurt or change Nvidia’s strategy for a bit. Jensen and Ian Buck will continue selling Vera (CPU) and Rubin (GPU) together. Intel also has a GPU solution where it uses Intel Xeon 6, as some customers still insist on x86.
If one assumes that hundreds of millions for ground-up SoC/CPU development is not an issue, the pool of available engineers—especially high-profile engineering fellows and senior leads—is incredibly small. The irony is that Gerard Williams III and John Bruno, the leads behind the Nuvia / Snapdragon X Elite engineering success, are possibly free. Manu Gulati, the third co-founder, was free for a few weeks before joining an AI startup.
With chip development, things often go south; there are no guarantees that everything will work as originally imagined, that things will be on time, or that the desired speeds and yields will be met. There is a lot of uncertainty, and on top of that, one needs to secure capacity in very crowded fabs.
It won’t be easy, but currently, the main company making arm data center CPUs is Ampere. From conversations we had at the “arm Everywhere” event, Ampere is competing in a lower performance bracket and serves hyperscalers rather than the AI data center market. I would not be surprised to see a major strategic turn in this SoftBank-owned company.
