Arm wrestles away 25% share of server market thanks to Nvidia's home-grown CPUs

Still far short of the 50% market share Arm infra chief was hoping for

Nvidia isn’t the only one riding the AI boom. During the second quarter, Arm CPUs captured a quarter of the server market, according to a recent Dell’Oro Group report.

That’s up from 15 percent of the server market a year ago, Dell’Oro analyst Baron Fung tells El Reg.

The driving force behind this trend is the adoption of Nvidia’s Grace-Blackwell rack-scale compute platforms like the GB200 and GB300 NVL72.

Each of the 120-kilowatt machines is equipped with 72 Blackwell GPUs and 36 of Nvidia’s Grace CPUs. The 72-core chips, first introduced in 2022, are based on the British chip designer’s Arm Neoverse V2 architecture and have been optimized to maximize data movement by taking advantage of Nvidia’s custom NVLink-C2C interface.

The first of these systems began shipping in small volume late last year. A refreshed version of the racks based on Nvidia’s Blackwell Ultra architecture began trickling out in the second quarter with neocloud operators like CoreWeave being among the first to take delivery.

To be clear, these machines shouldn't be confused with Nvidia's air-cooled B200 and B300 boxes, which are still relying on Intel CPUs to wrangle the GPUs.

According to Fung, a year ago, Arm’s server share was driven almost exclusively by custom cloud silicon like AWS Graviton. Now he says that Grace revenues are comparable to cloud GPUs.

Amazon Web Services (AWS) has been dealing in custom Arm silicon going back to 2018. However, it's only in the last few years that Microsoft and Google have taken the RISC instruction set seriously with their Cobalt and Axion CPUs, respectively.

As with Graviton, these chips are used for a combination of internal workloads and customer-facing cloud workloads.

As strong as Arm’s datacenter gains may be, 25 percent is a far cry from the 50 percent target Arm infrastructure chief Mohamed Awad was hoping to nab by the end of 2025.

Arm could see its market share grow as more chip designers bring server chips to market. Nvidia is already working on a new Arm-based CPU using custom cores called Vera.

Qualcomm and Fujitsu are also working on new and/or refreshed server chips, both of which have been qualified for use with Nvidia’s NVLink Fusion tech, which means we may see versions of Nvidia’s Superchip architecture based on more than just its own CPUs before long.

Arm’s ambitions aren’t limited to server domination either. Last summer, CEO Rene Haas predicted that, by 2029, half of all Windows PCs sold would be Arm powered.

Arm isn’t the only one seeing gains amid the AI boom. According to Dell’Oro, the ongoing AI expansion cycle drove the server and storage component market to 44 percent YoY growth in Q2.

SmartNIC and data processing unit (DPU) sales have roughly doubled since last year amid a shift toward Ethernet for AI compute clusters. At the same time, Fung tells us that custom AI ASICs are now shipping in volumes comparable to GPUs, though he notes that GPUs still account for the majority of accelerator revenues.

Looking ahead, Dell’Oro expects the server and storage system component market to grow 46 percent in 2025. ®