NVIDIA Sells Remaining ARM Stake; Can Its CPU Architecture Fuel the Next AI Revolution?

NVIDIA has finalized the sale of its remaining stake in ARM, marking a significant shift from its earlier ambitions to acquire the company. This decision underscores the evolving landscape of AI hardware and may reflect broader trends in the industry.

NVIDIA Partners with Intel: A Signal of ARM’s Limitations in Agentic AI

The collaboration between NVIDIA and ARM has been instrumental in shaping the current AI infrastructure, particularly through the development of cutting-edge CPU architectures like the Grace Hopper and Blackwell series. As NVIDIA gears up to introduce its new Vera CPUs, the spotlight falls on ARM, particularly as questions arise regarding its future role in the AI ecosystem. Recent reports from Bloomberg reveal that NVIDIA’s offloading of its residual $140 million stake in ARM coincides with debates about ARM’s effectiveness in the rapidly evolving AI landscape.

As the demand for CPUs surges, particularly for agent-based inference workloads, the industry is witnessing a shift from GPU-centric computing to CPU-driven tasks. This transformation is illustrated by Intel and AMD’s rapid success in catering to hyperscalers, who are increasingly seeking robust data center CPU solutions. Compelling insights from GF Securities highlight the growing concerns about ARM’s architecture:

GFHK noted in their Q&A following their February report that ARM-based CPUs have relatively weak momentum in AI servers, attributing this to lower GPU scheduling efficiency compared to x86. They indicated that companies including NVIDIA plan to develop x86 CPU solutions in response.

– via Jukan (analyst at Citrini)

Several factors contribute to the perception that x86 architectures excel in managing agentic workloads. Notably, the emphasis on ‘single-threaded burst speeds’ in agentic scenarios often outweighs the benefits of multicore processing. In environments executing millions of microtasks per second, any latency—even milliseconds—can induce significant processing delays.

Grace Hopper architecture — Grace Hopper | Image Credits: NVIDIA

Moreover, a critical advantage for x86 processors lies in the entrenched ecosystems of enterprise data centers. These environments, built over years with specific firmware stacks and virtualization layers, predominantly align with x86 architecture. This entrenched infrastructure explains the soaring demand for server products from Intel and AMD, especially as hyperscalers transition through a significant upgrade phase. For NVIDIA, transitioning to an x86 server framework seems increasingly advantageous.

Intel Xeon processor in lab — Image Source: Intel

It is important to note that NVIDIA’s strategic pivot towards x86 is not merely speculative; the company is actively pursuing this route through its recent collaboration with Intel. This partnership aims to develop x86-equivalent solutions within NVLink-integrated server racks. The decision to divest from ARM appears to be primarily a financial maneuver, lacking direct implications on NVIDIA’s future product direction.

Currently, NVIDIA’s Vera CPUs are based on ARM technology. However, the likelihood of the company diversifying its CPU offerings to incorporate Intel’s x86 ecosystem in future product lines, such as Feynman, remains on the horizon.

Source&Images