NVIDIA Sells Off Remaining ARM Shares in $140 Million Exit

NVIDIA has sold its remaining ARM stake worth about $140 million, according to SEC filings, signaling a financial reset amid shifting AI CPU dynamics.

NVIDIA exits ARM investment as AI CPU race intensifies

According to Wccftech, NVIDIA has divested its remaining stake in Arm Holdings, valued at roughly $140 million. The move formally closes the chapter on NVIDIA’s earlier attempt to acquire ARM outright, a deal that collapsed under regulatory pressure.

The sale comes at a time when CPUs are gaining renewed importance in AI infrastructure, especially for inference and emerging agent-based workloads.

ARM is still central to NVIDIA’s AI roadmap

Despite the divestment, ARM architecture remains deeply embedded in NVIDIA’s data center strategy. The company’s Grace Hopper and Blackwell platforms rely on ARM-based CPU designs, and the upcoming Vera CPUs will also use ARM cores.

NVIDIA’s next-generation Vera chips are expected to power future AI server systems, while Rubin GPUs are projected to drive HBM4 memory demand across hyperscale deployments.

At the same time, NVIDIA CEO Jensen Huang recently teased “chips the world has never seen” ahead of GTC, fueling speculation about future architectural shifts.

CPUs become critical in AI inference and agent workloads

While GPUs still dominate AI training, inference workloads increasingly depend on CPU-side tasks. These include API calls, orchestration, memory lookups, tool execution, and agent coordination.

Agentic AI systems often process millions of microtasks per second. Even minor CPU delays can leave GPUs idle, creating costly bottlenecks in high-performance clusters.

Analysts suggest ARM-based CPUs may struggle with certain AI server use cases, particularly where GPU scheduling efficiency and tight x86 ecosystem integration matter.

x86 regains momentum in data centers

Hyperscalers are currently upgrading server fleets, boosting demand for x86-based CPUs from Intel and Advanced Micro Devices. Enterprise data centers remain heavily optimized around x86 firmware, virtualization layers, and compiled software stacks.

Some reports indicate NVIDIA is exploring deeper x86 integration, potentially through partnerships with Intel for NVLink-based server racks. Such a move could improve CPU-GPU coordination in large AI clusters.

Still, NVIDIA describes the ARM stake sale as a financial decision rather than a strategic withdrawal from ARM. Vera CPUs remain fully ARM-based, and no immediate architectural pivot has been confirmed.

Beyond data centers, AI hardware momentum continues to accelerate. NVIDIA’s Rubin GPUs are expected to increase demand for next-generation HBM4 memory, intensifying competition across the supply chain.

Meanwhile, Intel projects that more than 50% of PCs could qualify as AI PCs in 2026, reflecting growing emphasis on on-device inference and dedicated NPUs.

As AI workloads evolve from large-scale pre-training toward inference and agent-driven systems, CPU architecture choices may become just as critical as GPU performance. NVIDIA’s ARM divestment highlights how fluid that competitive landscape remains.