NVIDIA has closed—at least financially—one of the most iconic chapters in its recent history: the company has divested its last stake in Arm Holdings. According to regulatory filings, the sale includes 1.1 million shares valued at approximately US$140 million and was completed during the fourth quarter of 2025, leaving its holdings in Arm at zero.
This move is not just accounting. It comes at a time when the industry is reevaluating the role of the CPU in the new wave of agentic AI: assistants capable of chaining steps, calling tools, querying memory, making API requests, and coordinating entire workflows. This “choreography” isn’t solely GPU-based. In many real-world scenarios, overall latency depends on less glamorous components—queues, orchestration, networks, caches, data serialization—where the CPU once again becomes central.
A memory-driven shift: from a failed acquisition to total exit
To understand why this sale has generated so much noise, we need to look back. NVIDIA had agreed to acquire Arm in 2020 for around US$40 billion, but the deal fell apart in 2022 following intense regulatory pressure in various markets. Since then, Arm continued its stock market journey, and NVIDIA maintained a minority stake that is now fully liquidated.
The official (and most cautious) interpretation is portfolio optimization: freeing capital, simplifying exposure, and turning the page. But the context has many viewing this decision as an indirect signal within a broader debate: which CPU architecture “fits best” as AI shifts from training to production, agents, and millions of micro-tasks per second.
ARM vs x86: the inference-driven debate returns
For years, Arm was associated with efficiency and scalability, and NVIDIA embraced this vision strongly: its designs with Arm CPUs appear in data center and supercomputing products. Yet, in the agentic era, nuances are emerging that complicate this picture.
One part of the technical argument revolves around many agent scenarios prioritizing performance peaks per thread and response times, as well as seamless integration with enterprise ecosystems where x86 remains the “native language”: hypervisors, toolchains, libraries, legacy environments, base images, historical optimizations, and decades of operational experience. In other words, for a company, the cost of switching architectures isn’t just about the server—it’s everything surrounding it.
The practical takeaway for system and development teams is that bottlenecks rarely announce themselves clearly. Sometimes the GPU is “idle,” waiting for an agent’s decision; other times, the problem lies in the pipeline (queues, I/O, network, tokenization, compression, RPCs, caches), where the CPU sets the pace more than many budgets have assumed two years ago.
Vera, Rubin, and ARM continuity… with an open door to x86
This is where the debate gets interesting: NVIDIA does not appear to be abandoning Arm in its products, even if it has exited ownership.
In its public roadmap, the company has introduced the Vera Rubin platform, with the “Vera” CPU based on Arm (with a custom design focused on large-scale AI performance). In other words, architecturally, Arm remains part of NVIDIA’s data center vision.
At the same time, signs of strategic rapprochement with x86 are emerging. For example, a significant collaboration with Intel has been announced, including technology integration and industrial partnership, which many analysts interpret as a way to reduce friction with the x86 universe, especially in deployments where compatibility and inertia are as important as raw performance.
Translated to data center language: although NVIDIA’s “own” CPU might still be Arm in certain systems, it’s plausible that the ecosystem will evolve toward more heterogeneous racks and platforms, where accelerators, high-speed networks, and x86 or Arm CPUs coexist depending on the use case, software requirements, vendor, and operational reality.
Implications for sysadmins and developers
For technical professionals, this news is less about actions and more about decisions:
- Platform planning: if your AI strategy involves agents (not just chat), start measuring CPU as a critical resource: latency per call, queues, P95/P99, thread saturation, and tooling overhead.
- Compatibility and technical debt: in legacy-heavy organizations, x86 will continue to dominate due to organizational friction. Arm can shine, but migrating runtimes and pipelines isn’t “just recompiling”.
- Hybrid architectures: the sensible trend is toward more mixing, not less. Less dogma (“GPU-only / Arm-only”) and more engineering (“this needs peak per thread, this needs throughput, this needs compatibility”).
- Strategic purchases for a 3-year horizon: agentic AI is shifting cost centers—training is just part; serving is a major factor; and serving requires end-to-end efficiency.
Ultimately, selling the last of its Arm shares doesn’t doom any architecture. What it does do is send a clear message: The AI race is no longer just on the GPU. When the market understands this, the CPU ceases to be “business as usual” and reemerges as a strategic battleground.
Frequently Asked Questions
Why does NVIDIA’s sale of Arm shares matter?
It completes the divestment after the failed acquisition attempt and comes at a moment when the industry is reassessing which CPU architecture suits large-scale agentic AI and inference.
Does agentic AI require more CPU than traditional AI?
It generally increases dependence on CPU for orchestration: calling tools, APIs, memory, queues, and coordination. In many deployments, the total latency and perceived performance rely heavily on this “glue”.
Does this mean NVIDIA will abandon Arm CPUs for data centers?
Not necessarily. NVIDIA’s roadmap for Vera Rubin continues to include Arm-based CPUs, implying technical continuity despite a change in ownership.
What should sysadmins watch when deploying AI agents?
Latency metrics (P95/P99), thread saturation, I/O and network timings, internal queues, and operational costs related to compatibility (containers, libraries, hypervisors, toolchains)—because bottlenecks may lie outside the GPU.