At the Supercomputing ’25 conference, Arm and NVIDIA announced that Arm has joined the NVLink Fusion ecosystem, marking the support of this technology by two major microarchitecture developers and a total of four CPU developers.
For NVIDIA, this means that Arm’s customers will be able to develop processors that can work in conjunction with NVIDIA AI accelerators, while Arm will also be able to design CPUs based on NVIDIA systems that compete with NVIDIA’s own or Intel processors.
NVIDIA’s Data Center Product Marketing Director, Dion Harris, stated: “Arm is integrating NVLink IP so that its customers can build CPU SoCs that connect to NVIDIA GPUs. With NVLink Fusion, hyperscale data centers can significantly reduce design complexity, save development costs, and bring products to market faster. The addition of Arm customers provides more options for specialized semi-custom infrastructure.”
Arm is a diversified large company whose business includes ISA and IP licensing, as well as developing custom CPUs and system-on-chips (SoCs) for large clients. For each type of business, support for NVLink Fusion brings specific benefits.
As an IP provider, by directly integrating NVLink IP into its architecture product portfolio, Arm can provide its licensees with a ready-made path to build CPUs that can natively access the NVIDIA AI accelerator ecosystem. Theoretically, this makes Arm-based designs more attractive to hyperscale data centers and sovereign cloud builders who wish to obtain custom CPUs and want compatibility with market-leading NVIDIA GPUs in the AI and high-performance computing (HPC) fields. Previously, NVIDIA’s Grace CPU was the only processor supporting NVLink connections that could be compatible with NVIDIA GPUs.
Although NVIDIA positions Arm solely as an IP provider, Arm itself, as a manufacturer developing CPUs for hyperscale data centers and sovereign organizations, has gained the ability to compete directly in NVIDIA-based systems. With native NVLink Fusion integration, future server CPUs designed by Arm can compete head-to-head with NVIDIA’s Grace and Vera processors, as well as Intel’s Xeon processors, in systems centered around NVIDIA GPUs. With NVLink Fusion, Arm CPUs are expected to become first-class participants in rack-level NVLink solutions.
Furthermore, support for NVLink Fusion also enhances Arm’s position as an ISA licensor, as it makes Arm architecture more inherently attractive to hyperscale data centers and chip designers looking to customize CPUs that closely integrate with NVIDIA GPUs. By ensuring that Arm-based CPU designs can work with a consistent NVLink interconnect architecture (rather than being limited to PCIe) in conjunction with NVIDIA GPUs, Arm gains ecosystem appeal and future relevance that competing architectures like x86 and RISC-V currently cannot match. Undoubtedly, this poses a risk to both AMD and Intel. The former is almost disinterested in supporting NVLink, while the latter is years away from building a custom Xeon CPU that supports NVLink for NVIDIA’s rack-level systems. However, considering factors like chip development cycles, by the time Arm CPUs supporting NVLink are ready, Intel’s custom Xeon CPUs may also be prepared.
Of course, Arm’s support for NVLink Fusion also benefits NVIDIA, as it significantly expands the pool of CPUs that can natively use NVLink in NVIDIA-centric AI systems without NVIDIA having to build all these CPUs themselves. By enabling Arm’s licensees (such as Google, Meta, and Microsoft) to directly integrate NVLink into their SoCs, NVIDIA can ensure that future Arm processors will be architected around NVIDIA GPUs or at least be compatible with them. On one hand, this may reduce the appeal of open alternatives like UALink; on the other hand, it will also diminish the overall attractiveness of AI accelerators from companies like AMD, Broadcom, and Tenstorrent.