During several announcements during SC25, one was that Arm is joining the NVIDIA NVLink Fusion ecosystem. This will be one to watch as it changes how accessible the NVLink Fusion technology will be.
Arm Joins the NVIDIA NVLink Fusion Ecosystem
Arm joined Fujitsu, Intel, and Qualcomm as CPU partners. From a market perspective, this is a significant development. We covered this during the Computex 2025 keynote, but NVLink Fusion serves as an onramp to NVIDIA platforms for non-NVIDIA silicon.
Fujitsu is building Japan’s future FugakuNEXT supercomputer with NVIDIA for a target of 2030. It makes sense that Fujitsu is using NVLink Fusion given the project. Qualcomm is notable because it has announced new integrated AI racks utilizing its accelerators, and the HUMAIN MOU also calls for Qualcomm data center CPUs. Intel and NVIDIA inked a big deal for Intel to become the x86 CPU provider for NVIDIA systems, which is even more important now that we just broke Intel decided to cancel its mainstream next-gen Xeon server processors. The future NVIDIA x86 Superchip will need NVLink fusion to link the x86 CPUs sourced from Intel and the NVIDIA GPUs.
Arm is different. As Arm has transitioned from a core IP provider in the datacenter to effectively helping organizations stamp out CSS-based solutions, it provides a large ecosystem where it is much easier to develop new chips. Arm is making its AMBA CHI C2C (Coherent Hub Interface Chip-to-Chip) interface with NVLink Fusion. This makes a lot of sense. For a hyperscale client building its own Arm chips, one could see how they might want to use their own Arm designs to be the CPUs powering AI clusters with NVIDIA GPUs.
For its part, NVIDIA is working to standardize many of the components in these AI systems to increase reliability and uptime. To that end, its next-generation NVIDIA Vera will utilize custom Arm cores instead of the Arm Neoverse V2 cores that power current NVIDIA Grace platforms.
While NVIDIA is pushing its ecosystem largely to Arm, the inclusion of Arm adds another layer to how companies can design custom CPUs and interface with NVIDIA GPUs.
Final Words
This is one of those announcements that is notable because it shows that there is interest in connecting Arm CPUs with NVIDIA GPUs outside of just using NVIDIA’s Arm-based CPUs. That makes a lot of sense. It also makes a lot of sense why Arm would want to enable this since you can think of how specific use cases and customers might want an option to use a different CPU but still use NVIDIA GPUs.
In a way I find this a bad development. There was a strong push for open industry standards based on Infinity Fabric ‘UALink’. All the big players AMD, ARM, Intel and many more were participating. But now one by one they seem to jump ship. It is a bit concerning.