Nvidia unveils NVLink Fusion, which lets clients use NVLink to couple non-Nvidia CPUs or accelerators with Nvidia's GPUs in their rack-scale setups to boost AI

Nvidia Corp. Chief Executive Officer Jensen Huang outlined plans to let customers deploy rivals' chips in data centers built around …

Bloomberg 2025-05-19

Context & Ripple Effects

Nvidia’s earlier AI-system strategy paired its accelerators with platform software and server partnerships, including the HGX-2 multi-GPU platform. NVLink Fusion extends that rack-scale approach by making the interconnect a bridge to third-party compute rather than only an Nvidia-controlled component.

The immediate follow-on coverage shows why that matters: Qualcomm’s planned data-center CPU link to Nvidia GPUs turns interoperability into a route back into a CPU market dominated by Intel and AMD. Later Arm and SiFive integrations suggest the interface can become a common attachment point across CPU architectures.

First-order effects

Data-center customers can combine Nvidia GPUs with selected non-Nvidia CPUs or accelerators through NVLink Fusion, widening their component choices while retaining Nvidia GPUs at the center of the rack.
Nvidia shifts NVLink from a primarily internal platform advantage toward an ecosystem interface, giving partner chip designers a defined path to attach to Nvidia-centric AI systems.

Second-order effects

CPU and accelerator vendors gain an incentive to support NVLink Fusion rather than require customers to choose a fully separate AI platform; Qualcomm’s announced GPU-linked CPU plan is an early example.
System builders and cloud operators can evaluate more heterogeneous rack designs, while Nvidia can preserve demand for its GPUs even when customers source surrounding compute elsewhere.

Third-order effects

If partner adoption broadens, high-speed interconnect compatibility could become a key control point in AI infrastructure: component diversity may increase, but the rack’s architecture can remain anchored on Nvidia’s GPU and networking ecosystem.
The move is part of a broader contest over whether AI data centers organize around closed full-stack platforms or interoperable, heterogeneous systems; the practical degree of openness will depend on which vendors implement the interface and on deployment terms.

The trend: AI infrastructure is moving toward heterogeneous rack-scale systems in which interconnect standards determine who can participate—and who retains platform control.

Discussion

@edzitron.com Ed Zitron on bluesky
it's so funny seeing this stuff get covered in the mass tech media because of AI. don't remember bloomberg covering AMD crossfire :/ [embedded post]
@benbajarin Ben Bajarin on x
What is SUPER interesting about @NVIDIA announcing NVLink Fusion, is the making a compelling case for their network backbone for semi custom infra using whatever parts you want including CUSTOM CPUS and CUSTOM ASICS. NVIDIA has a play even in the custom semi trend.
@iancutress @iancutress on x
Announcing NVLink IP for licensing through partners. Alchip Marvell Mediatek Enable any ASIC or cpu to integrate into NVIDIA's ecosystem [image]
@neiltwitz Neil Shah on x
@BradyWang10 ... NVIDIA announces “NVLINK Fusion” a “semi custom” system leveraging NVLINK components switches, NICs, Spectrum-X - mix and match with any special ASIC or compute with NVLINK IP This is big!! Great ecosystem play leveraging NVLINK infra #Computex #GTC25 [image]
@neiltwitz Neil Shah on x
@BradyWang10 ... Meet the new NVIDIA RTX Pro Server for classical server applications in addition to running Omniverse and multimodal Enterprise AI agents #Computex #GTC25 [image]
@maxwinebach Max Weinbach on x
Nvidia's NVLink Fusion is a pretty big deal. NVLink for anything. Nvidia wants NVLink to be the industry standard for all networking, using chiplets to connect NVLink to any custom silicon
@benbajarin Ben Bajarin on x
NVLink Fusion, fits nicely into something Jensen said a few years ago in a QA I was in. @NVIDIA is happy to sell you a completely integrated system, or just parts of the system. He wants there to be an NVIDIA part for everyone no matter big or small.
@iancutress @iancutress on x
Announcement: World's most advanced physics engine, Newton, will be open-source in July, Super real time. Integrated into MuJuCo and Isaac. GPU accelerated of course. [image]
@theaustinlyons Austin Lyons on x
Bring your ASIC, or bring your CPU [image]

Chronicles