Nvidia NVLink Fusion Is More Noise Than Substance

Nvidia (NVDA) CEO Jensen Huang announced at Computex 2025 keynote that Nvidia is opening its NVLink for adoption by other players. Of course, as with any Jensen exhaust, this has taken a life of its own and being pitched by some as the death knell for Advanced Micro Devices (AMD) led UALink interface.

Lost in the hype is that Nvidia threw open NVLink for adoption OVER 3 YEARS BACK in 2022. Noticeably, there have been no public takers to that effort. There is a reason for this. Nvidia wants to control NVLink spec and its future. And, no AI Accelerator designer, whether it is a merchant semiconductor outlet like AMD or Intel (INTC), or a hyperscaler designing custom solutions wants to be at the mercy of Nvidia.

Two years after Nvidia tried to get the industry to adopt NVLink, in May 2024, the industry coalesced around an AMD led UALink consortium. That was as sound a repudiation of Nvidia’s NVLink effort as any.

By forming a UALink consortium, AMD, Broadcom (AVGO), Cisco (CSCO), Google (GOOG) (GOOGL), Hewlett Packard Enterprise (HPE), Intel, Meta (META) and Microsoft (MSFT) announced that they have aligned to develop a new industry standard dedicated to advancing high-speed and low latency communication for scale-up AI systems linking in Data Centers. A few months later, Amazon (AMZN) Web Services, Astera Labs (ALAB) added their names to the list. Subsequently, Apple (AAPL), Alibaba (BABA), and Synopsis (SNPS) also joined the group. Broadcom has taken a bit of a backseat along the way for undisclosed reasons but the list of companies supporting or adapting UALink has grown substantially since then. The current list reads like the who-is-who of hyperscalers and IC supply chain (except Nvidia) and can be found here.

If we look at Nvidia’s Computex NVLink Fusion PR, we see that the current supporters are: MediaTek, Marvell, Alchip, Astera Labs, Synopsys and Cadence. All these names, except for Alchip, can also be found on UALink consortium website. Note that Synopsis and Cadence are tool and IP providers and not end customers and MediaTek was already using NVLink as part of its chip partnership with Nvidia.

No hyperscalers, who are the customers who are designing the AI ASICs, are in the NV Link list. So, what is going on here?

Not much, it seems. NVLink was made available years back and there were no end customer takers then and there no takers now (with the possible exception of Fujitsu).

The Nvidia business model for NVLink Fusion gives some clues. The following image is from Serve The Home.

As can be seen from the image, Nvidia is licensing (??) the NV Link technology for customers who are developing CPUs or GPUs but not both. In other words, all nodes MUST contain at least one Nvidia solution – either a CPU or an accelerator.

What? This “must have Nvidia” requirement completely defeats the plug and play aspect of interfaces such as NVLink. It is hard to see any hyperscaler adopting NVLink Fusion unless it is a part of a larger strategic agreement with Nvidia. The NVLink Fusion alone is not going anywhere with these players.

A niche CPU developer like Fujitsu and Qualcomm who may be limited in what they can develop, or implement may be attracted to the NVLink option hoping to connect to Nvidia infrastructure, but it is hard to see many going on this route. To the extent CPU developers go this route, they will be dependent on Nvidia which has a long history of being ruthless to its partners. The odds very much favor the so-called partners getting burnt here.

As far as accelerator customers go, which accelerator customers would want to connect to an underpowered Grace or Vera CPUs which are likely to lag their AMD competitors by a generation or two? It is unclear what benefits partners can get by connecting to Grace or Vera.

Companies like Alchip and Marvell can offer NV Link fusion in their ASIC libraries, which is what they seem to be committing to, but which hyperscaler customer is likely to use the IP? The prospects of hyperscalers adopting NVLink IP seem sketchy at best.

All things considered, NVLink fusion is an interesting effort by Nvidia to rope along sundry players to its NVLink infrastructure but is not likely to make much headway. And there is little evidence at this point that this can be a competitor to UALink let alone being a UALink killer.