In a development that could disrupt the artificial intelligence hardware world, Nvidia today at Computex 2025 disclosed plans to sell its cutting-edge chip-to-chip interconnect technology, NVLink Fusion, to other companies that make semiconductor parts.
This unprecedented move is intended to promote collaboration that will increase the pace of innovation in high performance AI in a way that could not easily be achieved by the industry: from data, to design, to testing models, to producing products, and training, all segments of the AI technology will be covered.
NVLink Fusion, the newest generation of Nvidia’s in-house NVLink technology, opens the door to ultra-fast, power-efficient inter-chip communication.
The rapid flow of data between disparate servers is for now a specialized requirement of the big leagues of internet operation, but it is becoming more important as computers grow ever more bloated with fuel for the current era’s stupendous AI workloads, for tasks like training enormous language models and running detailed simulations.
Previously, NVLink technology has mostly been used inside Nvidia’s own products, including its higher-end GPUs and CPU-GPU package systems like the GB200.
Announced today during Nvidia’s press keynote at CES in Las VEGAs, the partnership with Taiwan-based TSMC will give other companies using TSMC’s 16-nanometer process technology the ability to create their own custom AI hardware that works with the company’s years-old Volta processing unit.
NVLink Fusion will offer substantially higher bandwidth than prior forms of interconnect opening up new and exciting future multi-chip design opportunities, he added.
Several leading semiconductor companies have already announced their intention to incorporate NVLink Fusion. Marvell Technology and MediaTek will be some of the first companies to feature the new technology on their forthcoming chip designs.
Nvidia also mentioned that Alchip, Fujitsu, and Qualcomm were key partners for this effort. This widespread adoption indicates the industry’s acknowledgement of the severe requirement of improved chip-to-chip communication in the age of the rapidly developing AI.
Nvidia licensing NVLink Fusion would be a huge deal, analysts claim. By opening up its interconnect technology, Nvidia is not just making sure it’s the leader and maker of the leading chips, but the enabler of the larger ecosystem around AI hardware.
Such a strategy could promote increased interoperability between different forms of processors and accelerators, open the way to more and more powerful AI systems.
And, it could ultimately reinforce Nvidia’s market-leading position in AI by making its technology a key part of a much larger pool of hardware solutions.
Nvidia will still develop and sell its own chips, but licensing NVLink Fusion creates new revenue, and reinforces its power over the future of AI infrastructure.
In a speech, besides introducing the Nvlink Fusion, Huang also disclosed that Nvidia is planning to build a new headquarters in Taiwan, emphasizing the importance of the island in the global semiconductor manufacturing network.
He also gave a preview of the DGX Spark, a new desktop AI workstation for AI researchers that extends the scope of Nvidia beyond the data center.
Nvidia’s strategic move is a key point in its transformation from a GPU company to a company that is taking holistic, strategic control of the hardware of AI.
By enabling others to build upon its leading interconnect technology, Nvidia hopes to speed up AI innovation and further strengthen its place at the center of this disruptive technology.