At Computex 2025, Nvidia unveiled a major advancement in its AI hardware strategy with the launch of NVLink Fusion, a new version of its high-speed chip interconnect technology. Unlike previous iterations, NVLink Fusion will be sold to other chip designers, allowing them to build powerful custom AI systems by linking multiple chips together. Companies such as Marvell Technology and MediaTek have already announced plans to adopt the technology for their own custom chip efforts. NVLink has long been used in Nvidia’s own products—such as the GB200, which combines two Blackwell GPUs with a Grace CPU—to enable rapid chip-to-chip communication necessary for AI workloads.
Nvidia CEO Jensen Huang made the announcement during his keynote at the Taipei Music Centre, the venue for this year’s Computex exhibition, which is hosting around 1,400 exhibitors. He also revealed plans for Nvidia to build a new Taiwan headquarters in the northern suburbs of Taipei, reaffirming the company’s strong ties to the region. Huang reflected on Nvidia’s evolution from a video game graphics chipmaker into the dominant force behind the AI revolution, noting that what used to be presentations focused almost entirely on graphics chips are now cantered on AI systems and applications.
In addition to NVLink Fusion, Huang provided updates on Nvidia’s AI chip roadmap, which includes the Blackwell Ultra, due later this year, followed by the Rubin series and then the Feynman processors expected in 2028. He also announced that the company’s new desktop AI system, DGX Spark, designed for AI researchers, is now in full production and will be available within a few weeks. These developments illustrate Nvidia’s strategic shift from simply building AI models to powering their deployment and real-world application, solidifying its position at the centre of the global AI ecosystem.
