Nvidia to sell NVLink Fusion to speed AI chip communication industry-wide
Developed at Nvidia’s Israel R&D center, the new NVLink Fusion interconnect will be licensed to chipmakers like Marvell and MediaTek to enable faster, scalable AI systems.
Nvidia announced on Monday that it will begin selling a new chip-to-chip interconnect technology, NVLink Fusion, designed to accelerate the communication between chips essential for building and deploying artificial intelligence systems.
The company unveiled NVLink Fusion, developed over recent years at its Israel R&D center, as the latest version of its NVLink technology. For the first time, Nvidia will license the tech to third-party chipmakers, enabling them to build powerful custom AI systems with multiple interconnected chips. Among the early adopters are Marvell Technology and MediaTek, which plan to use Fusion in their next-generation chip designs.
NVLink, developed over recent years at the Israel R&D center based on the acquisition of Mellanox, which Nvidia bought for $6.9 billion in 2019, has been a foundational component in systems like the GB200, which pairs two Blackwell GPUs with a Grace CPU.
Nvidia CEO Jensen Huang made the announcement at the Taipei Music Center, the venue for Computex 2025, which runs from May 20 to 23. Alongside the product reveal, Huang said Nvidia will build a new headquarters in the northern suburbs of Taipei, further cementing the company’s investment in the region.
Related articles:
In his keynote, Huang recounted Nvidia’s transformation from a video game graphics chip maker into the world’s most dominant AI chip designer, especially since the launch of ChatGPT in 2022 catalyzed explosive demand. “At one point, 90% of our presentations focused on graphics,” he said. “That’s no longer the case.”
He also highlighted Nvidia’s expanding product roadmap, including: The Blackwell Ultra, expected later in 2025. The Rubin series, followed by Feynman processors slated for release in 2028. A new desktop AI system called DGX Spark, aimed at researchers and now in full production with shipments expected within weeks.
Nvidia has also been designing CPUs compatible with Microsoft Windows using Arm Holdings architecture, as previously reported by Reuters.
Computex 2025 will feature more than 1,400 exhibitors and marks the first major gathering of chip and computing executives in Asia since former U.S. President Donald Trump threatened sweeping new tariffs aimed at reshoring tech manufacturing to the United States.
Last year at Computex, Huang inspired a local media frenzy dubbed “Jensanity,” as crowds mobbed the CEO at every appearance, an energy Nvidia is clearly hoping to harness once again.
