Nvidia to Accelerate AI Chip Networking

 Nvidia launched an updated iteration of its NVLink technology, dubbed NVLink Fusion, and will license it to external chipmakers for constructing custom multi‑chip AI systems.

Marvell Technology and MediaTek have already committed to integrating NVLink Fusion into their bespoke chip designs.

Originally developed years ago, NVLink enables the transfer of vast data volumes between chips, exemplified by Nvidia’s GB200 module that pairs two Blackwell GPUs with a Grace CPU.

Chief Executive Jensen Huang announced the licensing plan at the Taipei Music Center during the Computex AI exhibition, running from May 20 to 23.

Huang also revealed Nvidia’s intention to establish a Taiwan regional headquarters in Taipei’s northern suburbs.

In his keynote, Huang recounted Nvidia’s evolution from a graphics‑chip specialist to a leading supplier of AI processors and supportive software.

“At one point, my presentations spent 90% of the time on graphics chips,” Huang noted, “but that has changed.”

Since ChatGPT’s 2022 debut, Nvidia’s processors have underpinned a surge of AI innovation across the technology sector.

Reuters has reported that Nvidia is developing CPUs compatible with Microsoft Windows and incorporating Arm Holdings technology.

Last year at Computex, Huang drew crowds in Taiwan in what the local press dubbed “Jensanity.”

At March’s annual developer conference, he detailed Nvidia’s strategy to shift focus from training large AI models to deploying AI‑driven applications.

He unveiled upcoming AI chip lines, including the Blackwell Ultra arriving later this year, followed by the Rubin series and Feynman processors due in 2028.

Nvidia also introduced DGX Spark, a desktop AI system for researchers, which Huang said is now in full production and “will be ready in a few weeks.”

With 1,400 exhibitors expected, Computex marks the first large‑scale Asia tech gathering since President Donald Trump threatened tariffs to incentivize US manufacturing.