NVIDIA Confirms Rubin AI Architecture in Production at TSMC

NVIDIA’s next leap in computing has officially begun. During his visit to Taiwan, CEO Jensen Huang confirmed that Rubin, the company’s most advanced AI architecture to date, is now in the works at TSMC.

Just months after launching its Blackwell Ultra GB300 AI servers, NVIDIA is already looking ahead to an even bigger generational shift.

According to Huang, six Rubin chips have already taped out and entered trial production at TSMC’s fabs. These include brand-new CPUs, GPUs, an updated NVLink switch, and even a silicon photonics processor – signaling that NVIDIA intends to revamp its entire tech stack. The company is also introducing its first chiplet-based design, paired with a 4x reticle size upgrade over Blackwell, setting the stage for dramatic performance gains.

Rubin will lean on next-generation HBM4 memory for its R100 GPUs, replacing today’s HBM3E, and will be built on TSMC’s cutting-edge 3nm (N3P) process with CoWoS-L packaging. This ground-up redesign suggests Rubin could be as transformative for the industry as the Hopper generation was just a few years ago.

While Rubin is not expected to hit the market until 2026 or 2027, anticipation is already high. NVIDIA’s aggressive product cadence has left rivals scrambling to keep up, and Huang’s comments in Taiwan reinforce the idea that Rubin represents more than incremental progress – it’s a complete rethink of compute architecture. Whether you’re watching from the AI, data center, or gaming side of the industry, Rubin looks poised to redefine expectations once again.