NVIDIA Rubin GPU and Vera CPU both taped out at TSMC with data center platform launch targeted for 2026.

Nvidia's chief financial officer Collette Cress has announced that the company's next-generation data center-grade GPU codenamed Rubin and CPU codenamed Vera have been taped out and are "in fab," meaning their silicon is currently being produced by TSMC. "The chips of the Rubin platform are in fab," said Collette Kress during the company's earnings conference call with financial analysts and investors. "The Vera CPU, Rubin GPU, CX9 Super NIC, NVLink 144 scale-up switch, Spectrum X scale-out and scale-across switch, and the silicon photonics processor [for co-packages optics]. Rubin remains on schedule for volume production next year." This announcement indicates that Nvidia's next-generation data center platform for AI is on track for introduction in 2026.

Tape-out represents a crucial milestone in semiconductor manufacturing when the final and verified chip design is sent from the design team to a chip producer for manufacturing. At this stage, the physical design place-and-route process is complete, with the chip layout fully optimized for performance, power, area, timings, and having passed all verification checks. The design is then converted into a format containing the exact geometric patterns of transistors and interconnects, which chipmakers like TSMC use to create photomasks for actual chip production. Nvidia tends to simulate its designs using its own supercomputers to ensure that the very first silicon from the fab works and achieves performance and power milestones.

The significance of reaching tape-out lies partly in the extreme costs and risks involved in subsequent stages. Since mask creation is extremely costly—often tens of millions of dollars for advanced nodes—tape-out represents a critical point of commitment. If mistakes are discovered after the mask printing stage, it almost certainly requires a new re-spin and tape-out, adding months of delay and tens of millions of dollars in additional costs. The fact that Nvidia's partners have successfully created photomasks and put the Rubin GPU, Vera CPU, and various scale-up and scale-out switching ASICs into production indicates that the designs have passed critical validation thresholds.

Once Nvidia receives actual chips back from TSMC, the company will begin bring-up and debug processes. Typically, if the first silicon implementation is sound and no re-spins are required, a complex chip may enter production within 9 to 12 months. However, given that the Rubin platform consists of multiple chips that must work together, it will take additional time for Nvidia and its partners to verify that all the processors function in concert as planned.