TSMC Remains Exclusive for NVIDIA's H100 Chips, Intel's Foundry Progress Limited

On June 1st, NVIDIA CEO Huang Renxun, during his keynote speech at the Taipei Computer Exhibition, expressed the company's commitment to achieving supply chain diversification. He also stated that NVIDIA is open to future collaborations with Intel for the integration of artificial intelligence chips. As a result, there is speculation that Intel may become a foundry for NVIDIA's high-end chips like the H100.

Huang Renxun clarified that the H100 is exclusively manufactured by TSMC and there are no plans to add a second wafer foundry. The main reason behind this decision is the complexity of the design and manufacturing process. Huang Renxun explained, "It is already difficult to accomplish it once; it would be even more challenging to split the production between two foundries."

NVIDIA and TSMC have already initiated their collaborative plans for 3/2nm processes. Huang Renxun confirmed for the first time that the next generation of AI chips will be ordered from TSMC. At least in the AI GPU series, TSMC will continue to be the sole supplier for several more years.

It's worth mentioning that apart from ordering TSMC to manufacture the A100/A800 and H100 chips, NVIDIA has also transitioned the RTX 40 series back to TSMC's 4nm process as their primary partnership with Samsung focuses on memory production.

NVIDIA's H100 GPU adopts TSMC's 4nm process, while the A100 uses a 7nm process. The top-tier H100 Tensor Core GPU, based on the Hopper architecture and manufactured on TSMC's 4nm process, features the fourth-generation Tensor Cores and Transformer engines. It is equipped with 80GB of VRAM, providing a 2x increase in performance per kilowatt compared to the A100 GPU.

Currently, TSMC does not have any significant competitors in sub-7nm processes for AI chip manufacturing. Although Intel has reentered the foundry market, they have not made any meaningful progress in this area yet.