TSMC and Nvidia team up in response to the challenge from DeepSeek

As you well know, DeepSeek is an innovative free and open-source Chinese AI model, which has challenged existing notions about hardware requirements in artificial intelligence (AI) training.

DeepSeek DOWNLOAD

Their approach has caused a notable drop in the market value of several U.S. tech companies, traditionally dominated by NVIDIA’s advanced hardware.

Despite DeepSeek claiming to use only 2,048 NVIDIA H800 chips, analysts have questioned this figure, suggesting that the actual number of GPUs involved could be around 50,000, creating uncertainty about the model’s costs.

60,000 wafers per month during 2025 to avoid bottlenecks

According to a report by SemiAnalysis, it is estimated that among these GPUs there are at least 10,000 H100 and another 10,000 H800, in addition to trimmed versions like the H20.

The AI model from DeepSeek, which has 671 billion parameters, allegedly cost 5.6 million dollars, a figure that might not reflect the actual expenditure. This situation has sown skepticism in the industry about the viability and the genuine resources of the Chinese company.

In response to this context of uncertainty, TSMC, the Taiwanese semiconductor giant, has announced a five-year expansion plan aimed at increasing its integrated circuit manufacturing capacity using its advanced CoWoS packaging technology.

This expansion is partly a reaction to the growing demand for NVIDIA’s Blackwell microarchitecture GPUs, which will allow the company to better meet the needs of its customers and strengthen its position in the market against rivals like DeepSeek.

DeepSeek DOWNLOAD

The new TSMC plants, located in Chiayi, Taiwan, will be crucial in this growth. With the capacity to process up to 60,000 wafers per month during 2025, TSMC is actively preparing to avoid future production bottlenecks and ensure its competitiveness in an increasingly challenging market.