TSMC currently cannot keep up with the demand for Nvidia’s AI chips. The production backlog will last at least 18 months, which could lead to a shortage of stocks.

CEO Mark Liu sounds the alarm about Nikkei Asia. TSMC currently has to make every effort to meet eighty percent of the demand for Nvidia GPUs, of which it is the exclusive manufacturer. In the coming months, there will be shortages of both the most powerful H100 and the slightly less advanced A100 GPUs, both of which are currently in high demand for running AI models.

Packaging

The situation may remind you of the chip shortage during the coronavirus pandemic. The problem, however, is not the number of GPUs TSMC can roll off the assembly line, but rather the insufficient capacity for packaging chips. To be clear, this doesn’t mean TSMC is short on boxes. In this context, packaging refers to the linking of all components in the chips.

This is a very important step in the production process, especially for powerful chips like the Nvidia H100 GPUs. A chip contains millions to billions of tiny transistors that transmit power and data. The more transistors a chip has, the more powerful its computing power is, is the simplified rule of thumb. TSMC’s process for packaging chips is called CoWoS.

Eighteen months

Liu expects his company will need a maximum of 18 months to reduce the production backlog. TSMC has already pumped $3 billion into a new facility that will focus on packaging technology. The company also wants to improve its CoWoS technology to package chips more efficiently so that more transistors are not necessarily needed to increase the processing power of chips in the future.

Analysts suspect there is more afoot. The demand for AI hardware will not immediately fall, the hype around generative AI is growing towards its peak. If production cannot keep up now, the backlog is likely to catch up. Eighteen months can quickly turn into several years, as we learned from the last chip crisis.

In addition, TSMC is one of the few on the market that can develop chips powerful enough to support large AI models such as GPT-4 on behalf of Nvidia. Of course, there are also AMD and Intel building GPUs, as well as new players like Amazon and Google that want to build their own chips, although they won’t cause an immediate shift in the market.

Nvidia has not yet responded to TSMC’s statements. The H100 is a generous cash cow for the company: it makes tens of thousands of euros in profit with every GPU sold. As a result, the company generates billions in profits. Nvidia certainly won’t want to return to last year’s situation.