Nvidia tried to get a dedicated line for 3D packaging of AI chips from TSMC, but it didn’t work out

For its ability to deliver compute accelerators to customers, Nvidia depends on TSMC not only for silicon wafer processing, but also for chip packaging using its unique CoWoS spatial packaging method. Attempts by the management of the first company to obtain dedicated TSMC production facilities for these needs were unsuccessful, according to rumors.

Image source: NVIDIA

This information is published by the Mirror Media resource, citing details about the program of the visit of Nvidia founder Jensen Huang to Taiwan in June of this year. According to sources, then he had a meeting not only with the retired founder of TSMC, Morris Chang, but also with the current head of the company, C.C. Wei. The head of Nvidia at that moment, according to rumors, asked TSMC to allocate a separate production line for the needs of his company, on which AI chips of this brand would be packaged, but was refused by representatives of the Taiwanese contractor. The outcome of the negotiations created some tension between the companies, sources say, but current chairman CC Wei did his best to smooth over the consequences.

At a recent quarterly conference, TSMC management admitted that the company will not be able to meet demand for the production of components for artificial intelligence systems until at least 2026. At the same time, the Taiwanese manufacturer still cannot find a balance between supply and demand, but is trying to rationally determine the amount of necessary capital costs. Apparently, investments in a production line for packaging Nvidia chips at this stage seem irrational to TSMC management. Moreover, the profit margin in this service sector is close to the company average, without providing any impressive advantages. As noted, TSMC’s refusal was motivated by possible consequences for the company’s relationships with other clients who would also want to achieve certain privileges. By maintaining a level playing field for all customers, TSMC can ensure a more predictable scaling of production capacity.

In the past, as noted by sources familiar with TSMC’s business practices, the company provided certain privileges to large clients. For example, Apple at one time asked to provide it with dedicated chip production lines, and TSMC agreed to this, but at that time the Taiwanese manufacturer was highly dependent on Apple orders and could not neglect this opportunity to optimally load its pipeline. In the case of Nvidia, the situation is noticeably different. It is expected that TSMC will not be able to meet the market demand for chip packaging capacity using the CoWoS method even by the end of next year, as demand will grow at a faster pace. This market position, close to a monopoly, allows TSMC to more firmly defend its interests in negotiations with customers. This is noticeable even in the statements of CC Wei, who recently admitted that he would like to take more money from Nvidia for TSMC services.