Taiwan’s TSMC produces cutting-edge graphics processors for Nvidia, and their mutual dependence has only intensified during the artificial intelligence boom. As The Information notes, at the same time, there has been growing tension and mutual dissatisfaction in the relations of long-time partners, and problems with preparing for mass production of chips for Blackwell computing accelerators are to blame.
Image source: NVIDIA
There were no signs of trouble in March when Nvidia’s Blackwell accelerators were unveiled, but subsequent interactions between the two companies have fueled the controversy as preparations for mass production of this generation of products have not gone according to plan. Nvidia engineers discovered that the first samples of Blackwell chips received from TSMC were not able to operate stably in the conditions typical for server systems. It was believed that errors in the development of accelerators led to the appearance of such defects.
In addition, the customer was confident that the slow scaling of production volumes of Blackwell generation chips was caused by TSMC’s desire to use new chip packaging technology, which in practice turned out to be quite crude. At the same time, TSMC representatives reproached Nvidia for being in a hurry in preparing Blackwell for mass production, which could have caused the corresponding defects. Nvidia allegedly knew about the presence of defects, but continued to rush TSMC and did not dare to eliminate these defects for a long time. Nervousness in the relationship between the two companies began to be transmitted to investors, who expressed their uncertainty in a decrease in interest in buying shares of both.
The relationship between Nvidia and TSMC dates back to 1995, and until recently they have developed quite harmoniously, turning the former into the latter’s second largest client after Apple. According to some estimates, last year Nvidia orders accounted for up to 11% of TSMC’s total revenue, which corresponds to approximately $7.73 billion in monetary terms. In this sense, Apple could claim all 25% of TSMC’s last year’s revenue.
Nvidia previously turned to Samsung for services to produce gaming video cards, but this experience could not always be called positive. However, rumors mention Nvidia’s intentions to once again entrust the production of new gaming GPUs to a South Korean contractor. This will allow Nvidia to be expected to cut chip production costs by 20% or 30% compared to TSMC.
It is generally accepted that TSMC founder Morris Chang instilled the practice of “punishing” disloyal customers with money. In this sense, Nvidia, moving from TSMC to Samsung and back, risks facing higher prices for its chips. This summer, information appeared about Nvidia’s request to build a dedicated line for chip packaging using the CoWoS method, but TSMC representatives did not take it seriously, citing the high level of capital costs. But they looked at the idea of increasing prices for the company’s services for Nvidia with approval.