Behind the delayed shipment of Blackwell: NVIDIA and Taiwan Semiconductor caught in "internal strife", one side blames packaging technology, the other side claims to be forced to rush work

When NVIDIA released the Blackwell architecture chip in March this year, they probably didn't expect so many twists and turns in producing chips with this most powerful artificial intelligence (AI) chip architecture. The latest revelations show that it has even caused a rift between NVIDIA and its key foundry partner, Taiwan Semiconductor Manufacturing Company (TSMC).

On Wednesday, October 16th, Eastern Time, according to the technology media The Information citing sources familiar with the matter, shortly after NVIDIA released Blackwell, they and TSMC began to blame each other. Just a few weeks after the release, NVIDIA engineers found during testing that the chips based on the Blackwell architecture would not function in the high-voltage environments common in data centers. Some NVIDIA engineers believe that this may be partly due to design flaws in Blackwell.

Some NVIDIA engineers also believe that the slowdown in production of Blackwell chips is due to TSMC adopting a new technology that packages different types of chips together. This explanation quickly began circulating among financial analysts following TSMC and NVIDIA.

In addition to the design of Blackwell, two TSMC employees stated that another factor contributing to the issues on NVIDIA's side is that NVIDIA rushed TSMC to complete the production process hastily, giving TSMC less time to solve problems compared to another major customer, Apple.

In early August this year, news of delays in the shipment of Blackwell chips shocked the market. At that time, reports indicated that due to design flaws, the most advanced AI chip in the Blackwell series would be delayed for three months or more, and the bulk shipment of Blackwell chips might be delayed until the first quarter of next year. Morgan Stanley also stated in a report that the production of Blackwell chips might pause for about two weeks, but could catch up in the fourth quarter of this year through TSMC's efforts.

Following the news of the shipment delays mentioned above, some disgruntled shareholders visited TSMC, and TSMC's investor relations department shifted the blame to NVIDIA.

The business relationship between NVIDIA and TSMC began in 1995, and the two parties can be said to have known each other from the very beginning. In that year, NVIDIA's founder and CEO Jensen Huang encountered a commercial bottleneck and wrote to his senior at Stanford University, TSMC's founder Zhang Zhongmou, for help. Shortly after, Zhang Zhongmou personally called Jensen Huang back. TSMC then took over and excellently completed the order to manufacture NVIDIA chips, helping NVIDIA quickly capture the market. Jensen Huang even turned this experience of contacting Zhang Zhongmou into a comic and gave it to Zhang Zhongmou as a gift.

Since then, NVIDIA and TSMC have had a long-term cooperation, forming a highly symbiotic relationship. NVIDIA relies on TSMC's advanced technology and expertise to bring chip designs to life, and has become TSMC's second largest customer after Apple. Analyst Dan Nystedt, who researches the semiconductor industry, estimated earlier this year that TSMC had $7.73 billion in revenue from NVIDIA last year, accounting for 11% of its annual revenue, while Apple, the top contributor, accounted for 25% of the revenue Strictly speaking, in the past nearly thirty years, the cooperation between NVIDIA and Taiwan Semiconductor Manufacturing Company (TSMC) has had its ups and downs. The news that broke out this Wednesday may not be too serious, ultimately just a footnote in the decades-long relationship between the two companies. Currently, TSMC occupies over 90% of the market for advanced chip manufacturing, and NVIDIA's launch of cutting-edge chips still relies on TSMC. Although there are signs that NVIDIA hopes to reduce its dependence on TSMC. For example, there are reports that NVIDIA is exploring a collaboration with Samsung to produce new gaming chips, hoping to receive a 20% to 30% discount on the same-generation chip manufacturing technology compared to TSMC.

The report this Wednesday also mentioned another factor that maintains the cooperation between NVIDIA and TSMC, which is Morris Chang. Although he retired in 2018, he still holds significant influence at TSMC. He places great importance on long-term customer loyalty. When some customers transferred part of their business originally given to TSMC to other manufacturers and then tried to return to TSMC, Morris Chang raised the prices for them.

Regardless of whether it affects the cooperation between the two companies, the news on Wednesday clearly indicates that as the AI business attracts more and more funding, the pressure NVIDIA puts on TSMC is increasing. In order to solidify its dominant position in the AI chip field, NVIDIA is committed to introducing a new architecture every year. TSMC has been working hard to expand the capacity of its latest chip packaging technology, but in this regard, TSMC's own efforts are limited.

Over three months ago, a report from a Taiwanese media outlet, Mirror Weekly, also exposed the contradictions in the development of the two companies under the AI boom. The media outlet stated that during a visit to Taiwan in June this year, Jensen Huang proposed to TSMC that he hoped to set up a CoWoS production line exclusively for NVIDIA outside the factory. A TSMC executive retorted on the spot, "Will NVIDIA pay for it? Does TSMC need to set up a wafer production line specifically for NVIDIA outside the factory?" The scene was very awkward, and it was only when TSMC Chairman Mark Liu stepped in to defuse the tension