Days after rumors of a strained relationship between Nvidia and TSMC regarding the delayed shipment of Blackwell chips, Nvidia CEO Jensen Huang clarified that the issue was "100% Nvidia's fault." This statement underscores the company's commitment to maintaining a strong relationship with its largest supplier.
According to Reuters, "We encountered a design flaw in Blackwell," Huang stated during a press conference on October 23. "While it was functional, this flaw led to low yield rates, and it was entirely Nvidia's responsibility."
He emphasized, "To make a Blackwell computer operational, we had to design seven different types of chips from the ground up and ramp them into production simultaneously."
"What TSMC did was assist us in overcoming those yield challenges and enabled us to resume Blackwell manufacturing at an impressive pace."
Huang's clarification came in response to a report from The Information on October 16, revealing that Nvidia engineers discovered the Blackwell chip could not operate properly in high-pressure data center environments during testing conducted weeks after its release. Some engineers attributed the issue to a design flaw in Blackwell, while others suggested that TSMC's recent adoption of CoWoS-L packaging technology—which integrates different types of chips—may have impacted production timelines. According to the report, two TSMC engineers indicated that Nvidia's eagerness for expedited production resulted in significantly less time allocated than what was provided for Apple.
This flaw contributed to the delays of the Blackwell chip, which were expected to last several months. According to The Information, these delays underscore Nvidia's pressure on TSMC amid rising AI investments. Nvidia aims to release a new chip architecture annually instead of the previous two-year cycle to sustain its dominance in AI chips, while TSMC struggles to scale its CoWoS capacity quickly enough to meet Nvidia's growing demands.
Huang's acknowledgment of the chip design flaw being Nvidia's responsibility indicates that the company does not have a strained relationship with its largest supplier. As of May 2023, TSMC accounted for 35.82% of Nvidia's costs, according to Bloomberg estimates—more than double that of Nvidia's second-largest supplier, Wistron, and significantly higher than the third-largest supplier, SK Hynix.
It's worth noting that TSMC has been making chips for Nvidia since 1995.
During the latest earnings call at the end of August, Nvidia said it had successfully addressed the mask change for its Blackwell GPU, ensuring no functional impact. The company is now sampling functional Blackwell samples in various system configurations and expects to start production in the fourth quarter of fiscal year 2024. The mask change did not require any significant modifications to Blackwell's design, and the company is confident in its ability to meet production timelines.
Nvidia added that the company anticipates shipping several billion dollars in Blackwell revenue. The production ramp for Blackwell is set to commence in the fourth quarter of fiscal year 2024 and will extend into fiscal year 2026. Demand for Blackwell platforms significantly exceeds supply, and this trend is expected to persist into next year.
Meanwhile, Foxconn has announced that GB200 shipments are anticipated in the fourth quarter. According to TF Securities, which conducted a survey of the two largest GB200 assembly manufacturers—Foxconn and Quanta—Microsoft's demand for the GB200 surpasses the total orders from other cloud service providers. Notably, Microsoft's orders for the GB200 in the fourth quarter of 2024 have surged significantly, increasing by 3 to 4 times.
Nvidia's supplier cost (%) | |
Company Name | Cost of Nvidia |
TSMC | 35.82 |
Wistron | 17.69 |
SK Hynix | 9.40 |
Fabrinet | 4.37 |
Source: Bloomberg, October 2024
