Nvidia GB200 chip has defects, delivery delayed
According to the US technology website The Information, due to design defects, the shipment time of Nvidia's next-generation Blackwell architecture series AI flagship chip GB200 will be delayed by at least three months or more, and mass shipments may be delayed until Q1 next year.
The Blackwell chip was originally scheduled to start mass production in October 2024. If it is postponed to April 2025 due to delays, it will directly affect Nvidia's quarterly earnings. At present, there are three types of chips known for the Blackwell GPU architecture: B100, B200 and GB200 super chips.
The GB200 chip contains two connected Blackwell GPUs and a Grace central processing unit. However, in recent weeks, when TSMC engineers were preparing for mass production, they found a design defect in the bare crystal connecting the two Blackwell GPUs. This defect will lead to a decrease in chip yield or output, and the usual practice is to stop mass production. Therefore, Nvidia had to adjust the chip design and cooperate with TSMC for new trial production before starting mass production.
TSMC originally planned to start mass production of the Blackwell series chips in the third quarter and to start mass shipments to Nvidia customers from the fourth quarter. However, due to the discovery of design defects, mass production had to be postponed to the fourth quarter, and mass shipments are expected to be postponed to the first quarter of next year. TSMC has reserved production capacity for mass production of GB200, but has to idle the production line until the problem is resolved.
In response, Nvidia responded to the media that the strong demand for Hopper chips and the production plan for Blackwell chips have not changed: "As we said before, the demand for Hopper is very strong, Blackwell sample trials have begun widely, and production is expected to increase in the second half of the year. Beyond that, we do not comment on rumors."
For Nvidia's customers, chip delays may increase their concerns about the timing of delivery. It is reported that technology giants including Google's parent company Alphabet, Amazon, Microsoft and Facebook's parent company Meta spent a total of nearly US$60 billion in capital expenditures in the second quarter, a year-on-year increase of two-thirds, a large part of which went to Nvidia.
The Information also reported that Meta has now placed orders worth at least $10 billion, while Microsoft has increased its order size by 20% in recent weeks. They originally planned to provide OpenAI with servers based on the Blackwell series of chips by January next year, but now it may be delayed until at least March next year.
In addition, the design flaw will also affect the production and delivery of Nvidia's NVLink server racks, because companies working on servers must wait for new chip samples before finalizing the server rack design.
In addition to product delays, the United States may launch an antitrust investigation into Nvidia, mainly involving a merger and acquisition case and Nvidia's business practices. The two investigations are independent but related, and both generally focus on the possibility of Nvidia building a moat around its GPUs.
The investigation is reportedly mainly due to complaints from Nvidia's competitors that Nvidia may abuse its market dominance when selling artificial intelligence (AI) chips. The US Department of Justice is reportedly investigating whether Nvidia forces cloud computing providers to buy multiple Nvidia products. The investigation also involves whether Nvidia will increase the price of network equipment to these customers once customers want to buy AI chips from Nvidia's peers such as AMD.
