In a recent revelation by The Information, Nvidia’s highly anticipated Blackwell architecture series AI flagship chip, the GB200, is experiencing a significant setback. The company has announced that due to critical design defects discovered prior to mass production, the shipment of these chips will be delayed by at least three months.
Nvidia unveiled the Blackwell series in March, with CEO Jensen Huang expressing confidence in May about commencing mass shipments later this year. The GB200 chip, featuring two interconnected Blackwell GPUs and a Grace central processing unit, was initially slated for mass production in the third quarter, with deliveries to Nvidia’s major clients—Microsoft, Google, and Meta—scheduled for the fourth quarter.
However, recent inspections by TSMC engineers revealed a design flaw in the chip’s bare die, impacting the connectivity between the two GPUs. This defect poses a risk to chip yield and output, prompting Nvidia to halt production and collaborate with TSMC on design revisions and trial runs.
This type of delay, while not unprecedented, is notably rare at such a critical stage. TSMC had allocated production capacity for the Blackwell series but now faces idle time as the design issues are addressed. As a result, mass production is expected to be postponed until the fourth quarter, with shipments likely to be deferred to the first quarter of the following year.
The delay poses significant disruptions for Nvidia’s major clients. Cloud service providers, including Google, Microsoft, and Meta, typically require around three months to integrate new chips into their computing clusters. Consequently, projects planned for deployment in early 2025 are now at risk.
Google has placed an order for over 400,000 GB200 chips, with costs potentially exceeding $10 billion when including server hardware. Meta’s order is similarly valued at a minimum of $10 billion. Microsoft has recently increased its order by 20%, originally intending to supply Blackwell-based servers to OpenAI by January. This timeline may now shift to March at the earliest.
AI companies, which had anticipated deploying large clusters of Blackwell chips in their data centers to enhance computing power for AI applications, now face uncertainty. The delay could impact the development of advanced models like ChatGPT and other AI innovations.
Adding to Nvidia’s challenges, the delay comes amid an ongoing antitrust investigation by the U.S. Department of Justice. This situation compounds the difficulties for Nvidia, already under scrutiny, and has negatively impacted the company’s stock, which had anticipated a substantial boost in revenue from the Blackwell series.
Nvidia has hinted at the possibility of producing a simplified version of the chip with a single Blackwell GPU to mitigate defects and expedite shipments. However, the company has refrained from commenting further on the specifics of the delay. For now, customers are testing the available samples, and Nvidia maintains that production is still on track for later this year.
Analysts had previously projected that the Blackwell chip would propel Nvidia’s data center revenue from $47.5 billion in 2024 to over $200 billion in 2025. This optimistic forecast now faces significant revisions due to the unforeseen delay.
Related topics: