NVIDIA could face delays in the release of its latest AI chips due to design issues, according to anonymous sources involved in chip and server hardware manufacturing cited by The Information. The delay could last up to three months or more, potentially affecting major customers like Meta, Google, and Microsoft. An unnamed Microsoft employee and another source claim that NVIDIA has already informed Microsoft of delays affecting the most advanced models in its Blackwell series of AI chips. As a result, significant shipments are not expected until the first quarter of 2025.
When contacted for comment, an NVIDIA representative would not respond to customer messages regarding the delay, but said that “production is on schedule” for later this year. The Information reports that Microsoft, Google, Amazon Web Services, and Meta declined to comment on the matter, while Taiwan Semiconductor Manufacturing Company (TSMC) did not respond to inquiries.
The manufacturing issue was discovered by TSMC and affects the processor die that connects the two Blackwell GPUs on the GB200.” — Data Center Dynamics
NVIDIA needs to redesign its chip, which requires a new manufacturing test at TSMC before mass production. Rumor has it they are considering a single-GPU version to speed up delivery. The delay is temporarily shutting down TSMC’s production lines.
SemiAnalysis’ Dylan Patel reports that Blackwell shipments will be significantly lower than expected in Q4 2024 and the first half of 2025. The shortfall is due to TSMC’s transition from CoWoS-S to CoWoS-L, which is required for NVIDIA’s advanced Blackwell chips. TSMC’s AP3 fab is currently dedicated to CoWoS-S production, while initial CoWoS-L capacity is being installed at AP6.
Additionally, NVIDIA appears to be prioritizing production of GB200 NVL72 units over NVL36 units. According to DataCenterDynamics, NVIDIA plans to ship more than 60,000 GB200 NVL36 units. Each of these supercomputers contains 36 GB200 superchips paired with 18 Grace processors and costs $1.8 million per unit. The larger GB200 NVL72 unit is said to cost $3 million. In contrast, the NVL72 design includes 72 GPUs, either in a single rack with 18 dual GB200 compute nodes or distributed across two racks, each containing 18 individual nodes. NVIDIA also offers standalone GPUs like the B100, which can cost between $30,000 and $35,000 per unit, and the GB200 superchips, which cost between $60,000 and $70,000 per unit.
