Nvidia Blackwell GPU Delays & Liquid Cooling Shortages

On August 13, according to a report by Economic Daily, Nvidia's Blackwell GPU, aimed at servers, is facing shipment delays due to design defects. At the same time, as GB200 servers transition from air cooling to liquid cooling, a critical component-the quick disconnect coupling (UQD)-is also experiencing a severe shortage, impacting the delivery schedules of Taiwanese AI server suppliers such as Wistron and Foxconn.

▲ Liquid-cooled Servers

The report states that manufacturers of quick disconnect couplings often have patent protection, and the production to supply process involves repeated validation, including OCP certification and client certification. Additionally, the original suppliers are reluctant to expand production, which lengthens the overall supply cycle when demand surges for such products.

It is worth noting that a recent survey report from Bank of America highlighted that the initial delay in Nvidia's Blackwell GPU is considered normal and reasonable, with mass shipments expected by early 2025. However, the entire industry is still grappling with severe supply shortages. This could extend the lifecycle of Nvidia's H100/H200 series products, which would benefit Taiwanese AI server manufacturers like Quanta, Hon Hai (Foxconn), Delta, and Wistron.

Bank of America provided two key conclusions after visiting relevant supply chains:

Firstly, due to the extensive R&D and testing required for the Nvidia Blackwell platform, initial delays are normal and reasonable. Taiwanese hardware supply chains confirmed that small-scale shipments will begin in Q4 of this year, with mass production and shipments expected in 2025.

Secondly, the AI industry is still facing serious supply shortages, and demand remains strong. Short-term supply issues are affecting shipments, but not demand. Over the past few weeks, U.S. hyperscale cloud service providers have increased capital expenditure. For example, Microsoft raised its projected capital expenditure for next year from $54 billion to $58 billion, emphasizing its continued investment in AI.

Bank of America believes that if Blackwell's shipments are delayed, Nvidia may extend the lifecycle of the Hopper series GPUs (H100, H200), which would benefit the related Taiwanese supply chain. For instance, Wistron holds a 95% market share in H100 and H200 server baseboards and approximately a 50% share in GB200 computing boards.

For hardware manufacturers, Bank of America sees cooling as the biggest challenge. GB200 is Nvidia's first chip to adopt a liquid cooling system, so the supply chain may need more time to improve the design and enhance performance. Despite reports of severe shortages of critical components like the quick disconnect coupling (UQD) required for the liquid cooling system, considering the high average price of GB200 servers ($2-4 million), the supply chain is expected to have sufficient funding and resources to address these issues.