GB200 Production on Track, AI Servers at Full Capacity
NVIDIA CEO Jensen Huang recently responded to rumors regarding delays in the delivery of the company's GB200AI servers during an interview with Wired magazine on December 18. Huang strongly dismissed the concerns, stating, "The GB200 is in full production, and everything is progressing smoothly." He further praised the company's latest Blackwell platform, highlighting its 30-fold performance improvement during inference tasks while balancing efficiency and energy savings.
Huang emphasized that more companies are racing to build AI capabilities, noting that a mere three-month delay could significantly impact the competitive landscape. His reassurances were seen as a stabilizing move for key partners involved in the production of the GB200 AI servers, including manufacturers like Foxconn, Quanta, Wistron, and Inventec. These companies had previously indicated a limited shipment of GB200 AI servers by the end of the year, with larger-scale deliveries expected in the first quarter of next year. Huang's comments are expected to strengthen confidence and support the anticipated ramp-up in production.
Quanta, one of the first companies to ship the GB200 AI servers, acknowledged that the validation process for high-frequency, high-performance servers often encounters various issues, from chip yield to installation testing. They assured that these challenges will be addressed as part of a phased supply chain optimization, ultimately ensuring robust mass production.
The Blackwell platform's AI chips are already being delivered, and market attention is now focused on the impending large-scale shipments of the highly anticipated GB200 servers. According to research firm TrendForce, however, there may be delays in the server cabinet deliveries, as key components such as high-speed interfaces and thermal design power (TDP) require further optimization. These adjustments could push the mass shipment timeline to the second or even third quarter of next year, potentially delaying shipments by three to six months.
In the interview, Huang also highlighted the impressive capabilities of the Blackwell platform. He explained that tasks that once took months to process can now be completed much faster, with model training times reduced to one-third or one-fourth of the original duration. What once required six months can now be achieved in about one and a half months.
Regarding the inference capabilities of Blackwell, Huang noted that the platform's approach focuses on long-term thought processes rather than zero-shot or one-shot learning. Essentially, AI generates multiple potential solutions in its "mind" and, using additional computational power, refines those solutions to provide the most suitable answers. This approach, known as "test time scaling" or "inference time scaling," allows Blackwell servers to deliver both high efficiency and energy conservation during the inference process.
