Reprinted from Caixin, copyright belongs to the original author. Please contact for removal if there is any infringement.
Following NVIDIA’s announcement at the end of August regarding the mass production of the Rubin architecture chips next year, NVIDIA discussed the progress of Rubin at the Goldman Sachs Technology Conference on September 8.
NVIDIA CFO Colette Kress stated that the Rubin chip is preparing for market entry, and the Rubin architecture will consist of six chips, all of which have entered tape-out.
Colette Kress mentioned that NVIDIA’s data center revenue for the second quarter of fiscal year 2026 includes various components, including the GB200 and B200 from the Blackwell architecture, as well as the GB300 from the Blackwell Ultra architecture. The GB200 and GB300 are still being shipped in the third quarter. Looking ahead, NVIDIA has already seen several gigawatts of demand related to Rubin before its market entry.
Colette Kress noted that discussions a year ago focused on the transition between the Blackwell architecture and others. Now that Blackwell has entered the market, including Blackwell Ultra, NVIDIA has had a smooth year. Another frequently discussed topic is whether there is still a significant need for computation during pre-training, post-training, and inference stages. It can be observed that there remains a huge demand for computation over the past year, a significant portion of which is driven by inference models.
Previously, Colette Kress mentioned during the earnings call following NVIDIA’s report at the end of August that capital expenditure related to data center infrastructure is expected to reach $3 trillion to $4 trillion over the next five years.
At the Goldman Sachs Technology Conference, she explained that discussing the capital expenditure for data center infrastructure over the next five years is to help the entire ecosystem understand the market’s importance. “We are talking about a new computing platform for the next few decades, and we need to start transitioning from a standard computing platform that has existed for two to three decades,” Colette Kress stated. She pointed out the actions of the four leading cloud service providers, noting that AI factories will become a significant part of the global industry, and sovereign AI demand is also crucial. Considering these factors, NVIDIA sees capital expenditure exceeding $3 trillion.
Colette Kress also addressed questions regarding the lifespan and replacement cycle of chips deployed in 2023 during the conference. She stated that NVIDIA has not yet seen significant changes; the previous generation Hopper architecture chips are still performing well. Many of NVIDIA’s customers have a depreciation period of 4 to 6 years, and many will continue to retain these chips in their data centers due to their high performance.
In the second quarter of fiscal year 2026, NVIDIA’s GAAP gross margin was 72.4%, and NVIDIA anticipates a GAAP gross margin of 73.3% for the third quarter. Regarding how to improve gross margins during product transitions, Colette Kress stated that the Blackwell Ultra version is running smoothly, allowing NVIDIA to continue focusing on cycle time, time to market, and lower costs, while also improving gross margins through product mix.
To ensure the supply chain keeps pace with NVIDIA’s iteration speed, Colette Kress responded that some supply chain vendors have spent a lot of time over the past 30 years understanding NVIDIA and improving overall supply. Some vendors need to build different production lines and consider production flexibility. Partnerships are one of the key factors in NVIDIA’s success, and she does not believe many other companies can establish such a supply chain.