NVIDIA Rubin 芯片：推动数据中心革命

$NVIDIA(NVDA.US) NVIDIA's next-generation Rubin chips have completed tape-out, and data center capital expenditure is expected to exceed SGD 3 trillion over the next 5 years.

During the recent Goldman Sachs technology conference, NVIDIA shared the latest progress on its Rubin architecture chips. CFO Colette Kress stated that preparations for the market launch of Rubin chips are underway, and all six chips under the Rubin architecture have already completed the tape-out stage.

According to Colette, NVIDIA’s data center revenue for the second quarter of fiscal year 2026 will include a variety of chip combinations. These include the GB200 and B200 chips based on the Blackwell architecture, as well as the GB300 chips built on the Blackwell Ultra architecture. In the third quarter, these chips will remain in shipping phase. As Rubin enters the market, the company estimates demand linked to Rubin-powered chips will reach several gigawatts of scale.

Colette also reviewed product developments in recent years. A year ago, discussions primarily focused on the Blackwell architecture and the transition between older and new architectures. However, Blackwell and Blackwell Ultra versions are now successfully launched. She noted that AI computing demand has rapidly expanded over the past year, with inference models driving significant growth in computation requirements.

During a previous earnings call, Colette revealed that global data center infrastructure capital expenditure is predicted to reach SGD 3-4 trillion over the next five years. At the Goldman Sachs conference, she further elaborated on this, emphasizing that this investment involves building entirely new computational platforms for the coming decades. This transformation will drive a generational shift away from conventional computing platforms developed over the last 20-30 years, toward next-generation systems. She highlighted initiatives by the world’s four largest cloud service providers, alongside the growing demand for AI factories and sovereignty-grade AI systems, as major catalysts accelerating this shift. Overall, NVIDIA projects that capital expenditures in the data center sector will surpass SGD 3 trillion.

Regarding chip lifecycles and replacement timelines, Colette stated that NVIDIA has yet to observe significant changes. Chips from the previous Hopper architecture generation continue to perform well, with many clients operating on a depreciation cycle of 4-6 years. These chips are expected to remain integral to data centers, meeting high-performance requirements in the near future.

As for NVIDIA's financial performance, GAAP gross margin for the second quarter of fiscal year 2026 reached 72.4%, with projections for the third quarter increasing to 73.3%. Colette highlighted that the stable performance of Blackwell Ultra has allowed the company to optimize product launch timelines and cost structures while offering an improved product mix, ultimately enhancing gross margins.

Finally, addressing supply chain coordination, Colette underscored that NVIDIA’s success rests heavily on its longstanding partnerships with suppliers. Some vendors have deepened their understanding of NVIDIA's needs and expanded production capacity over the past 30 years, while others have established new manufacturing lines to meet increasing demand. She emphasized that such in-depth collaborations have allowed NVIDIA to maintain steady progress despite rapid iteration cycles — a supply chain ecosystem that is rare in the industry and a key differentiator for NVIDIA’s competitiveness.