The Zhitong Finance App learned that Nvidia (NVDA.US) held the Berlin GTC conference earlier this week. After attending the conference, BNP Paribas was impressed by the scale and speed of the company's continued operations.
BNP Paribas analyst Carl Ackman wrote in his report to customers: “We have taken a few key points from the GTC conference in Berlin: 1) Nvidia continues to optimize at the AI factory level to accelerate gigawatt-level deployment; 2) Nvidia's CPO (Co-packaged Optics) promotion plans for horizontally expanded and vertically expanded networks are still being implemented in the second half of 2026 and the second half of 2027, respectively; 3) As data center deployments spread in Europe and Asia, Nvidia is investing heavily in the AI supply chain. We still look at Nvidia as the undisputed leader in AI infrastructure, and are expanding towards turnkey AI factory solutions, secure autonomous agents, and sovereign AI construction.”
Ackman gave Nvidia an “outperforming the market” rating, with a target price of $345.
At the conference, Nvidia showcased a number of upcoming products, including the Vera Rubin Ultra NVL576, Vera CMX, Groq 3 LPx, and Spectrum-X Photonics switches.
Additionally, Ackman said he saw the NVL72 rack with his own eyes, and the eight NVL72 racks formed a 576-GPU vertical expansion cluster through optical connections. “We think each NVL72 system may be equipped with 648 3.2T light engines, which is equivalent to 9 light engines per GPU or switch ASIC,” Ackman added. “Assuming an external laser source module provides light for each of the two light engines, this means that each NVL72 is equipped with 324 ELS modules, that is, each NVL72 system is equipped with 2592 ultra-high power continuous wave lasers.”
Another notable launch is DSX Pods, a Vera Rubin AI factory reference architecture that reduces GPU installation time from six months to 14 days. Nvidia is currently working with 17 partners and plans to deploy DSX Pods by the first quarter of 2027.
Other important announcements include the launch of the OpenShell reference architecture; plans to increase board production capacity 2.5 times over the next three years, with partners including Ibi Electric, Samsung, and Xinxing Electronics; reaching an agreement with Anqian; and installing a Google A5X instance based on Vera Rubin NVL72 in Europe.
“Google plans to deploy up to 960,000 Rubin GPUs, which will utilize the Nvidia CX-9 SuperNIC and connect to the TPUV8T supercluster via the Google Virgo network,” Ackman added. “We think 960,000 GPUs mean 13,300 NVL72 racks, or about 2.9 gigawatts of AI computing power capacity (excluding horizontally extended networks).”