BNP Paribas在GTC Berlin后维持英伟达跑赢大盘及345美元目标价，多为分析师对已发布路线图的复述

Nvidia ( NVDA ) held its GTC Berlin event earlier this week, and investment firm BNP Paribas walked away from the conference impressed at the size and speed at which the company continues to operate. “We walked away from GTC Berlin with a few salient takeaways: 1) Nvidia continues to optimize at the AI factory level to expedite GW-scale deployments; 2) Nvidia’s CPO push for scale-out and scale-up networks remains on track for 2H of CY26 and 2H of CY27, respectively; and 3) Nvidia is investing heavily in the AI supply chain as data center deployments proliferate across Europe and Asia,” BNP Paribas analyst Karl Ackerman wrote in a note to clients. “ We remain bullish on Nvidia as the undisputed AI infrastructure leader with expansion into turnkey AI factory solutions, secure autonomous agents, and sovereign AI buildouts.” Ackerman has an Outperform rating and a $345 price target on Nvidia. At the event, Nvidia showed off several upcoming products, such as its Vera Rubin Ultra NVL576, Vera CMX, Groq 3 LPX, and Spectrum-X Photonics switches. In addition, Ackerman said he was able to see the NVL72 rack in person, with 8 NVL72 racks connected via optics to form a 576-GPU scale up. “We think each NVL72 system could have 648 3.2T optical engines, which translates to 9 optical engines per GPU or per switch ASIC,” Ackerman added. “Assuming one external laser source module provides light to every two optical engines, this implies 324 ELS modules per NVL72, or 2,592 ultra-high power CW lasers per NVL72 system.” Another notable takeaway from the event was the announcement of DSX PODs, which is a Vera Rubin AI factory reference architecture that could cut the time it takes for GPUs to be installed, moving to 14 days from six months. So far, Nvidia is working with 17 partners to deploy DSX PODs by the first quarter of 2027. Other notable announcements include the introduction of the OpenShell reference architecture; plans for an increase in substrate capacity by two and a half times over the next three years, with IBIDEN, Samsung ( SSNLF ), and Unimicron; the deal with Amkor ( AMKR ); and installing Google ( GOOG ) ( GOOGL ) A5X instances based on Vera Rubin NVL72 in Europe. “Google plans to deploy up to 960k Rubin GPUs, which will leverage Nvidia’s CX-9 SuperNICs and connect to TPUv8t superpods via Google’s Virgo Network,” Ackerman added. “We think 960k GPUs implies 13.3k NVL72 racks, or ~2.9GW of AI compute capacity (not including scale-out networking).”