At GTC Berlin, Nvidia CEO Jensen Huang opened with a familiar pitch: AI infrastructure is scaling faster than sites can assemble it. The answer he offered is DSX Pod, a containerized version of Nvidia's DSX reference architecture built around Vera Rubin NVL72. According to Nvidia, each pod integrates seven liquid-cooled compute racks with 17 kilometers of Spectrum-X cabling and arrives factory-tested in transportable containers. It uses dry cooling rather than traditional evaporative cooling, and 800V DC power delivery.
Huang argued that shifting assembly into the supply chain shortens the time from compute arrival to token production from an estimated four to five months to around one month. He cited a 1GW datacenter costing about $50bn, with a three-month delay potentially costing $10bn in revenue. Nvidia said it is working with partners including Schneider Electric, HPE, Lenovo and Dell to deliver DSX Pod AI factories worldwide in early 2027.
The keynote also included updates to Isaac, Nvidia's robotics platform. The company unveiled Agile One SN, a 1.4-meter humanoid reference robot with five-finger hands for tactile manipulation, multimodal sensing, 2,070 FP4 teraflops of onboard compute and 128GB of unified memory. Huang said it would be available in Q1 2027, pre-assembled with integrated software.
On supply chains, Nvidia announced agreements with Amkor, IBIDEN, KYEC, Samsung Electro-Mechanics, SPIL and Unimicron to expand capacity in substrates, advanced packaging, testing and optical engines. It expects total semiconductor testing capacity to double from calendar 2026 to 2027, and targets a 15x increase in Taiwan-made optical engines by mid-2027 and another doubling by the end of 2028. Huang also used the keynote to argue that AI performance has improved 1 million times over 10 years, compared with roughly 100x under Moore's Law.