GTC Berlin Keynote: “GPU demand has really gone through the roof” as agents become more useful and widely adopted; reasoning, iteration and tool use are driving an “enormous” increase in token consumption. Huang described European AI activity as “really accelerating.”

- “GPU demand has really gone through the roof” as agents become more useful and widely adopted; reasoning, iteration and tool use are driving an “enormous” increase in token consumption. Huang described European AI activity as “really accelerating.”
- New factory-integrated DSX Pods target arrival-to-token-production time of roughly one month versus four to five months, reducing idle compute and CapEx. On-site deployment could fall to days once utilities are ready. Huang illustrated the stakes with a $50B, 1GW AI factory potentially generating $40B annually—making a three-month delay equivalent to $10B of deferred customer revenue, not Nvidia revenue guidance.
- Vera Rubin is in full production; its cable-free, fully liquid-cooled design cuts node assembly to under five minutes versus two hours for Grace Blackwell. Huang claimed another 30x performance improvement over Blackwell. Spectrum-X co-packaged-optics switches are also in volume production; DSX power controls target 30% less power for the same output.
- Nvidia invested nearly $10B across European startups and partnerships over the past year, including investments in 14 startups announced during the latest week’s European tour. Its European partnership base now exceeds 12,000, supporting expansion beyond hyperscalers into sovereign clouds, enterprises and industrial applications.
- OpenShell adoption has spread across frontier labs, operating-system providers, cybersecurity and platform companies since its announcement; Nvidia and Anthropic are partnering on Claude OpenShell agents. Hardware-isolated Sentry security runs on BlueField-4, extending Nvidia’s role beyond model inference into enterprise-agent deployment and security.
- Volkswagen’s van, JLR, Volvo and next-generation Mercedes vehicles are joining the Hyperion-compatible ecosystem. Nvidia aims to establish a common robotaxi architecture across automakers, autonomous-driving software providers and ride-hailing platforms, rather than a single-brand fleet.
- Agile 1, Nvidia’s reference humanoid developed with Agile Robotics, is scheduled for Q1 2027 with computing and software preintegrated. Huang expects humanoids to become the highest-volume robotics configuration; the reference platform is intended to remove several years of foundational integration work for developers.
- RTX Spark desktops, laptops and workstations are now available following four years of development with Microsoft, bringing local agent execution and CUDA-accelerated tools to PCs. The N1X-based platform offers 1 petaflop and 128GB memory; DGX Station with Windows is expected by year-end 2026.
- Nvidia completed its first commercial phone call on T-Mobile’s network with Nokia during the latest week, after nearly eight years of development. Huang sees telecom infrastructure evolving into distributed AI networks, describing roughly 100GW of existing distributed capacity as a future opportunity—not committed AI deployment.

**TTN Note: Before today’s Berlin keynote, Huang said Vera Rubin was “now in full production” during its earnings call on Aug 26th.