Berlin Pledge: Fortnight Builds for Modular Intelligence Campuses

The gathering in Berlin closed a week-long tour through Paris, London, Stockholm and Dublin with a single message: Europe should build intelligence, not merely rent it. NVDA says it placed nearly $10B across the continent over the past year, backed 14 young firms in a single week and now works with more than 12,000 regional partners. Nvidia (blogs.nvidia.com) carries the same figures in its live keynote record, so the scale claim holds up. The pitch frames models and tokens as goods Europe must produce rather than import.

The keynote stacks the work into five strata: energy feeds advanced systems, systems form factories, factories run models, models mint tokens, and tokens enter every application. NVDA claims a million-fold jump in training and inference efficiency across a decade, against roughly a hundred-fold from classical scaling laws. Blackwell runs agentic LLM work about 30 times faster than Hopper, the talk says, while Vera Rubin promises another 30-fold step over Blackwell.

A production Vera Rubin unit appeared on stage without visible cabling and fully liquid-cooled; one node assembles in under five minutes, down from two hours on the Grace Blackwell generation. Seventy-two components act as a single giant processor inside the NVL72 rack, joined by a new central processor and photonics switching. The DSX POD packs seven racks and 17 km of cabling into factory-tested containers, cutting site work from six months toward 14 days; CounterpointResearch notes 17 partners preparing deliveries for early 2027. At roughly $50B per gigawatt campus, each idle month burns serious revenue, which explains the urgency for teams serving agentic models and token demand.

Agents get five parts in this telling: models that mint tokens, a harness that organizes memory and tools, an isolated sandbox, enterprise knowledge and outside tools. The OpenShell runtime enforces policy at the level of files and network calls, keeps credentials outside the sandbox and checks every privilege change against a ceiling. A monitor on BlueField-4 watches agent reasoning outside the host and severs the link within milliseconds when intent drifts. Frontier labs, platform vendors and security firms have endorsed the design, the talk says.

The boldest forecast holds that every software vendor becomes an agent vendor, pairing each tool with specialized agents tuned for that tool. Models stop being standalone products and turn into a layer inside all applications. Subscription pricing could then give way to outcome-based fees for agents that complete tasks. Whether buyers accept that shift matters more than any demo.

Physical intelligence arrives through the Hyperion driving platform, now joined by four more marques: the VW van, a new JLR vehicle, the Volvo fleet and the next Mercedes generation. With the world driving about 1T miles a year, management sketches robotaxis near $1-2 per mile. Electrive (electrive.com) reports Bolt will gather data in Madrid during 2027 with Lucid vehicles, targeting at least 25,000 units across Europe, while Mercedes seeks German approval for city point-to-point assistance built on the same agent and model stack.

The finale shrinks the story to palm size: a compact system built with Microsoft over four years around the N1X part, pairing 1 PFLOP with 128 GB of memory. The same capacity once filled a 2016 cabinet that two people carried upstairs for a small lab. Europe leaves with a call to join the producing side of models and tokens with its 12,000-strong network, and the hall answered with applause.