NVIDIA's $10B Berlin bet: sovereign open models, 14-day AI factories, and a talking Jensen

Europe, don't just consume -- produce

The clearest line of Jensen Huang's GTC Berlin keynote, delivered Oct. 21, was an industrial-policy pitch: Europe should not merely consume AI, it should produce it. NVIDIA is backing the pitch with a commitment of nearly $10 billion to the European technology ecosystem, working with more than 12,000 partners. The company listed European customers including L'Oreal, Revolut, Spotify, Merck, Vodafone, Zara, Porsche, Nestle, and HSBC as already building on its stack -- a lineup designed to make sovereign AI look like a procurement decision, not a research project. m:…8373

The operational lever is the DSX POD reference design for Vera Rubin AI factories: modular containers that partners build, integrate, and validate before delivery, shrinking onsite deployment from six months to 14 days once site utilities are ready. NVIDIA says the DSX AI Factory Platform can also cut water consumption for cooling by up to 100% and generate the same AI output with up to 30% less power. For executives who lived through multi-quarter data-center builds, the 14-day figure is the number that will travel back to the boardroom. m:…8373

The open-model bundle

Alongside the infrastructure, NVIDIA expanded its open-model portfolio: Nemotron 3.5 Super VL and Bolt for language, the Kumo family of open-weight models for structured business data, Cosmos 3.5 world models, Alpamayo 2 Edge for autonomous vehicles, BioNeMo RFD4-Proteina for protein design -- which the company says generates designs about 13 times faster than RFdiffusion3 -- and Ising for quantum computing. The stated design principle is sovereignty-compatible: organizations can customize the models with local data and deploy them on infrastructure they control. m:…8087

The ecosystem packaging matters as much as the models. Nemotron models are built with data and research from the NVIDIA Nemotron Coalition, including European partners Black Forest Labs, H-Company, and Mistral, and enterprises including Dassault Systemes, SAP, and Shopify are already working with them. On agents specifically, Huang introduced the OpenShell Reference Architecture -- an open-source, policy-controlled runtime with kernel-level sandboxing -- with the message that the most important thing is making sure the agent is secure. Open models, open agent runtimes, containerized factories: it is a full stack for anyone who wants an AI economy without a foreign dependency. m:…8373 m:…8087

Japan runs the same play

Japan is not waiting for the European version. At Sakana AI's second annual engineer open house on Oct. 5, the company detailed its three-layer bid for Japan-based AI sovereignty: pre-training Japan-specific agent-native LLMs from scratch, post-training with Namazu to adapt overseas models to Japanese culture and customs, and orchestration with Fugu, which intelligently switches between models to hedge against geopolitical scenarios where overseas models become unavailable. m:…1338

Sakana differentiates itself by targeting decision-making support for finance and defense rather than operational tasks like meeting minutes, and its three teams -- Research, Applied, Product -- work as a single cycle. The company is collaborating with the Defense Intelligence Headquarters of the Ministry of Defense, is in discussions to sell the same solutions to overseas financial institutions, and plans to at least double its current engineering headcount. The pitch -- collective intelligence as a national capability -- rhymes with NVIDIA's Europe framing: sovereignty is becoming a product category with reference architectures. m:…1338

A talking Jensen as the demo

The most memorable artifact of GTC Berlin was not a slide but a booth: Synthesia's TJ, a cartoon version of Jensen Huang, deployed as an interactive avatar at STATION-Berlin from Oct. 21 to 22. Built with Synthesia's own video and voice models and trained and run on NVIDIA GPUs, TJ recommends conference sessions based on what attendees say they want to learn, with an LLM matching questions to the agenda and politely redirecting off-topic ones. Everything behind it ships as an Interactive Avatar API embeddable in apps or websites. m:…0805

It is easy to file TJ under conference gimmick, but the throughline with the week's announcements is real: the avatar is a showcase for sovereign, deployable, conversational interfaces -- the application layer on top of the AI factories and open models Huang was selling. The lesson for executives is that the sovereign-AI stack is now complete enough to demo: energy to tokens to a talking avatar, all on your own infrastructure. m:…8373 m:…0805