ATAILA Newsroom · Budapest ·
2026-W43
NVIDIA tells Europe to produce its own AI. We read GTC Berlin with ours.
The message from Berlin was one we make to Hungarian buyers ourselves: start from an open model, train it on your own data, and run it on infrastructure you control. This time it came from the company that sells the chips. We follow GTC Berlin the way we would want a customer to follow anything that matters to them: with our own tools, in our own network. This is our reading of what NVIDIA said, how we read the event, and which parts of the stack stay NVIDIA's.
The article we are responding to
“Nvidia woos wary Europeans with more open weights models”
theregister.com, 21 October 2026
What NVIDIA said in Berlin
NVIDIA GTC Berlin 2026 ran from 20 to 22 October: full-day workshops on 20 October, the conference on 21 and 22 October at STATION-Berlin. Jensen Huang gave the keynote on 21 October at the Tempodrom. NVIDIA's own live blog speaks of "more than 100 sessions". The session catalog, read on 22 October, lists 152 sessions on 16 stages.
Huang's line on Europe was short. "To be part of this AI economy, you don't want to just be consuming the AI; you want to be producing the AI," he said, and added: "We could do all of that here in Europe, and in fact, we've already started." NVIDIA's blog puts the commitment behind it at nearly $10 billion to the technology ecosystem across Europe, with 12,000+ partners, after a week of stops in Paris, London, Stockholm and Dublin.
The route NVIDIA proposes runs through open models. In Huang's framing, NVIDIA provides the models, the data and the scripts to reproduce them, and a company post-trains a model on its own data, which he called "sovereign intelligence". NVIDIA's open-models post of 21 October, by Kari Briski, says the models give organisations a foundation they can customise with local data and "deploy on infrastructure they control". A separate NVIDIA post on the European public sector makes the same case for governments.
The Register's Tobias Mann read this as a pitch to wary Europeans (21 October). The new Nemotron Retriever models for embedding and reranking are based on Nemotron 3.5, handle images and text across 34 languages, and come in sizes from one to eight billion parameters. A refreshed Nemotron Parse model extracts information from PDF documents; NVIDIA claims roughly 10.5 pages a second for the one-billion-parameter model on a single H100. The Register's point is the release practice: alongside the weights, NVIDIA routinely publishes the training data and the recipes, so a customer can inspect the model or build its own on the same architecture. It adds that NVIDIA is not alone in this; the Allen Institute has followed a similar strategy with its OLMo models.
Two infrastructure announcements came the same day. DSX POD is a reference design for Vera Rubin AI factories built by partners in modular containers. NVIDIA says it cuts deployment at a prepared site from six months to as little as 14 days, and partners are to deliver the units worldwide in early 2027. The Register notes that the 14 days exclude getting power, connectivity and the site ready. The OpenShell Reference Architecture is a design for safe AI agents built on OpenShell, a policy-controlled sandbox runtime. To implement it, The Register writes, organisations need NVIDIA hardware: Vera Rubin systems, Vera CPUs and BlueField-4 DPUs.
Two remarks frame the rest. Forrester's Paul Miller, quoted by The Register on 21 October, said NVIDIA "has invested almost $10 billion into Europe in the past year", but is not European, and that anything with critical ongoing dependencies on NVIDIA or any other non-European firm "is unlikely to meet the purest definitions of European tech sovereignty". And The Register reported on 22 October that Huang put the price of licensing NVIDIA's software stack at about $2,000 per GPU annually.
In Hungarian, Portfolio carried an item on 22 October, citing CNBC: Huang says demand for neocloud providers exceeds expectations but working capital is short, and in Europe scarce energy and high electricity prices make it harder. Portfolio mentions NVIDIA's "közel 10 milliárd dolláros" European package. It does not mention Berlin or GTC. We found no Hungarian item on the keynote itself.
nearly $10 billion
committed by NVIDIA to the technology ecosystem across Europe, with 12,000+ partners (NVIDIA keynote blog, 21 October); Forrester's Paul Miller says almost $10 billion invested in Europe in the past year (The Register, 21 October)
34 languages
images and text covered by the new Nemotron Retriever embedding and reranking models, offered from one to eight billion parameters (The Register, 21 October)
as little as 14 days
DSX POD deployment at a prepared site, down from six months, per NVIDIA; announced 21 October, deliveries in early 2027 (NVIDIA press release, 21 October)
about $2,000
per GPU annually: Huang's price for licensing NVIDIA's software stack (The Register, 22 October)
152
sessions on 16 stages in NVIDIA's GTC Berlin session catalog, read 22 October; 59 of them are of types NVIDIA does not usually record
20
day-1 sessions on NVIDIA On-Demand at about 21:40 CEST on 22 October, none from day 2 yet; the keynote and the pregame are on NVIDIA's YouTube channel (our check, 22 October)
How we read GTC Berlin
We follow the event through our own local AI tools. NVIDIA publishes the recordings: the keynote (1 h 34 min) and the pregame show (2 h 06 min) on its YouTube channel, the sessions on NVIDIA On-Demand, which on day 1 took 5 to 27 hours after a session ended. Each recording NVIDIA publishes goes through OpenVideo, our video-to-text application, which runs on our own AI nodes. It produces a timestamped transcript that tells speakers apart, with a summary, keywords and chapters. Nothing is uploaded to an external AI provider, and nothing leaves our network.
The transcripts go into a brain built for this event alone, a memory bank kept separate from every other event. Our agents answer questions over it: what was said about a topic, which sessions covered it, what a given speaker argued, with the session and the moment in the recording cited. AI Summit Budapest (7 and 8 September) was the first event we handled this way; our notes on it are in the newsroom. GTC Berlin is the second.
The state on the evening of 22 October: the event was set up in OpenVideo, and transcription had started with what NVIDIA had published. That was the keynote, the pregame, 20 day-1 sessions on On-Demand (one of them behind a free NVIDIA Developer login) and official extras such as the keynote films and NVIDIA Developer livestreams. Nothing from day 2 was public yet. The chat over the brain was still being wired to its per-event bank.
This is private use. NVIDIA publishes the recordings; we read them for our own understanding. We do not republish recordings or transcripts. When we refer to a session, we link NVIDIA's own pages: the session catalog, the keynote on YouTube, NVIDIA On-Demand.
The method is the point. A small team cannot sit in 152 sessions on 16 stages. It can read every session NVIDIA publishes, with tools that run where its data lives. That is the setup Huang described for Europe, applied to a conference.
The layer we run, and the layers we do not
What NVIDIA described for Europe is the layer ATAILA works in: open-weight models, adapted to the customer's own data, running on hardware the customer controls, inside the EU. ATAILA Cloud runs them managed and multi-tenant on our EU hardware. The Enterprise edition runs them single-tenant on the customer's own hardware. Both come with sandbox, dev, uat and prod environments and 99,9% availability. In September we argued the same direction from the buyer's side: bring the model to the data.
NVIDIA's release practice helps here. A model with published data and recipes can be checked, not only trusted. NVIDIA publishes the new Retriever models on Hugging Face, among other places, and in August we wrote about the reported NVIDIA deal for Hugging Face and why that concentration worries us. Open weights are only as open as the place they are downloaded from.
Here our argument is weak, and we say so. The silicon is NVIDIA's. The DSX POD and OpenShell reference designs are NVIDIA's, and the OpenShell design needs NVIDIA hardware underneath. The software licence Huang priced at about $2,000 per GPU a year is NVIDIA's. Miller's sentence about critical ongoing dependencies applies to us as much as to any AI cloud in Europe. The AI nodes that transcribe the Berlin sessions run on NVIDIA GPUs too.
What we own is narrower, and we keep the claim narrow: who operates the system, under which jurisdiction, and where the data goes on its way through. We drew the same line on 1 October about AWS's sovereign cloud test: sovereignty has layers, and nobody in Europe owns the silicon layer in this decade, including us.
What we concede, and what we will be asked
NVIDIA now hands out much of what an operator used to add. Models, data and recipes; reference designs for AI factories and for agent sandboxes. A sceptical reader will ask what is left for a company like ours. Our answer is the operation, the jurisdiction and the data path. That is a smaller claim than "sovereign AI", and it is the one we can back.
Our view of the event is a view of the recordings. 59 of the 152 sessions are expert sessions, labs, workshops and certifications that NVIDIA does not usually record. Day 2 was not public on 22 October. The brain answers from transcripts and slides, not from the room, and it knows only what NVIDIA chooses to publish.
The headline numbers are NVIDIA's. The nearly $10 billion, the 14 days and the 34 languages are vendor statements. The Register itself advises taking vendor claims with a grain of salt, and the 14 days start only once power and the site are ready.
Hungarian coverage is thin. The one Hungarian item we found, Portfolio's, is based on CNBC, says an AI assistant helped prepare it, and does not mention Berlin.
What we commit to: anything we write about a GTC Berlin session names the session, links NVIDIA's own page, and says when it comes from a transcript. We do not republish NVIDIA's recordings or our transcripts, and none of it leaves our network.
Sources, figures, and one disclosure
The story we are answering: Nvidia woos wary Europeans with more open weights models, Tobias Mann, theregister.com, 21 October 2026, 20:00 UTC. Opened and read in full. The address without the trailing article number returned 404 on 22 October; the link above is the working one.
Hungarian coverage: Az Nvidia vezére megszólalt az AI iránti keresletről, unsigned, portfolio.hu, 22 October 2026, 13:52. Opened and read in full. It summarises CNBC; the CNBC original was not opened. Portfolio notes that an AI assistant helped prepare the article.
Keynote facts and quotes: GTC Berlin 2026: Live Updates From NVIDIA CEO Jensen Huang's Keynote, Brian Caulfield, NVIDIA Blog, 21 October 2026, read in full. NVIDIA Open Models Advance Europe's Sovereign Intelligence Across Industries, Kari Briski, NVIDIA Blog, 21 October, read in full. NVIDIA Open Models Power a New Wave of European Public Sector Projects, NVIDIA Blog, 21 October, read in full. NVIDIA Cloud Partners Scale Across Europe, NVIDIA Blog, 21 October, read in full, not quoted. DSX POD: NVIDIA press release, 21 October, read for the deployment and availability wording; the companion European DSX release, same date, was read in part and is not quoted.
Further reporting by The Register, each read in full: Jensen Huang sends in the robot ducks to sell Nvidia's AI-everywhere vision (21 October), the source of the Forrester quotes and the note on site readiness; Nvidia picks OpenShell to tame AI gents (21 October), the source of the hardware requirement; Nvidia's Huang: Don't mention the AI-pocalypse (22 October), the source of the $2,000 figure.
Event counts: NVIDIA's session catalog API, read on 22 October 2026 (152 sessions, 16 stages, 59 of types not usually recorded). Recording status: NVIDIA On-Demand and NVIDIA's YouTube channel, checked on the evening of 22 October; the On-Demand count rose from 18 to 20 during that check, so a later reader will find more. OpenVideo's description is from openvideo.hu as read on 22 October.
Not opened: Bloomberg's keynote live blog (paywalled) and the CNBC interview behind Portfolio's item. Read but not used: the reports on the protest during the keynote, which this article does not cover.
The disclosure: this newsroom is drafted with AI models and checked by a person before it is published. Given the subject, it seemed right to say so.
The original article
“Nvidia woos wary Europeans with more open weights models”
theregister.com, 21 October 2026
- Read it live at theregister.com
- ATAILA archive, captured 22 October 2026
The live page has changed since we captured it.
← Back to the Newsroom Press enquiries: contact us