Nvidia used its GTC Berlin keynote on 21 October 2026 to put numbers on its European push. Its live blog of the event says the company has committed “nearly $10 billion to the technology ecosystem across Europe” and works with “12,000+ partners.” On stage, chief executive Jensen Huang also said Nvidia had announced investments in 14 startups over the past week, according to an automatic transcript of the stream and a post by Wall St Engine.
This is our wrap of the keynote: the moments that did not get their own piece, from a million-fold performance claim to a protest in the hall, plus links to every piece we published from the event. Huang’s pitch to Europe, as he put it: “To be part of this AI economy, you don’t want to just be consuming the AI; you want to be producing the AI.”
Business Pill · SOVEREIGN AI
A short Business Engineer explainer on sovereign AI: why countries and regions want their own AI capacity. It teaches the general idea only and says nothing about Nvidia or GTC Berlin.
Watch · The Keynote
NVIDIA GTC Berlin 2026 keynote with CEO Jensen Huang, 21 October, on NVIDIA’s YouTube channel. Quotations in this piece come from NVIDIA’s written materials and the outlets named, not from the spoken keynote.
The key insight: As we read it, Nvidia’s GTC Berlin keynote was a sales pitch for a whole stack rather than a single chip: Huang moved from AI factories and optics to agent security, open models, telecoms, cars and robots, and tied each to European customers and partners named in Nvidia’s blogs.
Nvidia’s live blog says the keynote came after a weeklong road trip across Europe: Paris, London, Stockholm and Dublin. Wall St Engine’s post on the keynote said “more than $10 billion invested over the past year”; Nvidia’s own wording, in the live blog and in its 13 October France post, is “nearly $10 billion.”
Nvidia’s country posts from the past week name 13 planned startup investments between them, on our count:
France (13 October). Doctolib, Entalpic, Genesis AI and Pigment.
UK (14 October). Boltz, Foresight, Hiverge, Prima Mente and Riverlane.
Nvidia says it has accelerated AI performance a millionfold over 10 years (its own measure; the post gives no benchmark). Huang set that against roughly 100x for Moore’s Law on stage, according to an automatic transcript. The 10,000x gap is our arithmetic. Sources: @nvidia on X; NVIDIA GTC Berlin keynote, 21 October 2026.
What Was Posted on X
With NVIDIA extreme co-design, weโve accelerated AI performance a millionfold over 10 years, and weโre not stopping there.
Nvidia’s official account posted on 21 October that ‘extreme co-design’ has accelerated AI performance a millionfold over 10 years. It is Nvidia’s own measure, and the post does not state the benchmark; our chart compares it with the roughly 100x Huang gave for Moore’s Law, as captured in an automatic transcript of the stream.
Jensen Huang just highlighted NVIDIA’s growing AI presence in ๐ช๐บ Europe, with over 12,000 companies in its ecosystem, more than $10 billion invested over the past year, and 14 new startup investments this week.
Wall St Engine posted that Jensen Huang had just highlighted over 12,000 companies in Nvidia’s European ecosystem, more than $10 billion invested over the past year and 14 new startup investments this week. Nvidia’s own live blog says ‘nearly $10 billion’, which is the figure this piece uses.
Clouds, CPUs and Confidential Computing
A same-day Nvidia blog on its European cloud partners carries most of the hard figures:
Mistral. Nvidia says Mistral Large 4 was “Trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral-operated European data centers,” and that Mistral’s expansion includes Vera Rubin infrastructure “planned for delivery in 2027.”
Nscale. In tests on GB300 NVL72 systems, Nvidia says Nscale fit “37% more GPUs and nearly 50% higher aggregate inference throughput within the same provisioned power budget,” and plans Vera Rubin with “expansion across five European markets.”
Nebius. The blog says Nebius “has adopted NVIDIA Groq 3 LPX” for production inference on its Token Factory.
CodeRabbit. Using a code-review routing model that Baseten built with Nemotron, CodeRabbit “improved routing agreement from 75.8% to 80.7% while reducing estimated inference costs by 50% for the evaluated workload,” according to the blog.
A separate Nvidia blog, dated 20 October, says cloud providers Nebius, Nscale and Verda are building with Vera, Nvidia’s CPU, and that Bull, Dell Technologies, HPE, Lenovo and Supermicro are building Vera-based systems. It also names Baseten-Blaxel, Daytona, E2B, Modal, Prime Intellect and Vercel, which run sandboxes for AI agents, as building on Vera.
A third blog introduces Confidential Bare Metal, which Nvidia calls “a new deployment option” that lets frontier model builders run sensitive work on third-party infrastructure “while retaining control of their models and data.” The model builder it names is Cohere, which “plans to integrate Confidential Bare Metal into Cohere Model Vault.” Google Cloud is extending support to its A4X Max instances, and the blog lists CoreWeave, Lambda, Microsoft Azure Infrastructure, Nebius, Nscale and others as supporting Nvidia Confidential Computing deployments.
What Huang Said on Stage
Nvidia’s official X account summed up one theme: “With NVIDIA extreme co-design, we’ve accelerated AI performance a millionfold over 10 years, and we’re not stopping there.” On stage, Huang set that against roughly 100x that he said Moore’s Law would have delivered over the same decade, according to an automatic transcript of the stream.
On agents, Nvidia’s live blog says Huang described them as thinking, reasoning, iterating and using tools, which drives token consumption as they become more useful. He listed five parts of an agent: a model, a harness that runs its work and memory, a sandbox that isolates the harness, knowledge and tools. “The most important thing is to make sure the agent is secure,” Huang said, according to the blog, which also puts Nvidia’s developer base at “more than 11 million developers.”
Huang also took on the argument that AI agents will put software companies out of business. He said the opposite would happen, and that tools companies would become agents companies, according to an automatic transcript of the stream. The slide behind him, titled “Enterprise Platforms Build Agents on NVIDIA Agent Toolkit,” showed 15 company logos, including SAP, Salesforce, ServiceNow, Siemens and Synopsys.
Keynote slide, NVIDIA GTC Berlin, 21 October 2026 (screenshot of the stream): “Enterprise Platforms Build Agents on NVIDIA Agent Toolkit”, with 15 company logos.
On telecoms, the live blog says Huang “cited 100,000 sites and about 100 gigawatts of distributed data centers”; on stage he presented the shift to AI networks as a prediction, according to the automatic transcript. Our T-Mobile piece covers the AI-RAN call he announced alongside it.
Huang also showed Nvidia’s Spectrum-X co-packaged optics switch and said it is in volume production, according to the automatic transcript. Nvidia’s January technical blog describes the product as “the world’s first fully integrated 512 lane 200G-capable co-packaged switch system.” Our Lumentum optics piece covers the wider demand for optical parts in AI data centres.
A Petaflop Then and Now
Near the end, Huang held up an RTX Spark laptop and compared it with the DGX-1 he delivered to OpenAI. “In the year 2016, I delivered 1 petaflops and 128GB … to a small company called OpenAI,” he said, according to Nvidia’s live blog. “This is 1 petaflops and 128GB.”
The comparison is not like for like. Nvidia’s April 2016 launch release rated the DGX-1 at “Up to 170 teraflops of half-precision (FP16) peak performance,” with “Eight Tesla P100 GPU accelerators, 16GB memory per GPU,” or 128GB in total on our arithmetic. Nvidia’s 16 October release rates RTX Spark at “1 petaflop of FP4 AI performance.” FP16 and FP4 are different number formats, so the two compute figures do not measure the same thing.
Huang also pitched a DGX Station that runs Windows. A same-day Nvidia blog says DGX Station for Windows is “expected from system manufacturer partners in December,” and gives DGX Station “748GB of coherent memory and up to 20 petaflops of AI performance.”
A Protest in the Hall
About a quarter of the way in, while Huang was outlining the Vera Rubin platform, the keynote was interrupted. DCD reported that what appeared to be Greenpeace activists heckled Huang over “environmental destruction” and over claims that Nvidia hardware was being found in Russian drones used in Ukraine. Both DCD and Tom’s Hardware reported a banner reading “Nvidia, clean up your dirty business.”
Huang told security staff to “go easy” on the activists and said, according to DCD, “Give her a chance. Everybody has a chance to say something.” He called the points they raised “all important issues, but let me get through my keynote.” The protesters were ushered out, DCD reported.
The Industrial Demos
A keynote video showed agents at work in industry. At Siemens’ Erlangen factory, the live blog says, “agents help prepare new products for production, testing plans in a digital twin before changes reach the factory floor.” At Dassault Systèmes, engineering agents adapt vehicle design, using simulation to find and fix a problem.
Nvidia’s industrial blog the same day adds figures:
Siemens. It is putting agents built on Nvidia’s FOX factory blueprint into Intelligence Center X “for estimated cost savings of 20%,” and at its Erlangen factory, the blog says, Siemens used Nvidia’s Cosmos tools “to reduce calibration time of inspection processes by 40%.”
Chips. “Benchmarks of S-Litho on imec’s model show up to 40x speedup,” and Infineon reports “a 12x speedup in 3D breakdown simulation” with Synopsys software on Nvidia hardware.
Software. Nvidia says the FOX blueprint is “now available in early access on GitHub,” and that it is releasing cuDSS for general availability, to speed up the calculations behind engineering simulations such as structural stress and fluid flow.
Factories. Audi tests battery-production concepts in a digital twin, and Schneider Electric and AVEVA are developing an AI factory digital twin “built for Nscale,” the blog says. Mercedes-Benz, Leonardo, Saab and Schaeffler are among the other companies it names.
Announced Today, Not on Stage
Nvidia also published a supply-chain release on 21 October that did not come up in the keynote, according to our automatic-transcript notes of the stream. It says “Amkor plans to increase its advanced packaging capacity by more than 300% by the end of 2028.” Our KYEC piece covers the rest of it, including Nvidia’s $1.1 billion capacity reservation advance to King Yuan Electronics for new U.S. testing facilities.
Update, 21 October: What Nvidia’s Partners Announced
Several Nvidia partners published their own GTC Berlin announcements on 21 October. Each line below is the company’s own claim, quoted from its release or post; we have not tested any of the figures.
Vertiv. In a release, Vertiv said its factory-built approach with Nvidia can, “under defined assumptions,” deliver “up to 30% more cumulative token output within 12 months of project kickoff with the same installed capacity.” It said Polar is using the model for “the first 41 MW phase” of its Herøya, Norway campus. Our DSX water and power piece covers the Nvidia design behind it.
Nebius. Nebius “today announced public live spot pricing for GPU compute,” its release says, alongside reserve auctions and a pilot “capacity-rebate program.” Our Nebius spot-price piece has the update.
ServiceNow. ServiceNow said Arc, “a personal long-running AI agent that can take on everyday work, is now generally available,” and that its AI Passport “will be integrated with NVIDIA OpenShell and NVIDIA Sentry.”
CrowdStrike. CrowdStrike said its “Falcon Guardian integration with NVIDIA OpenShell and NVIDIA Sentry will extend AIDR from agent runtimes to hardware-based enforcement.” Our OpenShell piece covers Nvidia’s reference architecture for these tie-ins.
CoreWeave. CoreWeave announced a Zero Trust AI Framework, “a customer-controlled security model that gives customers authority over access to their models, data, and runtime environments,” and Physical AI Launchables, which it describes as “a catalog of physical AI workloads packaged as Templates.”
Amkor. Nvidia’s supply-chain release says it and Amkor plan to target “a greater than 300% increase in interposer wafer capacity by the end of 2028” in South Korea. Our KYEC piece covers the rest of that release.
Aptiv. Aptiv said it is expanding its Nvidia work to deliver “scalable, production-ready automated-driving solutions” built on Nvidia’s Hyperion platform and Halos safety system, including running Thor-based compute on air cooling, which it calls “an industry first for Thor-based computing platforms.” Our Hyperion piece covers the carmakers Nvidia named.
TrendAI. Trend Micro’s TrendAI unit said it is bringing Nvidia’s Nemotron open models into a version of its Vision One platform for sovereign and private clouds, which it says “is available today.”
Cadence. Cadence said Nokia is running its AuraStack AI Super Agent on the Millennium M2000 AI Supercomputer and “is accelerating key design and analysis workflows by 5X.”
Mistral AI. Mistral posted on X that it is “announcing a new collaboration with @nvidia to bring AI to industrial engineering,” listing “faster simulation,” “open benchmarks and supporting datasets for physics AI” and “evaluations for agentic AI in CAE and EDA.” Its blog post says the work will develop “public benchmarks and open datasets for evaluating physics models.”
Cadence (sovereign cloud). In a separate release, Cadence “today announced the availability of a sovereign AI data center in Munich” that combines its OnCloud service, “NVIDIA Vera and Grace CPU systems” and its Millennium M2000 AI Supercomputer. It said its Reality Digital Twin Platform, in “a recent deployment,” helped “increase data center capacity utilization from 71% to 89%.”
KION. KION said a fleet of autonomous vehicles is now in daily operations at GXO’s site in Épinoy, France, and that its approach “has reduced the time required for on-site deployment by over 60 percent.” It said it is integrating Nvidia’s Halos blueprint into an AI-based safety concept.
Supermicro. Supermicro said its servers for Nvidia’s Aerial RAN Computer, built with Nokia’s RAN software, “are generally available now,” with a 2U ARC-Pro system planned for “general availability in the first half of 2027.” Our AI-RAN piece covers the operator side.
HPE. HPE said it is introducing an NVIDIA Vera CPU Rack, which it describes as having “256 NVIDIA Vera CPUs, support for 22.5K sandboxes, and 80% faster agentic CPU performance,” and that it will support OpenShell and DSX POD. It also said Schneider Electric plans to use HPE Private Cloud AI for “Industrial AI Agents at the Edge.”
Dell. According to Techzine’s roundup of what Dell shared for the event, its PowerEdge M9882 “uses 256 Vera CPUs in a 48U rack,” and Dell “is also introducing a Modular Data Center based on Nvidia’s DSX Reference Design.” We could not read Dell’s own release; our DSX POD piece covers Nvidia’s design.
IREN. IREN said (GlobeNewswire release) it “has established IREN Europe and secured the civil works license for its planned AI data center in Badajoz, Extremadura, Spain,” planned at “approximately 300MW of capacity, with initial operations targeted for 2028.” Our Merlin piece covers another European landlord’s AI data-centre exposure.
Update, 22 October: Day Two, Open Models at Banks, Retailers and Governments
Nvidia published four more blog posts early on 22 October listing European customers. Each line below is Nvidia’s description; we have not tested the claims. Two other posts from the same batch have their own coverage: our Cosmos piece (Porsche expects up to 90% fewer repeated test drives) and an update to our DSX piece (US grid queues).
Banks and payments (Nvidia blog). Revolut is scaling its PRAGMA foundation model, “servicing an initial 80 million customers.” Neobank bunq’s assistant on open weight models “matches the quality of leading closed models at 60% of the cost per answer.” Stripe sees “42x faster data pipelines and 92% cost savings” for its Stripe Radar fraud platform “in offline testing.” HSBC reports “5x faster data processing”, and Santander is an early adopter of Nvidia’s Kumo models.
Retail (Nvidia blog). On Shopify’s internal benchmark, its post-trained Nemotron model “surpasses a leading frontier model in accuracy while cutting serving costs by 90% and delivering responses 7x faster.” Mirakl has “onboarded more than 70 million products” with its catalog tool, and Otto Group extended its simulated warehouse work to its largest fulfillment centre, in Haldensleben, Germany.
Weather (Nvidia blog). StormScope EU, a new Earth-2 model trained on European data, predicts cloud conditions across “16 Meteosat channels at 10-minute intervals.” AXA is using Earth-2 to study cyclone risk across “more than 1,000 simulated seasons.”
Governments (Nvidia blog). Poland’s labour ministry is building an agentic system to “support more than 400 labor offices.” In France, “More than 2,500 eligible organizations” can use a sovereign workspace bought through the public purchasing body CANUT. In the UK, Locai Labs released “a dataset of 1 million synthetic personas.” In Germany, secunet, Nvidia, Red Hat and deepset are developing a sovereign cloud on a foundation whose security components are “authorized for use with classified information.”
As we read it, the common thread across the four posts is cost and control: bunq and Shopify report lower costs than leading closed or frontier models, while Mirakl and the government projects stress running models on their own or European infrastructure. That extends the open-models thread in our day-one open models piece. Healthcare items in the public-sector post are left out of this summary.
Connecting the Coverage
This is the hub for our GTC Berlin coverage, all published on 21 October. On infrastructure, our DSX water piece reported Nvidia’s claim that its designs can cut facility cooling water to near zero, putting the potential saving at 54 million liters a day for a 2-gigawatt deployment. Our DSX POD piece covered the containerised Vera Rubin factories Nvidia says can be deployed in as little as 14 days at a prepared site, and our Vera Rubin piece set out three versions of Nvidia’s 30x claim.
On agents and models, our OpenShell piece reported that Nvidia says enterprises can run agents from eight named model builders on its OpenShell design, and our open models piece covered SAP, Salesforce and others working with Nvidia’s Nemotron models.
On networks, cars and robots, our T-Mobile piece reported Nvidia’s claim of “the world’s first AI-RAN call on a commercial network,” our Bolt piece covered Bolt’s choice of Nvidia’s Alpamayo Stack software for its European robotaxi program, and our Agile Robots piece covered Agile ONE SN, the reference humanoid Nvidia showed on stage.
Before the keynote, our Germany pact piece covered the Joint Declaration of Intent that Germany signed with Nvidia at the Invest in Germany Summit on 19 October. Our Lumentum optics piece reported, on 20 October, Lumentum’s chief executive saying the company is “almost sold out for 2029.”
The Structural Read
As we read it, the Europe numbers are a running total, not a new pledge: Nvidia’s wording of “nearly $10 billion” appeared in its France post on 13 October, and the 13 startups we could name were announced across three country posts over the past week.
As we read it, the hardest figures came from partners rather than the stage: Nscale’s 37% more GPUs within the same power budget, CodeRabbit’s halved inference costs for one workload, and Siemens’ 40% shorter calibration time, all as Nvidia reports them.
As we read it, several of the keynote’s biggest lines still need a written source: the Moore’s Law comparison, the volume-production claim for Spectrum-X optics and the 14th startup are stage remarks in the sources we used.
Jensen Huang, NVIDIA founder and CEO, at GTC Berlin, as quoted in NVIDIA’s live blog, 21 October 2026
“Long before the AI models emerge, the infrastructure has to exist”
Three Implications
FOR EUROPEAN COMPANIES Nvidia’s live blog quotes Huang urging them to produce AI rather than only consume it, and Nvidia’s industrial blog names Siemens, Audi, Mercedes-Benz and others using its tools in engineering and factories.
FOR AI CLOUD PROVIDERS Nvidia’s blogs name Nebius, Nscale, Verda, Crusoe and Mistral among those building on Vera Rubin or Vera CPUs, and Nvidia says Nscale fit 37% more GPUs within the same power budget in tests.
FOR INVESTORS The figures that carry dates and amounts are in the written releases, such as Amkor’s planned 300%-plus packaging expansion by the end of 2028, while several keynote figures rest on stage remarks only.
The framework describes AI as a layered stack and argues: “This entire ecosystem has been shaped by constraints, bottlenecks, chokepoints, and scarcity at every layer.”
As we read it, Huang’s five-layer cake of energy, chips, AI factories, models and applications maps onto that view, and the supply-chain release published the same day, on substrates, packaging and testing, is about easing constraints in the lower layers. The framework was published in May 2026 and does not discuss GTC Berlin.
What Is Not Established
Several figures in this piece rest on what Huang said on stage, as captured by an automatic transcript, rather than on a written Nvidia source: the roughly 100x for Moore’s Law, the volume-production status of the Spectrum-X optics switch and the 14 startup investments, of which we matched 13 to names. Nvidia’s million-fold claim is its own measure, and the post does not state the benchmark.
The RTX Spark and DGX-1 figures use different number formats. The industrial and cloud figures are as Nvidia reports them, from tests and early deployments. The 10,000x gap in our chart and the 128GB DGX-1 total are our arithmetic.
Business Engineer Framework
The Nine Layers of AI
A Business Engineer map of the AI stack, layer by layer, and of the constraints and chokepoints that shape each one.
Nvidia’s GTC Berlin keynote set out its European pitch in numbers: nearly $10 billion committed across Europe, 12,000+ partners and, by Huang’s count on stage, 14 startup investments in a week. As we read it, the keynote’s through-line was that Nvidia now sells the whole AI factory, from optics and CPUs to agent security, and wants European companies building on it. Huang closed by telling the audience, according to Nvidia’s live blog: “Everybody can build agents.
Every company in Europe will become an AI company … and we’re going to have a thriving AI economy here in Europe.”
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. This piece is built on NVIDIA’s live blog of its GTC Berlin keynote on 21 October 2026, read in full, NVIDIA’s same-week blogs and releases named in the sources, two X posts, DCD and Tom’s Hardware reports on the protest, and a keynote slide as shown on Nvidia’s stream. Our desk notes of the keynote, made with an automatic speech-to-text tool, were used to find moments; remarks taken from them are paraphrased and not quoted, and the stream itself is the primary record.
Counts and gaps labelled as such are our arithmetic. Healthcare announcements from the event are not covered here. We did not contact Nvidia, Greenpeace or the companies named. Nothing here is legal, financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
