The elves are packing

Talk to people in Germany these days, whether in boardrooms, at conferences, in the school café or at the supermarket checkout, and the same tune keeps playing. Everything is getting worse.

The mood is downright tolkienesque: Middle-earth at the end of the Third Age. The last elves are heading for the Grey Havens, the light is fading, the great ones are leaving. Those left behind stand on the shore, not quite sure what they are still staying for.

With one difference: nobody here is sailing.

It's always someone else's fault

Why is that? The answers vary, but almost all of them share one thing: the cause lies elsewhere.

Politics. Bureaucracy. The EU and its tethered bottle caps. The polycrisis. Foreigners. Hardly any of these explanations contains the word “we”. How convenient.

And when nothing else works, there is always data protection. The word is dropped, everyone nods gravely, the project is off the table. Rarely does anyone ask which article of the GDPR actually stands in the way. On closer inspection, little usually remains that a bit of work couldn't solve.

Data protection is rarely the reason. Most of the time it is the alibi.

Maybe it's much simpler

Of course there are structural problems, and they have to be tackled at a structural level. Still, it is worth looking at the individual. Because part of the problem may be closer to home than we'd like. Maybe the real question is:

Are we even up for it anymore?

There is plenty of room to speculate why the answer is so often lukewarm. Affluence fatigue, demographics, permanent crisis mode, doomscrolling.

When it comes to what needs doing now, though, the diagnosis doesn't matter. The treatment stays the same: simply want it again. Play offence again instead of administering things into oblivion.

That sounds like a calendar quote. It isn't. It can be observed, namely in the places where it works. These places could be called “islands of will”.

Etteln beats Hong Kong

Etteln is the southern district of the municipality of Borchen near Paderborn. Roughly 1,800 residents, half-timbered houses along the main road, a marksmen's club. At first glance, a village like a thousand others in eastern Westphalia.

At second glance, Etteln may be the most digital village in the world.

There is a village app for events, swaps and town hall updates. There is a digital ride-sharing bench. There is e-car sharing and e-cargo-bike sharing, digital information steles and a digital twin of the village, mapped by drone. A municipal data space is in the works. A quick search leads down a deep rabbit hole of village digitalisation.

Fibre? The people of Etteln laid it themselves in the outlying areas in 2020, in the middle of the Covid lockdown. 65 residents, around 30 kilometres, some 3,500 hours of digging over six months.

At the end of October 2024, Etteln won the IEEE Smart Cities Contest, run by the world's largest engineering association. Runner-up was Hong Kong.

No magic potion

Nobody in Etteln is brewing a digitalisation magic potion. No different data protection laws apply. Nor does the village sit under a dome that mysteriously keeps the world's polycrises out.

The difference shows at one of the information steles outside the community hall: the marksmen's club poured the foundation, the church association laid the cables, the local village association did the installation.

There are simply people there who are up for it.

And that is the good news. Changing laws takes years, reforming the EU takes decades. Wanting it can start tomorrow morning.

The elves are welcome to sail. We'll stay here and lay fibre.

Germany is the most innovative country on earth. When it comes to risk management.

I say that without malice, I am exactly like this myself. Show me a form and I will fill it in before I read what it is for. And I have spent the last two weeks reading the programme of the Digital Summit that runs today and tomorrow in the Wilhelm Hallen in Berlin-Reinickendorf. Twentieth edition; the first one was held in Potsdam in 2006 and was still called the IT Summit. This year's motto: "Shaping AI: safe, sovereign, competitive". The order is the message. Safe comes first, competitive last, and in between sits a word we have used so often over the past three years that it is slowly losing its meaning.

The roadmap

This afternoon the government presents its AI roadmap, the final report of the AI task force the digital ministry set up in June. Before that it passes through the National Security Council. Eighteen cross-ministry measures. Germany is to become a "sovereign AI nation" by 2030. Eighty percent of large companies and 55 percent of SMEs are to use AI, 90 percent of federal agencies are to run at least one application in production, AI compute capacity is to grow 300 percent over 2025. Add an AI safety institute, a national AI supervisory authority whose draft law has been on the table since February, and the wish to make the EU AI Act "simpler", right after the Digital Omnibus pushed its deadlines further out.

I have nothing against any single one of these measures. I have something against the sum. We are building the safety institute, the supervisor, the security council and the target picture before we have a single product any of it could refer to. It is like ordering the fire code, the house rules and the insurance broker for a house, and then wondering whether you might need a plot of land.

The bill arrived in late September

WIPO's Global Innovation Index came out on 29 September. Germany ranks 11th, out of the top ten for the second year running; in 2022 we were still eighth. Look at the sub-indicators and the whole country is written on one page. Rank 3 in human capital and research. Rank 5 in the share of engineering graduates. And then: rank 60 in the use of digital information technologies, three places worse than last year. Rank 36 in culture and policies supporting entrepreneurship. Rank 94 in inbound foreign direct investment.

We know everything. We use nothing. That is risk management in its purest form, and nobody does it like us. Eva Schewior, president of the German Patent Office, put it diplomatically: we need to get better at turning research results "into attractive products and business models, especially in applying digital technologies". The patent office. The people who make a living administering ideas are telling us we administer too much.

Merz asks Jensen

The scene I cannot shake off for two weeks now: the Chancellor asks Jensen Huang which rules the EU should set for AI. I have treated Nvidia generously in the last two columns and I stand by it: the Industrial AI Cloud in Munich and Blackwell capacity for the Mittelstand are the right call. But you do not ask the chip salesman how to regulate his own market. You would not even do that at a car dealership.

And the chip salesman has just had an interesting month himself. Nvidia closed Friday at 204 dollars, around 15 percent below its yearly high, after October shook the whole sector. The numbers themselves are absurdly good, revenue doubled last quarter. What the market is questioning is the financing behind it. SpaceX borrows 40 billion dollars through an Apollo-managed vehicle to buy Nvidia hardware. The chips are leased, the debt is securitised, the data centres get built before anyone knows who pays for them. Meanwhile Washington closes export loopholes to China, Beijing answers on 16 October with a draft of its own export controls on language models, chip designs at 14 nanometres and below, and compute clusters. In the middle sits Europe, which builds the industry's most important machine at ASML and, in the words of its own management, sells "essentially nothing in Europe".

That is the partner we buy capacity from and label sovereignty. Bought chips are capacity, not sovereignty. I wrote that last week and October only underlined it: capacity from a supplier whose share price hangs on a financing question is a concentration risk. And in a country that has an institute for every risk, this one appears in no roadmap.

The risk that comes with an invoice

There is one dependency that comes with three numbers. 274 million euros, 347 million euros, 481 million euros. That is what the federal government spent on Microsoft licences in each of the last three years, according to its own answer to a parliamentary question. Plus 38 percent in a single year, further price increases announced for 2026. A European Parliament study puts the money flowing from the EU to US tech companies at around 264 billion euros a year. Markus Beckedahl said the sentence on ZDF today that sums up the whole summit: "There is always a pragmatic reason to stay with the Big Tech monopolists after all."

A country that was actually good at risk management would have those three numbers at the top of every summit agenda. A single supplier, double-digit price increases, access by a foreign government confirmed under oath. Instead we debate whether the Deutschland-Stack should make open source mandatory or merely recommend it. The Open Source Business Alliance calls the current draft "fragments" with "too many loopholes for closed-source offerings". Second consultation round. Maturity model postponed. Risk management.

Kiel never found the pragmatic reason

And then there is Kiel. The state chancellery of Schleswig-Holstein has been running on Linux since September. More than 200 desks, around 20 still pending because of specialist applications, the ministries follow step by step, support comes from Dataport and Gonicus. The year before they had replaced Office with LibreOffice, Exchange with Open-Xchange, 44,000 mailboxes, video calls on OpenTalk. Nextcloud is coming, Active Directory is going. The state saves around 15 million euros in licence costs in 2026 alone. Dirk Schrödter, the minister who pushed it through, calls the goal a "future-proof, stable, secure and above all digitally sovereign working environment". No summit, no roadmap, no security council. A state that at some point simply stopped asking.

I have spent a lot of time on a Linux machine myself lately, with zero ideology, because CachyOS is simply faster on my hardware than anything I had before. And something struck me that never comes up in the sovereignty debate: every single Nvidia machine we are buying for the Mittelstand runs on Linux. DGX, the Industrial AI Cloud, every supercomputer in the Top500. The entire AI wave we are discussing in Reinickendorf today runs on open source. The only thing in Germany still paying for an operating system is the public sector. We already have the sovereignty, it sits in every rack. We just put half a billion euros of Office licences on top.

Why I still do not buy the "left behind" narrative

Because the numbers say something else the moment someone builds instead of administers. In the third quarter, 5 billion dollars of venture capital flowed into Germany, the strongest quarter since 2021, second in Europe behind the UK and ahead of France. Helsing 1.8 billion, Quantum Systems 1.2 billion, Proxima Fusion, The Exploration Company. In fintech, KPMG puts Germany at 1.6 billion dollars for the first half, again second in EMEA and the only large market bucking a region that fell 37 percent. And in the FT ranking of Europe's leading start-up hubs, UnternehmerTUM from Munich is number one for the third year running, followed by Start2 and BayStartUP, all three from Bavaria. Plus more hidden champions than any other country on earth.

So the money does come. It comes to people who build drones, fusion reactors, space capsules. It comes to a state that installs Linux instead of launching a third consultation round. It does not come to roadmaps.

That is why Germany is not left behind for me; I wrote that two weeks ago and I stand by it. We are the country with the third-best knowledge in the world and the sixtieth-best nerve to apply it. Kiel and the roadmap want exactly the same thing. The difference is the moment you stop asking the chip salesman for the rules and simply switch the machine on.

Last week I wrote here that Germany hasn't been left behind and is merely leaving its substance lying idle. And that this one week brings everything needed to change that: a date, a budget, participants. Now the week is over. Time to take stock.

What actually happened

Monday evening, Kraftwerk Berlin, red carpet. Merz opens the federal government's first investment summit, Huang sits next to him and talks about robots and autonomous cars as if he were a member of the cabinet. A Joint Declaration of Intent was signed, officially an initiative to develop cutting-edge AI models for Physical AI. Around a dozen German companies are involved, Siemens, Audi, Black Forest Labs. The federal government is reportedly putting in a three-digit million sum, Nvidia engineers and chips. Nobody named concrete figures. Huang mentioned in passing that Nvidia invested almost one billion dollars in Germany last year, and Merz called on German companies to open up their industrial data for AI.

Then Wednesday, Tempodrom. The queue at Askanischer Platz stretched all the way around the entire football pitch; the Tagesspiegel quoted someone who compared it to a Taylor Swift concert. Many didn't get in at all and watched the keynote on their phones, a few meters from the stage. Huang backed up the show with numbers: almost ten billion dollars invested in Europe last year, 14 startup investments on the tour through France, the UK, Sweden and Germany, more than 12,000 partnerships on the continent. Plus the DSX Pod, a complete data center in a container, prefabricated and validated, so that an AI factory is up and running in weeks instead of six months. First systems in early 2027. In parallel, Google and Nvidia announced clusters with up to 960,000 GPUs.

Twenty minutes in, in the middle of the Vera Rubin presentation, Greenpeace activists stood up and wanted to know what Nvidia is doing about its chips in Russian drones. Huang told security to take it easy, everyone was allowed to say something, he would talk to them later. Then he carried on. At the end, the little robot ducks came on stage and were supposed to take a nap; one just kept walking and then toppled over. Two of them pulled off the somersault. After that, an AI-generated Eurovision video with singing robots.

I watched it on the stream and loved the ducks most of all, precisely because they fell over. This is what Physical AI looks like right now: impressive, expensive, and when you say "take a nap," something topples over.

The sentence of the week

The most interesting sentence of the week appeared in an assessment of the Nvidia pact by the Science Media Center: buying chips gets you capacity, and sovereignty doesn't come with it.

Heise put it more sharply and wrote of sovereignty by Nvidia's grace: Nvidia uses Europe's call for independence to entrench itself as an indispensable supplier, thereby deepening the dependency further. In the same week, Bitkom measured that 96 percent of German companies source digital products from abroad and that a loss of these imports would end their business after 16 months on average.

That is all correct. And it is still the wrong yardstick for this week.

Europe does not have sovereignty in high-performance chips this decade. Anyone who attaches that as a condition to every cooperation builds nothing at all. The researchers at the Science Media Center say so themselves: anyone who, in pursuit of chip independence, cuts themselves off from the available cutting-edge technology today will also lose out in developing their own systems. The dependency of German companies on the US currently arises mainly from their use of US models. Chip purchasing, by comparison, is the smaller issue.

And that's where it gets interesting. Because that is exactly what Huang said on stage on Monday: data is a national resource that you keep in the country and turn into your own intelligence instead of exporting it. That same evening, Merz called on companies to open up their industrial data. The US corporation and the Chancellor are talking about the same lever. And the lever lies with neither of them.

Who owns the models?

The question this week left open is: who owns the models trained on the purchased chips?

If a supplier from the Allgäu contributes its twenty years of machine data to a Physical AI initiative, and the result is a model that Siemens licenses and Nvidia hosts, then Germany has bought capacity and given away its substance. If the same model belongs to the supplier, or to a consortium of twenty suppliers, then the pact has done exactly what it says on the label. Both are possible under Monday's declaration of intent. Nothing has been decided.

The 1,700 Mittelstand world market leaders I cited last week as an argument against being left behind are now the ones being addressed, well ahead of Nvidia and the Chancellery. The data has been sitting on factory floors for decades, never trained on, because there was no reason to. Now there is a reason, a container that is up and running in weeks, and a federal government that co-funds it. What's missing is the decision about who ultimately holds the key to the model. And no summit makes that decision; every company makes it for itself.

A salesman, yes. So what?

Under my post from last week, several people commented that Huang is just a salesman, that his statements serve the share price, supposed hope. All true. Nvidia won't have become one bit more European this week, and the keynote was a two-hour product catalog with ducks.

Except: a salesman who delivers rails doesn't decide what runs on them. That was true of Megaupload and it's true of GPUs too. Infrastructure has to pay for itself, and besides the corporation itself, any research that would otherwise have no computing time benefits from that as well. That a US corporation takes the stage in Berlin and recites to us the recipe for our own sovereignty is a little absurd. It is still the best news of the week. The ingredients are here. Nobody has taken them away from us. We've just never mixed them together.

I remain a fan of the company. As a ten-year-old I couldn't afford their cards; today they deliver data centers in containers to Berlin. What we do with them is not their job.

What he really wanted in Berlin

On Friday evening, the head of the world's most valuable company stood on the third floor of the Saturn store on Alexanderplatz, raffled off five laptops, signed autographs and celebrated with 500 people. Nvidia is worth more than five trillion dollars on the stock market. There were seven launch events worldwide for the new RTX Spark laptops, and Berlin was the only one Jensen Huang attended in person (Tagesspiegel).

There is more to it than a PR stunt: the raffle kicked off a week that could move quite a lot for Germany. Tonight, Friedrich Merz opens the Invest in Germany summit at Kraftwerk Berlin with around 200 CEOs, and Huang is the announced star guest, speaking with the Chancellor on stage (Berliner Zeitung, Leadersnet). That will be followed by a private conversation between the two, then a meeting with Mercedes CEO Ola Källenius. Starting tomorrow, GTC Berlin, Nvidia's developer conference, runs at the Tempodrom, with Huang's keynote on Wednesday (Nvidia).

The real reason for the trip is the AI pact that Merz and Huang plan to announce today. The federal government is putting in a nine-figure sum, Nvidia is supplying engineers and chips, and around a dozen German companies are taking part, including Siemens, Audi and Black Forest Labs. The goal is physical AI, meaning models for robots and autonomous vehicles, trained on German industrial data using computing capacity in Telekom's data center (Borncity, WinFuture).

Asked why Berlin, Huang himself gave the clearest answer: Germany is Nvidia's largest European site, and it has telecommunications, robotics, autonomous vehicles and plenty of AI research. And he continues to call the auto industry, which everyone is currently condemning, extremely important for Nvidia, from styling to simulation and manufacturing to the technology inside the vehicle. Someone who talks about a country this way has not given up on it. He wants to sell there.

The remark that stuck

What stuck from this week above all was a single remark from London. At a press briefing on Wednesday, Huang said, in essence, that nothing pained him more than seeing one of the most technologically advanced countries fall so far behind. He suggested the reason was the European habit of overthinking everything, and added that Germany has so much talent: just get to work (Leadersnet, GameStar).

I have looked at that remark for three days, and I find it wrong in its choice of words and right in its intent. Grading countries is an analyst's job; Huang is the best salesman of his generation, on a European tour through Paris, London, Stockholm and Berlin (Handelsblatt). A salesman tells the customer that he urgently needs something, and needs it now.

You can see this in what he said on ZDF two days later: German mechanical engineering and the auto industry are exactly the capabilities that need to be combined with AI (ms-aktuell). Nobody praises the substance on Friday of a country they wrote off on Wednesday. The London remark was urgency, the Berlin remark was the offer. The two belong together, and both aim at the same thing: speed.

And honestly, in Nvidia's case I buy the role. Over 30 years, the company has turned graphics cards for gamers into the infrastructure for the entire AI world, and for years Huang has been wearing the same leather jacket and carrying the same message. Anyone selling Blackwell chips has an interest in Germany getting going. In this case, that interest is identical to ours.

This is not what left behind looks like

Being left behind means you have lost your substance. Germany is leaving its substance unused, and that is a completely different problem with a completely different solution. A few figures that got lost in this week's excitement:

• Germany invests around 3.1 percent of its economic output in research and development and has about 1,700 mid-sized world market leaders, the hidden champions Huang raved about in London (Webnachrichten / IPPEN).
• Since September, Jülich has been home to JUPITER, Europe's first exascale supercomputer and, according to Merz, number four worldwide. 500 million euros, nearly 24,000 Nvidia GH200 Grace Hopper superchips, with waste heat warming the campus (EuroHPC, Börse Express).
• Since this year, Telekom's Industrial AI Cloud has been running in Munich: around 10,000 Blackwell GPUs, up to 0.5 exaFLOPS, one billion euros, entirely privately financed and taken from idea to operation in six months (Telekom, heise). Six months. That is hardly the pace of a country that has been left behind.
• Black Forest Labs from Freiburg sits in the pact alongside Siemens and Audi. It is one of the most relevant image model teams in the world, founded by people who previously worked on Stable Diffusion at Stability AI, and it comes from the Black Forest rather than Palo Alto.
• And Nvidia itself has its largest European site here, ahead of London or Paris. Huang said so himself in Berlin on Friday (Tagesspiegel).

Computing power, talent and industrial know-how are all there. What is missing is the willingness to connect it all without three committees and two expert reports. The hardware is in Jülich and Munich. The data sits on the factory floors of 1,700 world market leaders. What is missing in between is the courage to simply start and let the first model be bad.

I see this every week in my own surroundings. Between the Berlin lab and the Chiemgau, where I am sitting right now, there are hundreds of businesses whose machines generate data that no one has ever fed into a model. It would have been possible long ago; nobody ever asked for it. That is untapped potential, and you unlock it with a concrete occasion.

Why the pact is an opportunity

The opportunity lies in the part of AI where Germany can still win: physical AI. With language models, the train has left for the US and China, and we have to be honest about that. With robots, factories, vehicles and machines that move things in the real world, the winner is whoever has the data on how the real world behaves. That data sits in German factory halls, test benches and vehicle fleets, and that is exactly what the pact targets, with robotics and autonomous driving at its core (Borncity).

The second point is sovereignty, and here I find Huang's definition more useful than most contributions to the German debate. He says, in essence: Nvidia sells chips and does not offer end-user services; the services should run in Germany so that control over data and the ability to act stay here (Tagesspiegel, Dr. Web). Of course, that is a chip seller's definition, and the supply chain remains a dependency. But Europe will not build its own GPU manufacturing in the next ten years, and anyone who waits for it will lose the application layer along with it. Better Blackwell chips in Munich under German operation than no models at all.

The third point is the timeline. If the federal government comes in with a nine-figure sum, Siemens, Audi and Black Forest Labs take part, and the computing capacity in the Industrial AI Cloud is already in place, then the project needs one thing above all: mid-sized companies that put their data in. That is the real invitation of this week. The pact is the scaffolding; the businesses have to put up the walls.

And yes, Nvidia earns money at every single step. That is no argument against it. Nobody asked Deutsche Bahn to lay tracks free of charge when the country was industrializing. Huang is delivering the rails for the next industrialization, and he is delivering them to Berlin, of all places, instead of Paris or London. That can be read as a compliment, even if it was packaged as a reprimand.

The kick in the pants

Germany has had enough wake-up calls over the last ten years. What helps now is a kick in the pants in the best sense, meaning a concrete occasion with a date, a budget and participants. That is exactly what this week in Berlin is: summit today, pact today, GTC from tomorrow, keynote on Wednesday. Anyone with responsibility in an industrial company has no excuse for the next four days not to take a look.

For two years with EXECUTE, I have done nothing else but build bridges between the Berlin lab and Bavarian mid-sized companies, and I know both sides. The Berlin scene likes to talk about sovereignty and builds little. The Mittelstand builds every day and never talks about AI, because it is too busy staying a world market leader. The pact is made for the second group, and they don't know it yet.

My suggestion to anyone who has a factory hall, a test bench or a fleet: watch the keynote on Wednesday and ask yourselves with every example whether your machine would have the data for it. The answer is almost always yes. Then call someone who knows how to get the data into a model. No expert report, no advisory board, just a pilot on one machine, with results in twelve weeks.

Huang is right about the overthinking. He is wrong about being left behind. You are left behind when you can no longer catch up, and Germany can, with JUPITER in Jülich, 10,000 GPUs beneath Munich and 1,700 world market leaders who have never touched their data. On Friday the man raffled off laptops on Alexanderplatz, today he negotiates with the Chancellor, and on Wednesday he delivers the keynote at the Tempodrom. There will be no bigger invitation than this.

Johannes Lex is co-founder of EXECUTE in Berlin and works from the Chiemgau on industrial, robotics and venture projects. This text is part of the Monday column, which appears here every week.