US bans DeepSeek from federal agencies, Mistral Timestral

Samantha: Welcome to The State of Tech, The European Edition, Tuesday October twenty, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: the US bans Chinese AI model DeepSeek from federal agencies, Meta reveals its gigawatt Prometheus supercluster in Ohio, Google and Synaptics put multimodal AI into tiny edge chips, Germany pushes for billions in fresh AI investment ahead of talks with Nvidia, and two things to try yourself: the free writing app FocusWriter alternative Manuskripte, and the open-source route planner BikeCitizens. Let's start with Washington drawing a hard line on Chinese AI.

US Homeland Security bans Chinese AI model DeepSeek from all federal civilian agencies.

Samantha: The US Department of Homeland Security is formally banning the Chinese AI model DeepSeek from federal civilian agencies, citing what it calls multiple threat vectors to government IT systems. The order will appear in the Federal Register and tells agencies to exclude DeepSeek from future procurements and remove any existing deployments. Crucially, the ban reaches further than the model itself. It also covers third-party products and services built on top of DeepSeek, which drops a very awkward question on every federal supplier: can you actually prove what is inside your software?

Bob: And that supply-chain reach is really the sting in the tail. Many commercial AI products quietly fine-tune or wrap open-weight Chinese models, often without flagging it to customers. Federal buyers now need a paper trail showing which base model sits underneath each tool. Expect compliance teams to send urgent questionnaires to vendors this week. Washington is treating DeepSeek less as a single app and more as a contamination risk running through the software stack, similar to how it handled earlier bans on specific Chinese telecom and surveillance gear.

Samantha: For Europe, this lands in the middle of an ongoing debate about AI sovereignty. European governments have not moved to ban DeepSeek, but several national cyber agencies have already warned public bodies to be cautious with Chinese generative models, especially on sensitive data. A US federal ban raises the pressure on Brussels and national capitals to clarify their own line. It also matters commercially, because European firms that sell software to US government customers will now need to document their model provenance in detail or risk being locked out of contracts.

Bob: There is a second order effect worth watching. If US federal buyers demand model-level transparency, that standard tends to spread. Large European enterprises, especially in finance, defence and healthcare, often mirror US federal procurement rules in their own vendor checks. So a bank in Frankfurt or a hospital group in the Netherlands may soon ask the same question: which foundation model is really powering this assistant, and where were its weights trained? DeepSeek itself has not publicly responded in detail, and the practical enforcement timeline for existing federal deployments is still unclear from the order as published.

Samantha: The bottom line is that geopolitics is now baked into every AI purchasing decision, not just the headline model choice. For European listeners this is a nudge to look at their own supply chains before regulators or customers do it for them.

Mistral Goes Beyond Words With Timestral

Bob: Okay, so Mistral just dropped something that isn't a chatbot. It's called Timestral, and it's four foundation models built for time-series data.

Samantha: Which is a bit of a left turn for a lab that made its name on language models.

Bob: Totally. Time series is basically anything where you've got numbers arriving over time. Energy demand, machine sensors, stock prices, hospital admissions. Instead of training a separate model for each one, the idea is you pre-train a big general model and then point it at your data.

Samantha: Same trick as language, really. Learn the general patterns first, then adapt. Whether it actually works for forecasting is a fair question, because plenty of simple statistical methods are stubbornly hard to beat.

Bob: That's the thing I'd watch. Benchmarks will tell us more than the press release. But the strategic move is interesting. Everyone's fighting over who has the best chatbot, and Mistral is saying, hang on, a huge amount of real business value sits in boring operational data.

Samantha: And Europe has a lot of that. Manufacturing, grid operators, logistics, automotive. These are industries where nobody wants to ship their raw sensor feeds to a US cloud if they can avoid it.

Bob: Right, so if the models are open and you can run them yourself, that's a real selling point. A utility could forecast load on its own servers, no data leaving the building.

Samantha: There's also the cost angle. These models don't need to be giant. Forecasting doesn't need a trillion parameters, so you could plausibly run them on modest hardware, which makes them attractive to mid-sized companies, not just the big players.

Bob: Which fits Mistral's whole pitch. Efficient, practical, European. They're not trying to out-spend the American labs on raw compute, they're picking ground where being smaller and open is an advantage.

Samantha: Though I'd be cautious about reading too much into one release. We don't yet know how the four models differ, or how they stack up against existing time-series models from Google and others. Announcing is easy, getting adopted is the hard bit.

Bob: Fair. But even as a signal, it's worth noting. The next wave of AI might be less about chatting and more about quietly predicting when your factory machine is going to fail.

Samantha: Less glamorous, probably more profitable.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.

Google and Synaptics launch a tiny multimodal AI chip for wearables and sensors.

Bob: Google and chipmaker Synaptics have formally launched a new microcontroller aimed squarely at edge devices that have very little power, memory or budget to spare. It combines Google's Coral neural processing unit, Synaptics' Torq neural engine and standard Arm processor cores on a single chip. The target market is wearables, industrial sensors, smart home gear and what the industry calls ambient AI, meaning devices that quietly listen, watch or measure in the background. The pitch is simple: run multimodal AI, so voice and vision together, directly on the device, without round-trips to the cloud.

Samantha: And this matters because today most smart devices either phone home constantly or run very limited on-device models. A hearing aid that understands conversation context, a factory sensor that spots a failing motor from its sound, or a doorbell that recognises a delivery versus a stranger, all of that works far better if the AI sits on the chip itself. Faster response, lower bandwidth, and the data stays local. Google is also signalling that it wants its edge chips to work smoothly with its larger Gemini models in the cloud, so developers can split tasks between the two.

Bob: For European industry this is directly relevant. German and Nordic manufacturers already lead in industrial sensors and smart building gear, and keeping data on the device helps enormously with GDPR compliance. If sensitive audio or video never leaves the sensor, the privacy risk shrinks dramatically. The catch is availability. Google and Synaptics have announced the chip and talked about performance per watt improvements, but concrete pricing, volume availability and which European reference designs will ship first were not detailed in the launch materials.

Samantha: So the headline for European product teams is that the toolbox for on-device AI is getting serious, with named partners rather than just demos. For listeners thinking about smart home or wearable products, expect devices over the next year to advertise local AI as a feature, not a buzzword. It will not replace cloud AI, but it will quietly change what your watch, your doorbell or your hearing aid can do on their own.

Germany's Chancellor Merz pushes for major AI investment ahead of Nvidia partnership talks.

Bob: German Chancellor Friedrich Merz is calling for a sharp increase in artificial intelligence investment in Germany, ahead of expected talks about a potential partnership with Nvidia. Merz framed AI as strategic for the country's technological future, pointing to the global race between the US and China and arguing Germany cannot afford to stay a spectator. The exact scope and funding of a possible Nvidia deal has not been detailed publicly, so for now this is a political signal rather than a signed contract. But the direction of travel is clear.

Samantha: Germany's position matters because it is Europe's largest economy and home to much of its industrial base, from automotive to chemicals to machine tools. Those sectors are exactly where AI could drive the next productivity wave, if the compute, data and talent are in place. An Nvidia partnership would most likely involve dedicated AI compute capacity, possibly through a German sovereign cloud or a national supercomputer upgrade. Nvidia has struck similar flavoured deals with the UK, France and Saudi Arabia over the past year, so the template exists.

Bob: There is a European tension here too. Berlin pushing a bilateral deal with a US chipmaker sits uneasily alongside Brussels' push for European technological sovereignty and homegrown chip capacity. Critics will argue that pouring money into Nvidia hardware deepens the dependency the EU is trying to reduce. Supporters will counter that Europe has no realistic domestic alternative at frontier AI scale right now, and waiting means falling further behind. Both views have merit, and the eventual deal structure, whether it includes local manufacturing or research commitments, will decide which side is more right.

Samantha: For European businesses the practical signal is that national AI budgets are moving from talk to allocation. Expect more announcements like this from Paris, Rome and Warsaw in the coming months, often tied to specific hardware vendors. The question for industry is whether that capacity actually reaches small and mid-sized firms, or stays locked inside a few flagship research projects. Historically in Europe, the answer leans toward the latter, and that is what trade groups are already lobbying to change.

Manuskripte is the free app that turns your Mac into a quiet place for writing long texts, with a focused, minimal workspace.

Bob: Time for two things you can actually try today. First up, for anyone wrestling with a long document, a thesis, a report, a book draft, there is a free writing app called Manuskript worth a look. It is made by an independent developer in Europe, and it runs on the Mac. The core idea is simple: give writers a calm workspace that keeps the screen free of clutter, with no toolbar maze and no full office suite getting in the way. You open it, you write, and the app steps aside.

Samantha: Manuskript is popular with people who write a lot because it keeps the interface minimal. A clean page, a typewriter-style focus, and a distraction-free full-screen mode mean the words stay at the centre. It also lets you work with Markdown, so you can structure headings and lists with a few keystrokes instead of hunting through menus. Users praise the quiet design and the fact that your text stays portable, so you can open it in other editors later. There is nothing to install on a server and no subscription to sign up for.

Bob: The practical pitch for European listeners is that Manuskript is free, works offline, and your draft stays on your own machine unless you decide to move it somewhere else. It is not an AI writing tool and does not try to be. It is a quiet workshop for people who write a lot and want their files to stay their own. Give yourself ten minutes to write a first page, and see if it clicks.

Samantha: It is also light enough to run on older Macs, which is handy if you are not ready to buy a new machine just to write. Reviewers often mention that it strikes a nice balance: enough to handle a serious draft, few enough buttons to stay out of your way.

BikeCitizens is the free cycling app and route planner that builds bike-friendly routes default apps ignore.

Bob: Our second pick is for the cyclists listening, and it is called BikeCitizens. It is a free cycling app and web route planner, built by a European team that is dedicated to cycling in cities. The twist is that it is designed around bikes rather than cars, so instead of pushing you onto the fastest road, it favours cycle paths, quiet streets and calmer alternatives. Mainstream map apps tend to treat cycling as an afterthought. BikeCitizens puts it first.

Samantha: The company is based in Austria and its roots are in urban mobility, so it works best in cities and towns across Europe where cycling infrastructure is well mapped. You enter your start and destination, and it suggests a route along bike lanes and low-traffic streets. You can then follow it with turn-by-turn voice navigation on your phone, which keeps your eyes on the road rather than the screen. The app is free to download for iOS and Android, and there is also a route planner you can try right in your browser.

Bob: Cycling communities often praise it as the planner that finds the calm way through a busy city, rather than the one that sends you down a main road beside the traffic. It is aimed at everyday riders, so it is not a tool for mountain bike trails or long hiking tours. But for commuting, exploring a new city on two wheels or planning a relaxed weekend ride, it is genuinely useful. Install it on the phone, set your destination, and you have a bike-first navigation setup without paying a cent.

Samantha: Today we covered: US Homeland Security bans Chinese AI, Mistral Goes Beyond Words With Timestral, Google and Synaptics launch a tiny, Germany's Chancellor Merz pushes for major, Manuskripte is the free app that, en BikeCitizens is the free cycling app.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech, the tech world in 15 minutes.