DeepMind quits over military AI; US eyes outbound AI curbs

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

Bob: And I'm Bob Russell. Today: a Google DeepMind engineer walks out over military AI contracts, global AI firms pile into Indian data centres, Washington tightens the screws on overseas AI investment, and a sub-twenty-five-hundred-dollar humanoid robot lands on the market. Plus two things you can try yourself: the free open-source map tool Qwant Maps alternative GraphHopper Routing, and the privacy-friendly reading app Omnivore. Let's start with that resignation at DeepMind.

Google DeepMind engineer quits in protest over military AI contracts, echoing Project Nimbus.

Samantha: A research engineer at Google DeepMind, Sofia Liguori, has resigned and gone public about it. She says she can no longer work for a company that is deepening its military AI partnerships, and she accuses the internal culture of discouraging employees from pushing back on how the technology gets deployed. Liguori spent more than three years at DeepMind. The trigger, according to her, is a classified agreement between Google and the US Department of War. She draws a direct line to Project Nimbus, the earlier cloud contract that caused so much friction with staff.

Bob: And this is not an isolated moment. We have seen a pattern now: engineers at Google, Microsoft, and Amazon raising objections to defence work, sometimes staging walk-outs, sometimes leaving quietly. What makes Liguori's exit different is that she is naming the problem publicly, and she is naming it at DeepMind, which has long cultivated an image as the research arm with stricter ethical guardrails. Her central claim is about process: that there is no real mechanism for staff to challenge how a model ends up being used once a contract is signed. That is a governance question, not just a values question.

Samantha: The context matters too. The US Department of War, the renamed Department of Defense, has been accelerating AI procurement across targeting, intelligence analysis, and logistics. Classified contracts by definition keep oversight thin. Civil society groups argue that if even the researchers building the models cannot see where they end up, external accountability becomes almost impossible. Google has not detailed the scope of the agreement Liguori refers to, and we should be careful: her description is her own account, not a confirmed company statement.

Bob: For Europe this lands on sensitive ground. European research institutes, universities, and startups collaborate extensively with DeepMind and Google Research. If a growing share of the parent company's work flows into classified military programmes, European partners will face pressure to clarify what their joint projects feed into. The EU AI Act carves out military uses, but dual-use research sits in a grey zone. Expect fresh calls in Brussels and in national parliaments for stricter transparency rules on cross-border AI collaborations, especially where public money is involved.

Samantha: The bottom line: one resignation does not change a contract, but it does change the conversation. And it puts every European lab working with Google DeepMind in the position of having to answer uncomfortable questions from their own staff.

Global AI firms race to India for data centre capacity as power constraints bite elsewhere.

Bob: Our second story. Global AI companies are increasingly turning to India for data centre capacity, and not just to serve Indian users. Operators there say international clients are using Indian facilities to serve customers worldwide. The driver is simple: power and permits. In the US, in Ireland, in the Netherlands, grid connections for new hyperscale sites have multi-year waiting lists. India, with its growing renewables build-out and faster permitting in states like Maharashtra and Tamil Nadu, has become an available alternative.

Samantha: Operators report that conversations often start at twenty megawatts of capacity and scale up quickly from there. For context, twenty megawatts is roughly the electricity use of a small town, and it is now the opening bid for a single AI customer. That is a shift. Two years ago those numbers would have described an entire mid-sized data centre campus. This is the physical footprint of generative AI showing up in infrastructure contracts.

Bob: SAMANTHA LAWRENCE: 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.

Samantha: For Europe this is a double-edged story. On one hand, European AI startups that cannot secure capacity at home now have a credible overseas option without going straight to the US. On the other hand, every megawatt that lands in Hyderabad instead of Frankfurt is a megawatt of economic activity, tax base, and strategic autonomy that Europe loses. Brussels has talked a lot about sovereign AI infrastructure. This story is a reminder that talk does not pour concrete.

Bob: There is also a data governance angle Europeans need to think through. If a French AI startup trains its models on Indian soil, which data protection regime applies to the training data? GDPR follows the data subject, not the server, but enforcement across jurisdictions is messy. Operators are moving faster than regulators here, which is the familiar pattern. The practical takeaway: expect India to become a genuine third pole in AI compute alongside the US and China, with Europe watching from the sidelines unless the permitting situation changes fast.

US Treasury proposes sweeping new rules restricting AI and supercomputing investments abroad.

Samantha: Third story, and it is a regulatory one with real teeth. The US Department of the Treasury has proposed new rules restricting American investment in artificial intelligence, supercomputing, hypersonic systems, and biotechnology in countries designated as of concern, principally China and Russia. This expands on restrictions first introduced in 2023. The goal, as Treasury frames it, is to stop US capital and know-how from strengthening the military capabilities of rival states.

Bob: The mechanism matters. These are outbound investment controls, which is a relatively new tool in the American economic policy kit. Export controls stop products from leaving. Outbound controls stop money and expertise from flowing in. The proposed rule would cover venture capital, private equity, and joint ventures. For American investors it means due-diligence paperwork and, in some cases, outright prohibition. For target countries it means a narrower pool of foreign capital for their frontier tech sectors.

Samantha: The European angle is uncomfortable. Washington has been pushing allies to adopt parallel rules. The Netherlands, Germany, and France are already in conversations about their own outbound investment screening frameworks. If the EU moves in the same direction, European pension funds and sovereign wealth vehicles with exposure to Chinese AI firms will face new disclosure and possibly divestment requirements. If the EU does not move, European investors could become the preferred backchannel, which Washington will not tolerate quietly.

Bob: There is also a competitive dimension. Some European AI startups quietly welcome tighter US rules because it reduces the capital flowing to their Chinese rivals. Others worry about retaliation: China has its own list of countermeasures, and European firms with operations in Shanghai or Shenzhen could get caught in the crossfire. The rule is still a proposal, with a consultation period ahead, so the final shape is not fixed. But the direction of travel is clear, and European policymakers will have to pick a lane within months, not years.

Faraday Future launches a humanoid robot for under twenty-five hundred dollars.

Samantha: Our fourth news story, and we are shifting from policy to hardware. Faraday Future Intelligent Electric, through its robotics subsidiary, has launched FF Nova, a humanoid robot priced under twenty-five hundred dollars in the US. That price point is the news. Most humanoid robots announced over the past two years, from Figure to Tesla's Optimus, carry price tags in the tens of thousands or are not for sale at all. Faraday Future is undercutting the market by an order of magnitude.

Bob: There are big caveats. Faraday Future has a long history of ambitious announcements that have not always matched delivery, and the detailed capabilities of FF Nova remain thin. The company also introduced a smaller model called FF Nova Mini, a curriculum programme for schools, and a developer programme for reinforcement learning. That suggests the strategy is less about selling a working domestic butler and more about building a developer ecosystem around a cheap platform. Think of it as the Raspberry Pi approach to humanoid robots: get the hardware into enough hands that software catches up.

Samantha: For Europe this matters because affordable hardware changes who gets to experiment. Universities, vocational schools, and small robotics startups across Germany, Italy, and the Nordics have been priced out of serious humanoid research. A sub-twenty-five-hundred-dollar unit, even one with limitations, makes classroom and workshop experiments feasible. European robotics clusters in cities like Pisa, Munich, and Delft could get a boost in research output purely because the entry barrier drops.

Bob: The flip side is a safety and standards question. Europe has strict rules on machinery, including robots operating in homes and workplaces. A cheap humanoid imported from a US consumer brand will have to meet CE marking and the Machinery Regulation, which came into force in 2027. Expect regulators to look closely at how Faraday Future handles liability for a robot that falls, pinches, or misinterprets an instruction. Cheap is only useful if it is also safe.

GraphHopper Routing turns any browser into a free, privacy-friendly route planner for cars, bikes and walks across Europe.

Samantha: To close out, two things you can actually use today, and the first is a free routing tool worth bookmarking. GraphHopper Routing is an open-source European route planner you can use directly in a browser at graphhopper dot com. It plans routes for cars, bikes, walking, and even trucks across Europe, using OpenStreetMap data. There is no account, no tracking cookies by default, and no advertising clutter.

Bob: What makes it genuinely useful is the detail. For cycling it offers routes that favour quiet roads and dedicated bike lanes, which the usual mainstream apps often ignore in favour of the fastest line. For walking it respects pedestrian paths and footbridges that car-focused services skip. You can drag the route to force detours, compare travel times across modes, and export the result as a GPX file for your bike computer or hiking GPS. For someone planning a weekend ride through the Ardennes or a city walk through Prague, that export function alone is worth it.

Samantha: Users report that it loads fast even on slow connections and that the European coverage is unusually good for a free service. The project is run from Germany and the underlying code is open-source, which means the routing logic is auditable. For context, GraphHopper also licenses its engine to commercial customers, which funds the free public version. That is a sustainable model, not a loss leader waiting to be shut down.

Bob: Two practical tips. First, if you need offline maps for a trip without reception, GraphHopper itself is a web tool, but you can combine it with an offline maps app on your phone by exporting the GPX file. Second, the walking profile has a scenic option worth trying if you have time to spare. It avoids busy roads even at the cost of a longer route, which is exactly what you want on a weekend. It works on any browser, on desktop or mobile, and costs nothing.

Samantha: Give it a try next time you are planning anything more interesting than a direct commute. It is one of those quiet European tools that does exactly what it promises.

Omnivore is the free open-source read-later app that saves articles for calm, ad-free reading.

Bob: Our second thing to try is for anyone drowning in open browser tabs. Omnivore is a free, open-source read-later app. You install a browser extension or share articles from your phone, and it strips out the ads, the pop-ups, and the cookie banners, leaving you with just the text and images. You can then read on the web, on iOS, or on Android, and highlight passages as you go.

Samantha: The main appeal is calm. Modern news sites are visually hostile: autoplaying videos, sticky headers, and newsletter prompts. Omnivore reduces an article to its reading core. For listeners who commute on a train and want to catch up on long reads without eating mobile data, there is also an offline mode. Articles sync down once, and you can read them without reception.

Bob: There are two features worth highlighting. First, PDF support. You can save PDFs the same way you save articles, which turns Omnivore into a lightweight research tool for students and anyone working through white papers. Second, newsletter forwarding. You get a dedicated email address, and newsletters sent there appear as clean articles instead of cluttering your inbox. For people who subscribe to Substacks or industry newsletters, that is a quietly powerful workflow.

Samantha: The project is open-source, which means the code is auditable and there is no advertising business model behind it. Users praise the clean interface and the speed. It is not the only app in this category, but it is one of the few that is genuinely free, actively maintained, and works across every major platform. Setup takes about five minutes: install the extension, create an account, and start saving.

Bob: Today we covered: a Google DeepMind engineer resigning over military AI contracts, global AI companies flocking to Indian data centres, the US Treasury proposing new overseas investment restrictions, Faraday Future's sub-twenty-five-hundred-dollar humanoid robot, and two things to try yourself: GraphHopper Routing and Omnivore.

Samantha: Today we covered: Sofia Liguori's resignation from Google DeepMind over military AI, the global rush for Indian data centre capacity, Washington's proposed curbs on overseas AI investment, Faraday Future's cheap humanoid FF Nova, and two things to try yourself: GraphHopper Routing and Omnivore.

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.