Google TPU 8 takes on Nvidia, Pentagon pressures Anthropic, Hinton warns UN on AI

Samantha: Welcome to The State of Tech, The European Edition, Thursday April twenty-third, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: Google's new TPU 8 chips taking aim at Nvidia, a turbulent day in global AI policy, Geoffrey Hinton sounding the alarm at the UN, a Chinese firm promising home robots in 35 days, Microsoft's massive Australia investment, and a new survey showing tech freelancers split on AI. Let's start with Google's silicon push.

Google goes after Nvidia's crown.

Samantha: Google has unveiled its TPU 8 generation, and this is the most direct shot at Nvidia we've seen from Mountain View. Two variants: the TPU 8T for training, the TPU 8I for inference. That split matters — Google is now covering the full AI workload cycle with its own silicon.

Bob: And the split is strategic. Training chips need raw power and memory bandwidth. Inference chips need efficiency and low latency because they run live services for billions of users. By separating them, Google can optimise each for its actual job instead of selling one chip that kind of does both.

Samantha: Google Cloud customers get first access, and that's the commercial lever. Google is saying: come to us, and you won't pay the Nvidia premium. Nvidia's H-series and Blackwell chips have been the bottleneck of the entire AI economy for two years now.

Bob: The numbers behind that bottleneck are wild. Nvidia margins on data centre GPUs have been above seventy percent. Every hyperscaler — Amazon with Trainium, Microsoft with Maia, now Google with TPU 8 — is trying to claw that back. Not because they'll replace Nvidia fully, but because even shifting twenty percent of workloads in-house saves billions.

Samantha: And compute costs determine who gets to build frontier models. If TPUs genuinely deliver better price-performance for inference, that trickles down to API pricing, which trickles down to the apps European startups can afford to run.

Bob: That's the real story. An AI startup in Berlin or Paris isn't buying chips — they're renting compute by the token. If Google undercuts on Gemini inference because TPU 8I is cheaper to run, Anthropic and OpenAI have to match it. Price war at the infrastructure layer.

Samantha: For Europe there's the usual tension though. Cheaper American cloud AI is great for European enterprises in the short term, but it deepens dependence on US hyperscalers. The EU's sovereign cloud ambitions — Gaia-X, the whole push for European alternatives — don't get easier when Google Cloud keeps getting cheaper and faster.

Bob: And European chip efforts remain modest. SiPearl, the work around the European Processor Initiative — promising but nowhere near TPU 8 scale. Competition between Google and Nvidia helps European buyers. It doesn't help European builders.

AI policy gets messy on both sides of the Atlantic.

Bob: Onto a tangled policy day. In Washington, the House Foreign Affairs Committee advanced the bipartisan MATCH Act, tightening semiconductor equipment exports to adversaries. And in parallel, the Pentagon reportedly labelled Anthropic a security risk over its AI kill switch policies, threatening to blacklist the company from military contracts.

Samantha: The Anthropic story is fascinating because it flips the usual narrative. Anthropic has built its brand on safety — constitutional AI, responsible scaling policies, the whole Dario Amodei message. Now the Pentagon is saying: your safety features are the problem, because we can't have a model that refuses orders.

Bob: A kill switch in this context means the ability to shut down or constrain a model's behaviour in specific scenarios. The Pentagon wants AI it can fully control. Anthropic wants AI it can constrain on ethical grounds. Those two positions are not compatible when the customer is the Department of Defense.

Samantha: And meanwhile Google launched Workspace Intelligence, Gemini-powered, going directly at Microsoft Copilot in enterprise. So while governments tighten the screws, commercial AI keeps accelerating. That split is the defining tension of 2026.

Bob: The MATCH Act matters because semiconductor equipment — not just chips, but the machines that make chips — is where the real chokepoint sits. ASML in the Netherlands, Applied Materials, Tokyo Electron. If the US tightens export rules, European equipment makers get caught in the middle again.

Samantha: ASML has been here before. Every round of export controls means new licences, new compliance costs, and lost Chinese revenue. The Dutch government has pushed back, but mostly ends up aligning with Washington.

Bob: For European regulators, the Anthropic episode is a credibility test. The EU AI Act explicitly requires safeguards on high-risk systems. If the US military pressures firms to strip those safeguards, European rules become a refuge for safety-focused AI — or an obstacle, depending on who you ask.

Samantha: And European defence ministries will face the same dilemma. Do they want AI they can fully control, or AI with ethical guardrails built in? That debate hasn't really started publicly yet in Brussels.

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.

The godfather of AI wants a steering wheel.

Samantha: Geoffrey Hinton, Nobel laureate and the man who largely invented modern deep learning, spoke at the UN's Digital World Conference today. His line: AI is a very fast car with no steering wheel. It must be steered to serve societies, not undermine them.

Bob: Hinton left Google in 2023 specifically to speak freely about AI risks. Every intervention since has pushed harder on regulation. What's new today is the venue — the UN — and the timing. A UN Scientific Panel report feeds into the Global Dialogue on AI Governance in July, with all 193 member states at the table.

Samantha: This is the closest thing to a global AI framework we've seen. The Global Dialogue isn't binding, but it sets norms. And norms matter — they become the reference point for national legislation. Hinton is trying to make sure those norms have teeth.

Bob: His specific worry is autonomy. Models that set their own sub-goals, that optimise for outcomes humans can't fully inspect. He's been consistent: the path from today's frontier models to systems that act without meaningful human oversight is shorter than people realise.

Samantha: And there's a real gap between people who work on AI capabilities and people who work on AI alignment. Hinton keeps pointing out that the capability side is funded at hundreds of billions, while safety research is a rounding error. That imbalance doesn't fix itself.

Bob: Industry will argue self-regulation works. Hinton's point is that you don't let car manufacturers decide if seat belts are necessary. You legislate. And the UN track is where that legislation gets its intellectual grounding.

Samantha: For Europe, Hinton is basically endorsing the AI Act's philosophy. The EU has been criticised for regulating too early, but Hinton's argument is: too early is better than too late. Brussels gets to say: we told you so.

Bob: The harder question is enforcement. The AI Act has teeth on paper — fines up to seven percent of global turnover. In practice, national regulators are understaffed and the technical complexity is enormous. Hinton's warnings don't solve that, but they give regulators political cover to be tougher.

Home robots in 35 days. Really.

Bob: X Square Robot — a Chinese firm backed by Alibaba and ByteDance — has unveiled Wall-B, an embodied AI foundation model for home robots. The headline claim: deployment in everyday households within 35 days.

Samantha: That timeline is aggressive enough to be either a genuine breakthrough or a marketing moment. Embodied AI has been the hardest problem in robotics — factory robots do one thing in controlled conditions; homes are chaos. Toys on the floor, pets, kids, unpredictable lighting.

Bob: Wall-B is the first full implementation of what X Square calls the World Unified Model — vision, language, action, and physical prediction in one system. Physical prediction is the interesting bit. The robot simulates what happens next before it moves. That's how you pour a drink without knocking over the glass.

Samantha: Tesla's Optimus, Figure, 1X, Unitree — everyone is chasing humanoid or semi-humanoid home robots. What separates them now is the model, not the hardware. The bodies are converging. The brains are where the competition sits.

Bob: If X Square actually ships in 35 days, even to a limited pilot, that's a milestone. Consumer electronics robots like Roomba are narrow-function. Wall-B is being positioned as general-purpose, which is a completely different category.

Samantha: And the Alibaba-ByteDance backing tells you about market access. Those platforms put hundreds of millions of Chinese consumers in reach overnight. Scale at home first, then export.

Bob: Europe has strong industrial robotics — KUKA, ABB, Universal Robots in Denmark. Consumer embodied AI? Basically nothing at X Square's scale. Another category where Europe watches from the sidelines.

Samantha: And regulators haven't written rules for autonomous robots in private homes. Privacy — these things have cameras and microphones constantly. Liability — if the robot breaks something, who pays? Safety — what if it misjudges around a child? GDPR covers some of it. Most of it is uncharted.

Microsoft drops 25 billion Australian dollars on Australia.

Samantha: Microsoft has announced its largest ever global technology investment — 25 billion Australian dollars, about 18 billion US — committed to Australia by the end of 2029. Digital infrastructure, cyber defence, and workforce skilling. And a pledge to train three million Australians in AI by the end of 2028.

Bob: Three million is roughly one in nine Australians. That's not a training programme — that's national infrastructure. And it triples a previous Microsoft commitment, which tells you how quickly the AI skills gap is being priced into strategy.

Samantha: The investment is backed by a memorandum of understanding with the Australian government, aligned with the National AI Plan. That's the template — tech giant plus sovereign government, big number, workforce component to make it politically defensible.

Bob: The infrastructure piece is data centres and cyber. Australia is a strategic location for US tech firms — stable, English-speaking, Five Eyes partner, decent power supply. Cloud capacity there serves all of Asia-Pacific.

Samantha: You get your infrastructure built, your data residency solved, and a pipeline of trained workers. In exchange, the host country gets jobs, investment, and AI capacity. Both sides need each other.

Bob: And the skilling number is the part that defuses political backlash. Governments are under pressure about AI job losses. Microsoft saying we'll train three million people is a direct answer to that pressure.

Samantha: Europe has fragments of this — Microsoft investments in Germany, France, Spain. But no single country gets a 25 billion package, because no single European country has the scale or strategic clarity of Australia's national plan.

Bob: That's the argument for EU-level coordination. Twenty-seven separate deals mean twenty-seven smaller commitments. A unified European AI infrastructure and skilling deal could easily top this. Whether Brussels can pull that off politically is another matter.

Freelancers can't agree on whether AI is friend or foe.

Bob: Closing out with something more human. A new survey from The Accountancy Partnership polled over a thousand tech freelancers. The split: 43.1 percent think AI will positively influence their industry. 36.3 percent fear job displacement. And creative freelancers are the most pessimistic — 43.3 percent expect negative effects.

Samantha: That creative pessimism tracks with reality. Image generation, copywriting, video editing — those tools have genuinely changed the market. Rates for routine creative work have dropped. Rates for high-end creative direction have held up. AI flattens the middle.

Bob: The split is almost 50/50 optimist to pessimist, with a nervous middle. What's interesting is that freelancers who've actively integrated AI tools tend to be in the positive camp. The ones who haven't, aren't.

Samantha: Adoption equals confidence. Use the tools, see they extend your output, feel less threatened. Avoid the tools, watch clients ask about them in briefs, feel the ground shift.

Bob: Across the EU, self-employed and freelance workers are tens of millions of people. If a third of them genuinely fear displacement, that's a real political and economic signal. Especially in creative hubs — London, Berlin, Amsterdam, Lisbon.

Samantha: The encouraging reading: AI literacy and prompt engineering training change outcomes. Freelancers who upskill charge more, not less. The discouraging reading: upskilling costs time and money freelancers often don't have.

Bob: Most EU training programmes and subsidies assume an employer-employee relationship. Freelancers fall through the gaps. If Europe wants its gig workforce AI-ready, that infrastructure needs rebuilding for people who don't have an HR department.

Samantha: And the cultural bit — freelancers are individualists by definition. They don't join big training programmes. Any solution has to be bite-sized, on-demand, and genuinely useful from day one. That's a design challenge, not just a funding one.

Bob: And a nice note to end on: the freelancers who are thriving are the ones treating AI as a colleague, not a competitor.

Samantha: Today we covered: Google's TPU 8 chips taking on Nvidia, US export controls and the Pentagon versus Anthropic, Hinton's UN warning on AI regulation, China's Wall-B home robot reveal, Microsoft's 25 billion dollar Australia deal, and tech freelancers split on AI's impact.

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.