Musk's Terafab Intel Deal, GPT-5.5, SK Hynix AI Chips & Anthropic Europe Expansion

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

Bob: And I'm Bob Russell. Today: Elon Musk and Intel team up on a massive AI chip venture called Terafab, Chinese AI lab DeepSeek unveils its V4 models with a twenty-billion-dollar valuation in sight, OpenAI puts a number on how many US jobs face near-term automation, Europe's humanoid robot rollout stalls on regulation, Anthropic hunts for European data centre capacity, and OpenAI and Microsoft tighten their cybersecurity alliance. Let's start with Musk and Intel.

Musk teams with Intel on a trillion-watt chip dream.

Samantha: Tesla, SpaceX and xAI are forming a joint venture called Terafab, and the ambition here is genuinely eye-watering. They want to produce AI chips on Intel's 14A process, with a facility capable of one million wafers a month, targeting one terawatt of AI compute annually.

Bob: One terawatt. That's the number that's making chip analysts spit out their coffee this morning. For context, that is roughly an order of magnitude beyond what today's largest AI clusters pull combined.

Samantha: The structure is interesting. Tesla is leading a three-billion-dollar research fab in Austin. SpaceX is spearheading the broader Terafab project. Intel brings the chip design, fabrication, and packaging expertise. So Musk gets the volume and the vertical integration, Intel gets a flagship customer for 14A.

Bob: And Intel badly needs that anchor tenant. The 14A node is their shot at clawing back process leadership from TSMC, and until now the list of committed customers was thin. Musk writing cheques at this scale changes the narrative overnight.

Samantha: The strategic signal is that Musk no longer wants to depend on Nvidia for Tesla's self-driving fleet, Optimus robots, or xAI's Grok training runs. He wants custom silicon, at scale, owned end to end.

Bob: Which follows the Amazon, Google, Meta playbook, but pushed to an extreme. Those hyperscalers design chips and outsource fabrication. Musk is going after the fab itself. If it works, it's a template. If it doesn't, it's the most expensive lesson in semiconductor physics ever written.

Samantha: For Europe, the most obvious beneficiary is ASML. A fab at this scale needs extreme ultraviolet lithography machines, and ASML in Veldhoven is the only company on the planet that makes them. More Terafab demand means more orders for Dutch industry.

Bob: The flip side is less comfortable. European cloud providers like OVHcloud, Scaleway, and IONOS already struggle for GPU allocation. If Musk locks up a terawatt of custom compute for his own businesses, the global supply of leading-edge AI silicon gets tighter, not looser. European AI developers should assume access stays constrained through 2027.

Samantha: And there's a policy angle. The EU Chips Act was meant to bring twenty percent of global semiconductor production to Europe by 2030. Projects like Terafab, concentrated in Texas, make that target harder to hit.

DeepSeek V4 lands with a million-word context window.

Bob: Moving to China. DeepSeek has launched V4 Flash and V4 Pro, and the headline spec is an ultra-long context of one million words. The company is in talks with Tencent and Alibaba to raise fresh funding at a valuation north of twenty billion dollars.

Samantha: A one-million-word context puts DeepSeek in direct competition with Google's Gemini and Anthropic's Claude on long-document tasks. Think entire legal archives, full codebases, multi-book analysis in a single prompt. And DeepSeek is still positioning itself as open-source friendly, which Western labs largely are not.

Bob: That combination, frontier capability plus open weights, is what made DeepSeek a shock story in early 2025. They're doubling down. The twenty-billion valuation, if it closes, would make them the most valuable Chinese AI startup by a wide margin.

Samantha: For a long time the narrative was US versus China, with China catching up. DeepSeek's trajectory suggests parity on specific benchmarks, and leadership on cost efficiency. They train models for a fraction of what OpenAI spends.

Bob: And Tencent and Alibaba joining the cap table matters. It signals the Chinese tech establishment consolidating around a national AI champion, with state-adjacent capital behind it.

Samantha: For European developers, DeepSeek's open-source stance is attractive. You can self-host, you can audit the weights, you can avoid US export licence risk. That's appealing under the AI Act's transparency requirements.

Bob: But the trust question is unavoidable. Using a Chinese model for sensitive European workloads raises data residency, supply-chain, and geopolitical concerns regulators in Brussels and Berlin are actively debating. The EU hasn't banned Chinese AI models, but procurement guidance for public sector use is tightening.

Samantha: And it puts pressure on Mistral, Aleph Alpha, and the European champions. If DeepSeek is free, multilingual, and competitive on benchmarks, the business case for a paid European alternative needs to rest on sovereignty and compliance, not just performance.

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.

OpenAI puts eighteen percent on the automation risk line.

Bob: OpenAI has published an economic research report called The AI Jobs Transition Framework. The headline number: eighteen percent of US jobs face relatively higher short-term automation risk from AI. Forty-six percent are likely to see less immediate change.

Samantha: The methodology is what's novel. OpenAI's economists looked at 921 occupations using a multi-dimensional approach. Technical capability of current models, plus observed actual ChatGPT usage patterns. So it's not just theoretical, it's grounded in who's already using these tools and for what.

Bob: The insulated categories are predictable but worth naming. Teaching, healthcare, roles requiring physical presence, regulated professions, anything built on trust relationships. The exposed categories skew toward clerical, administrative, and routine knowledge work.

Samantha: There is something unusual about OpenAI publishing this. The company building the technology is quantifying its own labour market impact. Sceptics will say the eighteen percent number is conservative by design. Optimists will say it's the first honest internal estimate from a frontier lab.

Bob: Either way, it's now a reference point. Policymakers, unions, HR departments will cite this number. And the distinction between short-term automation risk and longer-term transformation is useful. Not every exposed job disappears; many just change shape.

Samantha: For European governments, the methodology is more useful than the US-specific numbers. Applying this framework to French, German, or Dutch labour markets would give ministries something concrete to build retraining programmes around.

Bob: Europe's starting position is different from the US. Stronger social safety nets, more active labour market policies, higher unionisation in exposed sectors. That cushions the transition, but it also slows adaptation. Works councils in Germany will push back hard on AI deployment that looks like headcount reduction.

Samantha: And the AI Act already requires employers to inform workers about AI systems used in HR and management. Reports like this one give unions the data to negotiate harder on how AI gets deployed, not just whether.

Europe's humanoid robots stuck in regulatory limbo.

Samantha: Fourth story. Europe's progress on humanoid robots is stalling, and the reason isn't technology. It's regulation. The Fraunhofer Institute for Manufacturing Engineering and Automation in Germany is flagging that safety rules and liability frameworks for humanoids simply don't exist yet at EU level.

Bob: Fraunhofer IPA is usually the bridge between academic robotics and German industry. When they say regulation is the bottleneck, that's not a complaint from researchers, it's a signal from the people whose job is getting this tech deployed.

Samantha: The specific gaps are around liability when a humanoid causes harm, certification standards for autonomous movement in shared human spaces, and insurance frameworks. The existing Machinery Regulation covers industrial robots in cages. It doesn't cleanly cover a two-legged machine walking around a warehouse with human colleagues.

Bob: Meanwhile, US companies like Figure and Agility are piloting humanoids with BMW, Mercedes, and Amazon. Chinese manufacturers Unitree and UBTECH are flooding the market with cheaper units. European robotics firms like Neura in Munich have the technology, but face a slower approval path to commercial deployment.

Samantha: So the risk is that Europe ends up importing humanoids certified under foreign standards, rather than setting its own. That's the same pattern we saw with consumer drones a decade ago.

Bob: This is the classic European trade-off. Careful regulation protects workers and consumers, but the clock is ticking. If Brussels takes three years to draft humanoid-specific rules, the market will already have standardised around non-European specs.

Samantha: And there's a worker protection angle. Unions want clarity on liability when a humanoid injures someone. Employers want clarity before they invest millions. Right now nobody gets what they need, and deployment stalls for everyone.

Anthropic wants European data centre space, fast.

Bob: Story five. Anthropic is scaling up its European hunt for data centre capacity. They're recruiting a transaction principal to negotiate deals in Frankfurt, London, Amsterdam, Paris, Dublin, and Nordic and Southern European markets.

Samantha: The context is the estimated six-hundred-billion-dollar AI infrastructure budget of major US tech companies for 2026. Anthropic is competing with Microsoft, Google, Meta, and Amazon for the same power and land. Europe is now a strategic theatre, not an afterthought.

Bob: The city list tells you everything about European data centre geography. Frankfurt for Germany and financial connectivity, London for the UK market, Amsterdam for the Dutch fibre hub, Dublin for tax and hyperscale density, Paris for France, plus the Nordics for cheap hydropower and cool climate.

Samantha: Hosting Claude inference in Europe means lower latency for European customers and, crucially, easier compliance with GDPR and the AI Act's data residency expectations. For enterprise customers in banking, healthcare, and the public sector, that's the difference between procurement approval and rejection.

Bob: And Anthropic is clearly targeting enterprise over consumer. OpenAI has ChatGPT brand recognition. Anthropic's bet is on deep integration with European corporations that need sovereign, auditable AI. Building local infrastructure is table stakes for that pitch.

Samantha: The upside is tangible. High-skill jobs in data centre operations, construction investment, tax revenue for host cities. Dublin and Amsterdam have seen what this looks like with Microsoft and Google.

Bob: The downsides are equally tangible. Ireland's grid is already strained by existing data centres. The Netherlands has paused new-build permits in parts of the country. Energy consumption and water use for cooling are genuine political issues now.

Samantha: And the sovereignty paradox remains. Anthropic in Frankfurt is still Anthropic, a US company subject to US law. Putting the servers on European soil helps, but doesn't resolve the fundamental dependency question.

OpenAI and Microsoft tighten the cybersecurity knot.

Bob: Final serious story before we close out lighter. OpenAI and Microsoft have expanded their cybersecurity partnership under a programme called Trusted Access for Cyber. Microsoft gains access to OpenAI's most cyber-capable models. In return, Microsoft hardens OpenAI's systems using its Secure Future Initiative infrastructure.

Samantha: It's a symmetrical arrangement. OpenAI gets Microsoft's global telemetry, threat intelligence, and security engineering. Microsoft gets frontier AI models tuned for defensive cyber operations, feeding into products like Defender and Sentinel.

Bob: The threat backdrop justifies it. Ransomware crews, state actors, and increasingly AI-augmented attackers are the baseline now. Phishing generation, vulnerability discovery, and social engineering are all being industrialised with language models on the offensive side.

Samantha: So the defence side consolidates. Microsoft already dominates enterprise endpoint security. Adding OpenAI's models deepens that moat considerably. For Chief Information Security Officers, it makes the Microsoft security stack even harder to avoid.

Bob: Which is great for protection and complicated for competition. Smaller European cybersecurity vendors like WithSecure or ESET face a widening capability gap they cannot close on their own.

Samantha: European organisations running Microsoft 365 and Azure, which is most of them, get upgraded protection largely for free. That's a genuine win against ransomware gangs targeting hospitals and municipalities.

Bob: The dependency cost is that European critical infrastructure is now defended by a US-US alliance. NIS2 and DORA push European firms toward rigorous cybersecurity, but the tools come from Redmond and San Francisco. European cyber champions exist, but none operates at this scale.

From serious to slightly absurd.

Samantha: And to close, something lighter. Think back to Musk's terawatt ambition at the top of the show. A terawatt is roughly the average electricity demand of the entire United Kingdom. For chips. To make chatbots smarter.

Bob: And a million wafers a month. Stacked, that's a tower taller than most European skyscrapers, made of silicon discs, appearing every thirty days.

Samantha: Meanwhile in Europe, we're still debating whether a humanoid robot is allowed to hand you a coffee in a factory canteen without a liability framework. The contrast is almost cinematic.

Bob: It captures the week perfectly. America builds terawatt fabs. China ships million-word context models for cheap. Europe writes careful rules about who is responsible when the robot spills the latte.

Samantha: And honestly, someone has to write those rules. When the first humanoid does spill coffee on a worker's laptop, everyone will be glad someone thought about liability first.

Bob: Fair. Careful isn't always slow. Sometimes it's just grown-up.

Samantha: Today we covered: Musk and Intel's Terafab mega-venture, DeepSeek's V4 launch and twenty-billion valuation, OpenAI's eighteen percent automation report, Europe's humanoid robot regulatory gap, Anthropic's European data centre push, and the expanded OpenAI-Microsoft cybersecurity alliance.

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.