Frontier AI Executive Order: Europe's Parallel Review Risk

Samantha: Welcome to The State of Tech, The European Edition, Wednesday June three, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: a White House executive order asking AI labs to hand over their most powerful models before launch, a Chinese factory churning out a humanoid robot every fifteen minutes, Microsoft's quantum chip redesigned by AI, SK Hynix doubling memory capacity to tackle the AI shortage, and to close out, two things you can actually try today: the TRiSM framework for spotting AI security gaps, and Intel's new rackscale AI infrastructure for running agents in your own data centre. Let's start with Washington.

Trump signs executive order asking AI labs to hand over frontier models before public launch.

Samantha: President Donald Trump has signed an executive order titled Promoting Advanced Artificial Intelligence Innovation and Security. It asks AI companies to voluntarily give the federal government access to their most powerful models up to thirty days before public release. The goal: let national security agencies stress-test these systems for cybersecurity and critical infrastructure risks before they hit the market.

Bob: The key word is voluntarily. There's no mandatory licensing, no preclearance, no formal approval gate. What the order does set up is a classified benchmarking process for what it calls covered frontier models. So OpenAI, Anthropic, Google DeepMind and others are being invited to share weights or access with the government for testing, not forced to.

Samantha: And that's a deliberate choice. The White House is framing this as partnership with industry rather than a top-down regulator. The bet is that companies will cooperate because the alternative, down the line, could be much heavier rules.

Bob: It's also a signal to Brussels. The European AI Act already covers general-purpose models with systemic risk, and there are reporting duties baked in. What's different here is the early-access angle: thirty days before launch, in a classified setting. That's closer to how nuclear or biotech research gets reviewed than how software usually ships.

Samantha: For European AI labs and the European subsidiaries of US firms, the practical question is whether models tested under this US process arrive in Europe with a kind of unofficial security stamp. Or whether European regulators will want their own look, creating two parallel review tracks.

Bob: There's also the competitive angle. Chinese labs aren't part of this. So the order effectively builds a Western pre-launch testing club, while Beijing keeps moving on its own track. That widens the geopolitical split in how frontier AI gets governed.

Samantha: The bottom line: this isn't binding regulation, but it sets a precedent. If the biggest labs play along, voluntary today becomes expected tomorrow. And if they don't, that itself becomes the news story.

EngineAI opens Shenzhen factory producing a humanoid robot every fifteen minutes.

Bob: Over to China, where EngineAI Robotics has just opened a new manufacturing base in Shenzhen and started mass deliveries of its T800 humanoid robot. The facility is twelve thousand square metres, designed for up to ten thousand units a year, with one robot rolling off the line every fifteen minutes.

Samantha: That cadence is the headline. We've spent two years talking about humanoid robot prototypes. This is a serial production line, with material inspection, assembly and after-sales maintenance all under one roof.

Bob: It puts EngineAI in a small group of Chinese firms moving from demo videos to actual factory output. The target customers are industrial: warehouses, logistics, light manufacturing, and eventually service roles where labour shortages are biting.

Samantha: For Europe, this matters in two ways. First, availability. If Chinese humanoids start arriving in volume, European industrial buyers get a cheaper option alongside players like Figure, Agility or the Norwegian firm 1X. Second, it puts pressure on European robotics startups to scale faster, because the cost curve is now being set in Shenzhen.

Bob: And there's a regulatory wrinkle. Humanoid robots in workplaces will fall under EU machinery and safety rules, and increasingly under the AI Act when they use general-purpose AI for perception and planning. So importing them isn't just a procurement decision, it's a compliance project.

Samantha: The takeaway: humanoid robotics is shifting from research milestone to industrial supply chain. Whoever masters production at scale gets to set the price.

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.

Microsoft says its AI-designed quantum chip Majorana 2 puts commercial quantum computing within reach by 2029.

Bob: Microsoft has unveiled a new quantum computing chip called Majorana 2, and the interesting detail isn't just quantum, it's how the chip was designed. The company says AI tools handled large parts of the materials science work, which let them use lead as a manufacturing material.

Samantha: Lead is unusual because it's water-soluble, which normally makes it a nightmare in chip fabrication. But it has properties that help with the topological qubits Microsoft is betting on. The reported result: a thousand-fold performance improvement on some parts of the chip.

Bob: Microsoft now expects commercially useful quantum machines by 2029. That puts them roughly in line with IBM's published roadmap. So the two big Western quantum players are converging on the same horizon, which makes that date harder to dismiss as marketing.

Samantha: The AI angle here is the bigger story. Using machine learning to search materials and chip designs is shortening cycles that used to take years of lab work. It's the same pattern we've seen with protein folding and drug discovery, now applied to hardware.

Bob: For Europe, there's a real stake. The EU has its own quantum flagship programmes, and countries like France, Germany and the Netherlands have national champions. The risk is that if Microsoft and IBM hit a commercial product first, European efforts end up as research, not industry.

Samantha: On the other hand, useful quantum machines by 2029 would open up applications in chemistry, logistics and cryptography that European companies have been preparing for. The post-quantum encryption transition, in particular, suddenly looks less abstract.

SK Hynix to double memory chip capacity by 2030 as AI supply crunch deepens.

Bob: SK Hynix is planning to double its memory chip capacity over the next five years. That's a major capital commitment, driven by what SK Group chairman Chey Tae-won describes as an AI memory shortage that could last until 2030.

Samantha: The bottleneck is high-bandwidth memory, the type stacked next to AI accelerators to feed them data fast enough. Every Nvidia, AMD or Broadcom AI chip needs a lot of it, and SK Hynix is the dominant supplier.

Bob: Doubling capacity sounds huge, but the demand side is growing just as fast. Hyperscalers keep raising their AI infrastructure budgets, and now sovereign AI projects in the Gulf, India and Europe are adding to the queue.

Samantha: For European data centre operators, this is genuinely consequential. Memory shortages have been pushing up the cost of AI servers and lengthening delivery times. A clear capacity roadmap from SK Hynix gives buyers something to plan around, even if relief is years away.

Bob: There's also a strategic point. Europe imports nearly all its advanced memory. The EU Chips Act focuses on logic chips and packaging, not memory. So when SK Hynix decides where to invest, that shapes Europe's AI economics from the outside.

Samantha: The bottom line: the AI boom isn't constrained by software or models right now, it's constrained by memory. And the company that controls memory supply controls the pace of the whole industry.

Bob: To close out, two things you can actually look at today. One for the security-minded, one for the infrastructure crowd.

Gartner's TRiSM framework offers a concrete checklist for spotting AI security gaps in your own organisation.

Samantha: Gartner has published a fresh warning that traditional cybersecurity is blind to four fast-moving threats: deepfakes, compromised AI applications, prompt injection, and software supply chain attacks. The useful part for listeners is the framework they recommend alongside it, called TRiSM, which stands for Trust, Risk and Security Management for AI.

Bob: What makes TRiSM worth a look today is that it's not a product, it's a free conceptual checklist. If you work anywhere near IT, security, or even a team that's quietly rolling out ChatGPT or Copilot, you can search for Gartner AI TRiSM and pull up the public summaries this afternoon.

Samantha: The framework breaks AI risk into four buckets: explainability and model monitoring, model operations, AI application security, and privacy. The point is to map every AI tool your organisation uses, including the shadow ones employees brought in themselves, and ask basic questions. Who trained it? What data goes in? Where do outputs go?

Bob: Prompt injection is the one most people underestimate. It's when an attacker hides instructions inside a document or webpage that your AI assistant then reads and obeys. If your team uses an AI agent that browses the web or reads email, that risk is already in your building.

Samantha: Deepfakes are the other practical one. Finance teams in particular are being targeted with fake voice and video calls from supposed executives authorising payments. The mitigation is boring but works: callback verification on a known number, and a second human approver for any transfer above a threshold.

Bob: What you can do today: open the Gartner TRiSM summary, pick the four threat categories, and run through your own AI tools against them. It's a thirty-minute exercise that gives you a real map of where you're exposed. Free, no signup, just a search away.

Samantha: And if you're a smaller business without a security team, the same framework works as a conversation starter with your IT provider. Ask them which of the four TRiSM areas they cover, and which they don't.

Intel's new rackscale AI infrastructure lets companies run AI agents in their own data centre instead of the cloud.

Bob: And to wrap up, something for anyone watching their cloud bill. At Computex this week Intel unveiled new rackscale AI infrastructure, basically pre-built racks of servers tuned for running AI inference and agent workloads on your own premises.

Samantha: The setup combines Intel's Xeon processors with accelerator chips from SambaNova, and pairs them with the next-generation Xeon 6+ processors aimed at agent-style AI. The pitch is straightforward: if you're running AI agents at scale, doing it in your own data centre can be cheaper and more controllable than renting cloud GPUs.

Bob: For European companies, the appeal is partly about cost, partly about data sovereignty. Running customer data through an in-house rack avoids a lot of GDPR friction that comes with sending it to a hyperscaler in another region.

Samantha: What you can actually do today: Intel has published configuration guides and reference designs on its developer site. If you're an IT lead or architect, you can pull those down, see the rack layouts, and compare the total cost against your current cloud spend. Free to download, no sales call required.

Bob: It's also useful as a benchmarking exercise. Even if you don't buy Intel hardware, the reference designs tell you how much memory, networking and power a serious AI agent setup actually needs. That helps you ask sharper questions of any vendor.

Samantha: The broader shift is that AI infrastructure is moving back on-premises for specific workloads. The cloud isn't going away, but for steady, predictable inference jobs, owning the hardware is starting to make financial sense again.

Samantha: Today we covered: Trump's executive order on frontier AI model access, EngineAI's humanoid robot factory in Shenzhen, Microsoft's AI-designed Majorana 2 quantum chip, SK Hynix doubling memory capacity, and two things to try yourself: Gartner's TRiSM framework for AI security, and Intel's rackscale AI infrastructure reference designs.

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.