EU AI Act Enforcement August 2026: What Companies Must Do

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

Bob: And I'm Bob Russell. Today: Meta's eye-watering AI spending spooks investors, SoftBank prepares a hundred-billion-dollar AI and robotics IPO, key AI regulation deadlines loom on both sides of the Atlantic, OpenAI plots a smartphone for 2028, Washington tightens the screws on China's chip industry, and Amazon may sell its Trainium AI chips to outside customers. Let's start with Meta.

Meta lifts 2026 capex to 145 billion dollars, shares plunge as investors question AI returns.

Samantha: Meta has just raised its 2026 capital expenditure forecast to between 125 and 145 billion dollars. That's a staggering jump, and Wall Street did not love it. Shares dropped sharply despite earnings being boosted by AI-driven advertising.

Bob: And the gap between those two numbers, twenty billion dollars, is itself larger than the entire annual revenue of most European tech companies. Mark Zuckerberg is framing this as a long-term bet on what he calls agentic AI, and he pointed to new models like Muse Spark as proof the spending is paying off.

Samantha: The money is going into infrastructure, data centres, custom silicon, and an aggressive talent grab. Meta is essentially trying to buy its way into the front of the AI race after lagging behind OpenAI and Google for parts of the last cycle.

Bob: At the same time, the company is dealing with mounting child safety litigation, which adds a financial and reputational headwind. Investors are asking a fair question: when does this spending actually convert into profit, rather than just bigger electricity bills?

Samantha: The bottom line is that we're watching the most expensive technology arms race in history play out in real time. And every hyperscaler is doing the same thing, which means the capital intensity of being a top-tier AI player keeps rising.

Bob: That has knock-on effects for the whole industry. If Meta, Microsoft, Google and Amazon are collectively spending close to half a trillion dollars a year on AI infrastructure, smaller competitors simply cannot keep up on their own terms. They have to specialise or partner.

Samantha: For Europe, Muse Spark and similar models will land in the EU market under the AI Act. So we're going to see a real test of how those compliance rules apply to frontier American models trained on this kind of compute.

Bob: And the child safety angle is highly relevant here. Brussels and several member states are already pushing hard on online safety and digital rights. A Meta that's distracted by US litigation may be less willing to fight European rules, which could actually help regulators get concessions.

Samantha: There's also the competitive question. European AI champions like Mistral or Aleph Alpha simply cannot match 145 billion dollars in capex. Their pitch has to be efficiency, sovereignty and trust, not raw scale.

Bob: Exactly the strategic positioning we're seeing emerge. Now, speaking of enormous AI bets, let's move to Tokyo.

SoftBank plans hundred-billion-dollar US listing for new AI and robotics venture called Roze.

Bob: SoftBank is reportedly setting up a new AI and robotics company called Roze, with plans to list it in the United States at a target valuation of a hundred billion dollars. The Financial Times says it could happen as soon as this year.

Samantha: Masayoshi Son is once again going big. The strategy is to offset SoftBank's enormous AI commitments, including its OpenAI investment, by creating a publicly traded vehicle that captures some of that upside. Roze is expected to focus heavily on data centre construction.

Bob: The structure here matters. By bundling AI and robotics into one listed entity, SoftBank can raise capital from public markets without diluting its existing Vision Fund stakes. And data centres are the picks-and-shovels play of this AI gold rush.

Samantha: It also gives Son a way to recycle returns. SoftBank has been selling down parts of its Arm stake and other assets to fund AI commitments. A successful Roze IPO would essentially refinance the AI portfolio at a much higher valuation.

Bob: If Roze actually prices at a hundred billion, it instantly becomes one of the largest pure-play AI infrastructure companies in the world. That changes how institutional investors allocate to the sector.

Samantha: And it signals that the public markets, not just venture capital, are now the funding engine for AI infrastructure. That's a meaningful shift after years of private mega-rounds.

Bob: For Europe, more global data centre capacity means cheaper compute over time, which helps European AI startups and researchers. But it also means more competition for engineers, power contracts and grid connections.

Samantha: European cloud providers like OVHcloud and Deutsche Telekom's T-Systems will feel pressure on pricing. And there's a sovereignty question: do European AI workloads run on Roze infrastructure in Virginia, or on European hardware under European law?

Bob: Which is exactly why the EuroStack debate keeps gaining traction. Public listings of foreign AI infrastructure giants only sharpen that conversation.

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.

EU AI Act enforcement begins August 2026 as US states roll out their own AI laws.

Samantha: The regulatory clock is ticking. The European Commission's enforcement powers for general-purpose AI model providers under the EU AI Act kick in on August 2, 2026. That means real fines, real audits, real consequences.

Bob: And it's not just Brussels. Colorado's AI Act takes effect on June 30, 2026, requiring developers and deployers to use reasonable care to avoid algorithmic discrimination, plus risk management policies and impact assessments. Several other US states are following.

Samantha: For general-purpose AI providers, the EU rules mean documentation requirements, transparency on training data, copyright compliance, and systemic risk evaluations for the largest models. Fines can reach seven percent of global turnover.

Bob: Colorado's approach is different but complementary. It focuses on high-risk AI used in consequential decisions like hiring, lending and housing. Deployers, not just developers, carry obligations. That's a US first at state level.

Samantha: The bottom line is that AI governance is no longer theoretical. Companies need legal teams, technical documentation, and impact assessments ready in months, not years.

Bob: And the divergence between jurisdictions is the real headache. A model deployed in Frankfurt, Denver and Tokyo now faces three different rule sets. Smaller companies will struggle with that overhead.

Samantha: For European businesses, August 2 is the date to circle. Any company using a general-purpose AI model from OpenAI, Anthropic, Google or Meta needs to know whether their provider is compliant, because liability flows downstream.

Bob: It's also a competitive opportunity for European AI companies that have built compliance in from day one. Mistral, for instance, has been positioning itself as the regulator-friendly option, and that pitch gets stronger after August.

Samantha: The interplay with US state laws also matters for European exporters. A European AI vendor selling into Colorado now has to meet both the AI Act and Colorado's requirements simultaneously.

OpenAI reportedly building an AI agent smartphone for 2028 with MediaTek and Qualcomm.

Bob: OpenAI is pushing further into hardware. Reports say the company is developing an AI agent-powered smartphone for a 2028 launch, working with MediaTek and Qualcomm on a custom processor. China's Luxshare Precision would handle system integration and manufacturing.

Samantha: This is OpenAI's most ambitious hardware play yet, beyond the Jony Ive collaboration we've heard about. The pitch is a phone built around AI agents that perform complex tasks autonomously, rather than today's app-centric model.

Bob: The supplier mix is interesting. MediaTek and Qualcomm together suggests OpenAI wants flexibility on silicon. Luxshare is a major iPhone assembler, so OpenAI is tapping the Apple supply chain playbook directly.

Samantha: A 2028 timeline is also realistic for a custom processor. Designing silicon, validating it, and building software around it takes years. This isn't a quick experiment, it's a serious long-term commitment.

Bob: If agents really can book your travel, manage your inbox and negotiate on your behalf, the home screen full of apps starts to look obsolete. That's the bet OpenAI is making.

Samantha: It's also a direct challenge to Apple and Google. Both control the existing mobile AI experience through Siri and Gemini. OpenAI inserting itself with dedicated hardware is a frontal assault on that duopoly.

Bob: For European consumers, this means another option in 2028, assuming it ships on time. But it also means another non-European platform shaping how Europeans interact with AI on their most personal device.

Samantha: European telcos will need to adapt. If agents handle tasks autonomously, data usage patterns change dramatically. And European regulators will have to figure out how the AI Act applies to an always-on agent making decisions on your behalf.

Bob: That's a genuinely novel regulatory question. Liability for an autonomous agent's actions is not yet settled law anywhere.

US blocks chip equipment shipments to China's Hua Hong Semiconductor.

Samantha: Washington has tightened the screws again. The US Department of Commerce ordered several chip equipment companies to halt shipments of certain tools to Hua Hong Semiconductor, China's second-largest chipmaker. Two specific Hua Hong facilities are the targets.

Bob: Those facilities are believed to produce China's most advanced chips, including parts critical for AI workloads. So this is a direct strike at China's AI hardware supply chain, not just generic semiconductors.

Samantha: The pattern is clear. Each round of US controls targets the next layer of Chinese capability, from SMIC to Huawei to Yangtze Memory and now Hua Hong. Beijing's response has been to accelerate domestic equipment development, but that takes years.

Bob: And the affected equipment makers, mostly American but with Dutch and Japanese partners, lose a major customer overnight. ASML and Tokyo Electron will be watching closely, because Washington often pressures allies to match its restrictions.

Samantha: This is no longer about chips alone. It's about who controls the foundational technology of the AI era. Every restriction reshapes global supply chains and forces companies to pick sides.

Bob: The unintended consequence is that China is now investing massively in indigenous lithography and etching tools. In ten years, the global semiconductor map could look very different.

Samantha: For Europe, this is uncomfortable. ASML in the Netherlands is the world's only maker of EUV lithography machines, and it's already restricted from selling its most advanced systems to China. More US pressure means more lost revenue for a strategically vital European company.

Bob: European automakers, industrial firms and AI developers also depend on stable chip supply. Any disruption flows through to manufacturing in Stuttgart, Munich and Eindhoven.

Samantha: Which is why the European Chips Act and its successor proposals matter so much. Building domestic capacity is no longer optional, it's strategic infrastructure.

Amazon may start selling its Trainium AI chips directly to outside customers.

Bob: And to close on a lighter and frankly fun note, Amazon is thinking about selling its Trainium AI chips to external customers. CEO Andy Jassy says it could happen over the next couple of years, potentially turning a 20 billion dollar in-house chip business into a 50 billion dollar one.

Samantha: The fun part is that Amazon's own internal demand already exceeds supply. So they're essentially saying, our chips are so popular with ourselves that we might let other people buy them too. That's a pretty good marketing pitch.

Bob: This puts Amazon in direct competition with NVIDIA, which currently dominates AI accelerators with margins approaching ninety percent on some products. Even modest price competition from Trainium racks could squeeze that.

Samantha: It also follows Google's similar move with TPUs being available through cloud rentals. The hyperscalers are quietly becoming chip vendors, which is a wild plot twist for an industry that NVIDIA seemed to own outright two years ago.

Bob: More options means lower prices, which means cheaper AI for everyone. Enterprises that have been waiting twelve months for NVIDIA H100s might suddenly have alternatives.

Samantha: And it forces NVIDIA to keep innovating. Competition is the best gift any market can get, especially one this concentrated.

Bob: For European cloud providers and enterprises, Trainium racks would be a genuine alternative. Sourcing flexibility matters enormously when you're trying to build sovereign European AI capacity.

Samantha: It also fits the European preference for diversified supply chains. Less dependence on one Californian chipmaker is something Brussels has explicitly been pushing for.

Bob: And honestly, the idea of Amazon, NVIDIA, Google and AMD all fighting for the same European AI customer in 2027 is a much healthier picture than where we are today.

Samantha: Today we covered: Meta's massive AI capex jump, SoftBank's Roze IPO plans, AI regulation deadlines in the EU and US, OpenAI's 2028 smartphone, US chip restrictions on Hua Hong, and Amazon's plan to sell Trainium chips externally.

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.