OpenAI ends Microsoft exclusivity: what it means for Europe

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

Bob: And I'm Bob Russell. Today: China blocks Meta's acquisition of AI startup Manus, a Taiwan court hands down a ten-year sentence over TSMC trade secret theft, OpenAI and Microsoft tear up their exclusivity deal, Anthropic finds AI agent quality quietly tilts negotiations, governments worldwide stall on scaling AI according to KPMG, and Google Cloud Next 2026 puts AI into every corner of the cloud. Let's start with the Meta-Manus block.

China blocks Meta's acquisition of AI agent startup Manus and bars founders from leaving the country.

Samantha: Beijing has officially blocked Meta's acquisition of the AI startup Manus. China's National Development and Reform Commission halted the deal after a regulatory review and reportedly ordered the parties to walk away. Meta agreed to buy Manus back in December for its AI agent technology. The company was founded in China but is now headquartered in Singapore, and that ambiguity is exactly what made this transaction so politically charged.

Bob: And there's a sharper edge to this. Two of the Manus co-founders have reportedly been restricted from leaving China while the review plays out. That's not a normal antitrust intervention, that's a national security signal. Meta says the transaction fully complied with applicable law and expects an appropriate resolution, but analysts read this as Beijing tightening control over the AI industry and slowing technology transfer westward.

Samantha: It also fits a pattern. Beijing wants AI talent and AI IP to stay inside China, even when the corporate vehicle has formally moved offshore. Singapore headquarters apparently doesn't insulate you if your founders, your engineers, and your training data trace back to the mainland.

Bob: This effectively kills the playbook of relocating an AI company to Singapore or the UAE and then selling to a US buyer. If China can claim jurisdiction through the founders, the chilling effect on cross-border AI M&A is enormous. Meta isn't just losing one deal, the entire category of acquisitions becomes more expensive and slower.

Samantha: For European buyers this cuts both ways. European companies eyeing AI startups with Chinese DNA now have to assume the deal can be blocked at the last minute, even after signing. And on the other side, Brussels has its own foreign subsidies regulation and FDI screening that increasingly looks at AI as a strategic asset.

Bob: Which means a Mistral or an Aleph Alpha trying to acquire abroad will face friction in three directions: the seller's home regulator, the EU's own screening, and US export controls if any American technology touches the stack. AI M&A in 2026 is no longer a commercial decision, it's a geopolitical one, and European boards need legal teams that understand all three regimes before they even start drafting term sheets.

Taiwan court sentences former Tokyo Electron engineer to ten years for stealing TSMC trade secrets.

Bob: Taipei has just delivered one of the toughest rulings yet under Taiwan's national security act. A former Tokyo Electron employee, who previously worked at TSMC, got ten years in prison for stealing semiconductor trade secrets. The court found he used his TSMC connections, after moving to the Japanese chip equipment maker, to help Tokyo Electron land additional contracts.

Samantha: Ten years is significant. Trade secret cases in most jurisdictions get civil damages or short sentences. Taiwan is treating semiconductor IP the way other countries treat espionage, and that reflects how central TSMC has become to the global AI buildout. Every Nvidia GPU, every advanced AI accelerator, runs through TSMC fabs. Protecting that process knowledge is now a sovereign priority.

Bob: And the AI boom is the backdrop here. Demand for advanced nodes has exploded, which means the value of leaked process recipes has exploded too. A few documents about extreme ultraviolet lithography integration can save a competitor years of R and D and billions in spending. Taiwan is essentially saying: if you take that, we will treat you as a threat to the state.

Samantha: The bottom line is that the legal perimeter around chip technology is hardening fast. Japan, the US, the Netherlands, and now Taiwan are all aligning their laws and enforcement around the same idea, leading-edge semiconductors are critical infrastructure.

Bob: For Europe this hits very close to home. ASML in Veldhoven sits at the same chokepoint, extreme ultraviolet lithography. The Netherlands has already tightened export controls and ASML has dealt with multiple IP theft cases over the past years. This Taiwan ruling gives European prosecutors a reference point for how seriously to take similar cases.

Samantha: It also raises questions about employee mobility within the European chip ecosystem. Imec in Belgium, Infineon in Germany, STMicroelectronics, the EU chips act ecosystem, all of them need clear contractual and legal frameworks for what engineers can take with them when they switch jobs. The principle that walking out with knowledge can become a criminal matter, not just a civil one, is now firmly established.

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 and Microsoft end exclusivity, allowing OpenAI models to ship beyond Azure for the first time.

Samantha: OpenAI and Microsoft have reportedly ended the exclusivity rights that defined their partnership for years. OpenAI's models, previously locked to Microsoft Azure, can now be distributed on other major cloud platforms. Microsoft keeps access to the models through 2032, but it loses the exclusive distribution moat that gave Azure such a competitive edge.

Bob: This is a structural change in the AI cloud market. Azure's pitch to enterprises was essentially, if you want GPT-class models, you come to us. That pitch is gone. OpenAI can now go to AWS, to Google Cloud, to Oracle, and to sovereign European clouds if it wants. Microsoft pivots from exclusive partner to one customer among many, although a very large one.

Samantha: And it explains why Microsoft has spent the last year diversifying. Heavy investments in their own MAI models, the Inflection team, custom silicon, partnerships with Anthropic. They saw this coming. The relationship has been visibly strained since OpenAI's governance crisis, and both sides clearly wanted optionality.

Bob: For enterprise buyers this is good news. The AI procurement conversation stops being about which cloud has the best model and starts being about price, latency, data residency, and governance. That's a healthier market.

Samantha: For Europe this could be transformative. OVHcloud, IONOS, Scaleway, the Gaia-X ecosystem, these players have struggled to offer frontier AI because the models lived behind Azure's wall. If OpenAI is genuinely willing to license to European providers under EU data residency terms, the sovereign cloud conversation suddenly has real substance.

Bob: There's also a regulatory upside. The EU AI Act and the Data Act both push toward portability and avoiding lock-in. An OpenAI that distributes more broadly fits that policy direction much better than the Azure-exclusive model. European procurement officers, especially in the public sector, finally get to negotiate from a position of choice rather than dependency.

Anthropic experiment shows weaker AI agents quietly lose negotiations without users noticing.

Bob: Anthropic ran an internal experiment called Project Deal, where Claude models acted as autonomous buying and selling agents in a real, incentivized marketplace among employees. The result: stronger AI models consistently delivered better commercial outcomes for their users. The unsettling part is that participants stuck with weaker agents often had no idea they were losing.

Samantha: That invisibility is the headline finding. In traditional markets, if you overpay for something, you usually figure it out eventually. With AI agents transacting on your behalf, the negotiation happens at machine speed, with reasoning you can't easily audit. You just see the outcome and assume it was reasonable.

Bob: And it points at a new kind of inequality. If wealthier consumers and larger companies deploy top-tier agents while smaller players use cheaper models, the price gap doesn't just persist, it compounds invisibly across millions of transactions. The frontier-model premium becomes a tax on everyone using the budget tier.

Samantha: This research moves the agentic AI conversation from hypothetical risks to concrete economic impact. We're no longer asking whether AI agents will negotiate on our behalf, they already do. The question is who benefits and who quietly subsidises the system.

Bob: For European regulators this is a meaningful flag. The Digital Markets Act, the AI Act, consumer protection rules, none of them currently address agent-versus-agent negotiation explicitly. If a small Portuguese retailer uses a basic AI agent and gets systematically out-negotiated by Amazon's frontier model, is that an unfair commercial practice? The legal frameworks aren't ready for that question.

Samantha: And there's a transparency angle. Europe could plausibly require disclosure of which AI agent class was used in a transaction, or mandate access to baseline-quality agents the way we treat universal banking services. It sounds far-fetched today, but a year ago so did mandatory chat interoperability, and now we have it under the DMA.

KPMG report finds half of governments plan AI at scale within a year, most are stuck in pilots.

Samantha: KPMG's Global Tech Report 2026 for the public sector landed today, and the picture is uncomfortable. Nearly half of governments worldwide say they plan to deploy AI at scale within the next year. The reality is that most are still trapped in pilot purgatory. The ambition is loud, the execution is slow.

Bob: The barriers KPMG identifies are familiar but stubborn. Departments don't coordinate. Data sits in silos that never integrate. And a huge slice of technology budgets still goes to keeping creaky legacy systems alive rather than funding transformation. You can't run a modern AI strategy on a 1990s mainframe with a chatbot bolted on top.

Samantha: And the talent gap matters too. Government salaries can't compete with private sector AI roles, so the people who could actually scale these projects end up at hyperscalers or AI labs. That leaves civil services dependent on consultants, which is expensive and rarely builds lasting capability.

Bob: When a bank can't scale AI, shareholders complain. When a government can't scale AI, people wait longer for benefits, healthcare appointments, asylum decisions, tax refunds. The stakes are different and the patience for failure is shorter.

Samantha: European governments face an extra layer here. They're not just scaling AI, they're doing it under the EU AI Act, which classifies many public sector uses, like welfare or migration, as high-risk. Compliance documentation, conformity assessments, fundamental rights impact assessments, that all sits on top of the basic integration challenge.

Bob: And there's the digital sovereignty thread. Member states want to use European cloud and European models where possible, which is politically right but technically harder because the tooling ecosystem is less mature. Estonia and Denmark are showing it can be done, but they're small and digitally native. Replicating that in France, Germany, or Italy at federal scale is a much bigger lift, and the KPMG numbers suggest most governments are nowhere near ready.

Google Cloud Next 2026 unveils eighth-generation Tensor chips and renames Vertex AI to Gemini Enterprise Agent Platform.

Bob: And to close on something a bit more fun, Google Cloud Next 2026 happened, and Google basically said: AI everywhere, all the time, in every product. The headline announcement is the eighth generation of Tensor chips, now split into two flavours, one optimised for inference and one for training. That split alone promises significant efficiency gains.

Samantha: They also rebranded Vertex AI to the Gemini Enterprise Agent Platform, which is a mouthful but tells you exactly where Google's heart is. Everything is now an agent. Build agents, manage agents, secure agents, deploy agents. There were also new AI security agents announced, basically autonomous defenders watching your cloud environment.

Bob: My favourite detail is the chip split. Google admitting that training and inference have such different requirements that you need different silicon is a quiet acknowledgement that the AI workload landscape has matured. Inference is now the bigger market by volume, training is bigger by intensity, and one chip can't do both well anymore.

Samantha: And the renaming is not just marketing. When you call your platform an Enterprise Agent Platform, you're telling developers that single-shot AI calls are yesterday's design pattern. Multi-step, tool-using, autonomous agents are the new default, and the tooling needs to reflect that.

Bob: For European customers Google Cloud has been making real efforts on data residency, with EU sovereign cloud offerings via partners like T-Systems. The new agent platform will need to demonstrate it works within those boundaries, particularly for regulated industries like finance and healthcare.

Samantha: Combine this with the OpenAI-Microsoft news from earlier and you see the pattern. The hyperscaler battle in 2026 is genuinely competitive again, and European buyers, for once, sit in a strong negotiating position. Wow moment of the day: AI agents are no longer a feature, they are the platform, and the cloud you knew last year is already obsolete.

Bob: And on that note, that's our six.

Samantha: Today we covered: China blocking the Meta-Manus deal, the ten-year sentence in Taiwan over TSMC trade secrets, OpenAI and Microsoft ending exclusivity, Anthropic's findings on AI agent quality in negotiations, KPMG on governments struggling to scale AI, and Google Cloud Next 2026 putting AI into every layer of the cloud.

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.