28 April 2026: OpenAI Frees Itself From Microsoft Lock-in (AM)

The Microsoft and OpenAI alliance has been quietly rewritten, freeing OpenAI to sell on Amazon Web Services and giving Microsoft a long fixed timeline plus more cash. China has blocked Meta’s $2B push into AI agents, a former DeepMind star has just raised £840m for a new British lab, and Google has reopened its free five day AI agents course. The shape of the AI market is shifting fast, and most of it now runs through agents.

Microsoft and OpenAI have renegotiated the partnership that defines half the AI economy. The two companies announced on Monday that OpenAI can now sell its products on Amazon Web Services, breaking a years long exclusivity arrangement that had been straining under OpenAI’s recent infrastructure deals. In return, Microsoft secures more cash through a revenue share agreement and a fixed end date, with a nonexclusive licence to OpenAI models and intellectual property running through 2032.

The trigger was OpenAI’s up to $50 billion compute commitment with Amazon, signed in February, which was on a collision course with the original Microsoft contract. OpenAI gets the right to spread workloads across more cloud providers. Microsoft gets recurring revenue and a defined runway.

For business users, ChatGPT Enterprise and the OpenAI API can now be procured through AWS Marketplace, which matters for any UK firm standardised on Amazon’s cloud. The old assumption that Azure is the default route to OpenAI no longer holds.

China has blocked Meta’s $2 billion acquisition of AI agent startup Manus after a months long antitrust review. Reuters and the BBC confirmed the decision late on Monday, with Beijing ordering Meta to unwind the deal it had signed earlier this year. Manus, the Singapore based but China rooted maker of one of the most capable autonomous web browsing agents, is now back to operating independently.

This is a bruise for Mark Zuckerberg’s plan to push Meta hard into agents that book travel, fill forms and run multi step research tasks on a user’s behalf. Manus’s product was the most polished consumer agent on the market when the deal was announced, and Meta had hoped to bolt it onto WhatsApp and Instagram inside a year.

The block is also a reminder that AI agent acquisitions are now being treated as strategic infrastructure deals by Beijing, in the same bucket as semiconductor takeovers. Expect every cross border AI deal involving a Chinese asset to face a similar review for the rest of the year.

A former DeepMind researcher has raised $1.1 billion to build AI that learns without human data. David Silver, the British computer scientist behind AlphaGo and AlphaZero, announced funding for his new lab Ineffable Intelligence at a $5.1 billion valuation, less than a year after he left Google DeepMind. The company is based in London and is hiring across reinforcement learning, evaluation and infrastructure roles.

The pitch is that the next leap in capability will not come from scraping more of the public internet, which is largely exhausted, but from systems that learn through self play and interaction with simulated environments. That is the same recipe that took AlphaGo from amateur to superhuman in a single training run.

For the UK, this is a meaningful signal. London now hosts a serious frontier lab that is not a Google subsidiary, with capital that puts it in the same league as Mistral in Paris. It also gives British engineers a domestic alternative to relocating to San Francisco for top AI roles.

Google and Kaggle have reopened registration for their free five day AI agents intensive course. The course runs in June 2026 and is aimed at developers who want hands on experience building autonomous agents using Gemini, function calling and the new agent development kit. The previous edition drew more than 250,000 sign ups and pushed Kaggle past Hugging Face on weekly active learners.

Sessions cover prompt engineering for agents, tool use, multi agent coordination and evaluation. There is a final capstone project that has historically been a useful CV item for engineers moving into AI roles. Registration is free, runs entirely online and is open globally including to UK and EU residents.

If you run a small business and have wondered how AI agents could automate parts of your workflow, this is the cheapest credible way to find out. The course materials remain online indefinitely after the live week ends.

Skye, a new AI home screen app for iPhone, has secured backing before its public launch. The app, built by Signull Labs, replaces the standard iOS home screen with an agent that surfaces what it thinks you need next based on context, time of day and recent activity. Investors include the Founders Fund and a clutch of former Apple executives.

The bet is that the next consumer interface for AI is not a chatbot you open but an ambient layer that sits above your apps. Skye does not run the apps for you. It just decides which ones to put in front of you and when. That places it firmly in the category of AI products competing with the iPhone home screen itself, which is a notable shift now that Apple’s own Apple Intelligence rollout has stalled in Europe.

Skye is invite only for now, with an App Store launch expected in May.

Here is everything else worth knowing from this morning’s AI news.

• OpenAI hits FedRAMP Moderate: ChatGPT Enterprise and the OpenAI API are now authorised for US federal agencies, opening a procurement channel that often sets standards for UK public sector buyers.
• OpenAI may be building a phone: A note from analyst Ming-Chi Kuo suggests OpenAI’s hardware ambitions could include a device where AI agents replace traditional apps.
• Mercor data breach: Roughly 4TB of voice samples from 40,000 AI training contractors at Mercor have leaked, raising fresh questions about voice data in model training pipelines.
• Meta signs space solar deal: A small but symbolic contract with Overview Energy for night time solar power beamed from orbit, aimed at AI data centre demand.
• Nvidia’s Raw2Insights for ultrasound: A new physics informed AI model for adaptive ultrasound imaging that could improve point of care scans in NHS settings.
• Open source agent tops TerminalBench: The Dirac CLI agent, built on Gemini 3 Flash Preview, scored 65.2 percent versus Google’s official 47.8 percent.
• Talkie language model: A 13 billion parameter vintage language model trained to write in 1930s English has launched for writers and historians.

This is a daily news update for informational purposes only. AI products and policies change rapidly. Verify details directly with providers before making decisions. Nothing here is financial or legal advice.

AI Daily is Cristoniq’s daily guide to developments in artificial intelligence, published every morning.