OpenAI Escapes Azure, Europe Pries Open Android

If You Only Read One Thing

AI's most important fight today is not over model quality; it is over who controls the doorway. OpenAI and Microsoft rewrote their pact because one cloud cannot contain OpenAI's ambitions, while Brussels is trying to make Android's AI layer interoperable before Gemini becomes the default gatekeeper. The stack is opening from private necessity and public force.

OpenAI Gets Multi-Cloud, Microsoft Gets Certainty

The old Microsoft-OpenAI deal was built for a world where OpenAI needed one patron with money, GPUs, and distribution. The new deal is built for a world where OpenAI needs every cloud it can get and Microsoft needs the relationship to stop looking like a single point of failure.

On Monday, OpenAI and Microsoft announced an amended agreement: Microsoft remains OpenAI's primary cloud partner, OpenAI products still ship first on Azure unless Microsoft cannot support the needed capabilities, and OpenAI can now serve customers across any cloud provider. Microsoft keeps a non-exclusive license to OpenAI models and products through 2032. Microsoft will no longer pay OpenAI a revenue share, while OpenAI keeps paying Microsoft revenue share through 2030, at the same percentage but subject to a cap. Axios reported that OpenAI still pays Microsoft 20% of revenue through 2030, now capped, and that Amazon distribution could be imminent.

Why it matters: Exclusivity used to be a moat. In frontier AI, it is becoming a capacity constraint. A cloud can win by owning the scarce partner when the partner is small enough to fit inside the cloud's capital plan. Once the partner is raising at an $852B valuation, courting Amazon, flirting with Qualcomm hardware, and defending an IPO path in court, exclusivity starts to cap both sides. OpenAI gets to tell enterprise customers that procurement, data residency, and latency no longer route through one vendor. Microsoft gives up the cleanest distribution lock but keeps the thing investors care about more: model IP rights through 2032, shareholder upside, Azure-first launches, and a capped economics schedule. The practical shift is that OpenAI is no longer a Microsoft product channel. It is an application layer negotiating with infrastructure suppliers.

Room for disagreement: This is not a divorce. Microsoft remains the primary cloud partner and the first launch surface, which means Azure still captures the early enterprise halo. OpenAI also traded away some economics to get the freedom; a capped revenue-share obligation is cleaner for an IPO, but it is still a claim on cash flows in a company already burning heavily. The stronger reading may be less "OpenAI escaped" than "both sides made the contract legible for bankers."

What to watch: Whether OpenAI models appear as first-party products on AWS or Google Cloud by the end of June. That would confirm the amendment is a distribution reset, not just legal cleanup before the Musk trial and IPO process.

Europe Tries To Open Android Before Agents Harden

The European Commission is not waiting for AI assistants to become the new browser. It is trying to write the access rules while the category is still forming.

On April 27, the Commission sent preliminary findings to Google under the Digital Markets Act specification proceeding it opened in January. The draft measures would require Google to give third parties effective access and interoperability with key Android capabilities, including the AI services that compete with Gemini. Interested parties have until May 13 to respond. The Commission's stated goal is straightforward: Android users should be able to integrate rival AI services into the phone without losing functionality that Google's own AI receives.

Think of a specification proceeding as Brussels turning a broad gatekeeper rule into a product checklist. The DMA says dominant platforms cannot use control of core services to foreclose rivals. The specification process asks the operational question: which APIs, device hooks, and system privileges must Google expose so a rival assistant can actually work?

Why it matters: The assistant is moving from app to operating-system layer. A chatbot you open voluntarily competes on quality. An agent that can wake by voice, read screen context, call other apps, use local files, and execute tasks competes on placement. If Gemini gets those hooks by default while Claude, ChatGPT, Mistral, or a European assistant has to live as an ordinary app, the mobile AI market hardens before users ever make a choice. Brussels is applying the old Android-search and browser lesson to agents: distribution is not neutral when the platform owner controls the invocation surface. The deeper point is that AI regulation is moving from model safety to interface competition. Europe may not build the leading model, but it can still shape where models are allowed to sit.

Room for disagreement: Google's security argument is real. Giving rival agents access to privileged device capabilities increases the attack surface and turns every interoperability mandate into a permission-design problem. If the Commission writes the checklist too broadly, it could make Android less secure while subsidizing low-quality assistant clones. If it writes it too narrowly, Gemini keeps the advantage and the proceeding becomes theater.

What to watch: Whether the final measures require equally effective access to voice activation, screen context, and app-execution hooks. Those are the three capabilities that separate a default assistant from a downloadable app.

The Contrarian Take

Everyone says: Big AI partnerships are breaking apart because everyone wants optionality. OpenAI wants AWS, Microsoft wants a multi-model strategy, and Google is being forced to share Android.

Here's why that's wrong (or at least incomplete): The platforms are not becoming less powerful. Their control points are becoming more specific. Microsoft gave up exclusivity but kept OpenAI IP through 2032 and Azure-first launches. Google may be forced to open Android, but the fight exists because Android's privileged hooks are valuable enough to regulate. GitHub's billing shift below says the same thing from the developer side: whoever controls the work surface controls the meter. The AI stack is opening at the edges while the surviving choke points become more explicit.

Under the Radar

GitHub turns Copilot into metered infrastructure. GitHub says all Copilot plans move to usage-based billing on June 1, with AI Credits consumed by token usage and code completions still included. Copilot code review will also consume GitHub Actions minutes for private repos. This is the end of the "unlimited agent" illusion: agentic coding is a cloud workload with a bill, not a flat SaaS feature.
Meta is buying an energy option from orbit. Meta signed a capacity reservation with Overview Energy for up to 1 GW of power from spacecraft that would beam infrared light to solar farms at night. Overview plans a 2028 low-Earth-orbit demonstration and a 2030 geosynchronous fleet. The useful signal is not that space solar is suddenly bankable. It is that AI demand has made even exotic power options worth reserving.

Quick Takes

Gas turbines are becoming an AI bottleneck. TechCrunch, summarizing BloombergNEF data, reports that combined-cycle gas power plant costs rose from under $1,500 per kilowatt in 2023 to $2,157 last year, while builds now take 23% longer. Data-center demand is expected to rise from 40 GW today to 106 GW by 2035, and gas-turbine equipment prices are projected to be up 195% over 2019 levels by year-end. The AI power trade is moving from land and permits to turbine queues. (Source)
A $1.1B seed round makes UK sovereign AI concrete. Former DeepMind researcher David Silver's Ineffable Intelligence raised $1.1B at a $5.1B valuation, with Sequoia, Lightspeed, Index, Google, Nvidia, the British Business Bank, and the UK's Sovereign AI fund involved. The lab's technical claim belongs in the AI briefing; the news angle is that London is turning DeepMind alumni into an industrial-policy asset. (Source)
The FTC put a number on social-platform fraud. Americans reported $2.1B in 2025 losses from scams that started on social media, up eightfold, with Facebook the largest origin point and WhatsApp and Instagram a distant second and third. Investment scams alone accounted for $1.1B. The regulatory risk for Meta is not just content moderation; it is payments-like fraud exposure on an attention marketplace. (Source)

The Thread

The common story is access. OpenAI wants access to every cloud because one supplier cannot satisfy its compute and enterprise-distribution needs. Rival AI services want access to Android's privileged hooks because a great model is less useful if the operating system gives Gemini the first move. GitHub wants access to the billing meter because agentic coding turns every long task into token and runner consumption. AI is not dissolving platforms. It is forcing the industry to say which platform privileges are contractual, which are regulatory, and which are simply infrastructure bills.

Predictions

New predictions:

I predict: OpenAI models will become directly purchasable through AWS Bedrock or Google Cloud Marketplace by 2026-06-30. (Confidence: high; Check by: 2026-06-30)
I predict: The European Commission's final Android DMA measures will require equally effective access to at least three AI-assistant capabilities: voice activation, screen context, and app action execution. (Confidence: medium; Check by: 2026-07-31)

Generated 2026-04-28, 03:38 ET.