Morning Brief · Wednesday

OpenAI Just Ended Its Exclusive With Microsoft. AWS Is the New Favorite.

OpenAI and Microsoft renegotiated their partnership — Azure's exclusivity is over, and OpenAI is launching Codex and Managed Agents on Amazon Bedrock. Musk returns for Day 2 of cross-examination after a Day 1 that left the courtroom largely unmoved. Apple previews four AI photo editing tools coming to iOS 27. And dozens of Baidu's Apollo Go robotaxis froze in Beijing traffic — a story that's been sitting quietly in the background until Bloomberg surfaced it today.

Mira Novian

April 29, 2026 · Morning Brief · ~8 min read

OpenAI ends Microsoft's cloud exclusivity — and immediately picks AWS as its next primary partner

In a coordinated announcement Monday, OpenAI and Microsoft amended their foundational partnership agreement, ending Azure's exclusive lock on OpenAI models. Under the revised deal: Microsoft remains OpenAI's "primary" cloud partner, but OpenAI can now serve its products on any cloud provider. Microsoft's license to OpenAI IP through 2032 becomes non-exclusive. Microsoft stops paying a revenue share to OpenAI. And OpenAI's payments to Microsoft continue through 2030 at the same percentage but subject to a total cap. Microsoft remains a major OpenAI shareholder.

The announcement landed in parallel with a separate reveal: AWS and OpenAI are launching an expanded partnership on Amazon Bedrock, bringing OpenAI's latest frontier models, Codex (the coding agent), and Amazon Bedrock Managed Agents powered by OpenAI — all in limited preview — to the world's most widely deployed cloud infrastructure. Enterprises will be able to access OpenAI models alongside Anthropic, Meta, Mistral, and other Bedrock providers through a single API. Ben Thompson at Stratechery, who interviewed both OpenAI CEO Sam Altman and AWS CEO Matt Garman ahead of the embargo, put it plainly: "It seems clear that OpenAI's focus is going to be on AWS." The calculus makes sense: Azure's exclusivity benefited Azure's differentiation but was actively hurting OpenAI, particularly as Anthropic's cloud-agnostic strategy drove rapid enterprise adoption on all three hyperscalers simultaneously.

aboutamazon.com ↗

This is one of those announcements that looks incremental on the surface but is structurally massive. Azure's exclusivity was a moat — it was the reason enterprises that didn't already have an AWS or GCP footprint were signing Azure contracts. That moat just got filled in. For OpenAI, the calculus is obvious: enterprises choose their cloud first, then their AI vendor. Locking to Azure was losing deals to Anthropic on Bedrock and Vertex every single week. Now OpenAI competes on merit everywhere. The more interesting question is what this means for Microsoft's AI narrative. Copilot and Azure's AI pitch was predicated partly on OpenAI access. That advantage is now table stakes rather than a differentiator. Microsoft appears to have negotiated out of the revenue share payments it was making — freeing its own PnL — in exchange for losing the exclusive. That's a sensible trade when your investment is already locked in via equity. But the Azure AI story is going to need a new center of gravity.

Legal

Musk enters cross-examination after a Day 1 the courtroom found more petty than persuasive

Day 2 of Musk v. OpenAI is underway in Oakland, with Elon Musk returning to the stand for cross-examination. Day 1 set a striking tone: Musk's direct examination left observers, including The Verge's Elizabeth Lopatto who was in the courtroom, cold. The narrative his attorneys tried to tell — OpenAI was founded as a nonprofit to save humanity from AI, Musk was its engine, and Altman betrayed the mission — kept getting derailed by Musk's instinct to brag. He spent significant time recounting his biography, claiming to work "80 to 100 hours a week," and describing how he came up with the idea, the name, recruited the key people, and "taught them everything I know." He delivered the line with a pause for laughter; one or two people obliged. Most of the courtroom was silent.

The more revealing moments came when Musk disclosed that crypto issuance had been discussed as a funding model at the founding stage ("I was against that because it sounded kinda scammy"), that he was "not averse to a small for-profit" structure from the beginning, and that he wanted a majority equity stake that would dilute over time because he was "providing all the money." He also identified Shivon Zilis as "my chief of staff and, uh, you know" — prompting audible laughter from someone who clearly knew more about that relationship than the jury did. Judge Yvonne Gonzalez Rogers has already reprimanded OpenAI's lawyers for taking inconsistent positions on the company's name's origin: "Do not take inconsistent positions in front of me." Cross-examination from OpenAI's attorneys today is expected to probe the gap between Musk's founding narrative and the contemporaneous documents.

theverge.com ↗

The trial is revealing something that will matter for the historical record regardless of the verdict: the actual texture of OpenAI's founding is a lot messier than either party's public story. Musk wanted equity that would dilute over time — that's not a pure altruist framing. He discussed for-profit structures and crypto issuance. He wanted a majority stake because he was supplying the funding. These are rational positions for a startup founder, but they're at odds with the "I sacrificed wealth to save humanity" narrative he's telling the jury. The defense's job on cross is to put those documents in front of the jury and ask what they reveal about Musk's actual intent. The thing is, even if Musk's founding motives were self-interested, that doesn't mean OpenAI's conversion was fair. Both things can be true. That's the interesting legal puzzle — and probably the reason the fraud claims were dropped. Hard to prove fraud when the founding documents are ambiguous enough that a reasonable person could read them either way.

Consumer AI

Apple's iOS 27 will add four AI photo editing tools to the Photos app this fall

Bloomberg's Mark Gurman reports that iOS 27 will introduce an "Apple Intelligence Tools" section to the Photos app editor, featuring four AI-powered editing capabilities: Extend, Enhance, Reframe, and Clean Up. Clean Up is the only one of the four currently available in the Photos app — it arrived with Apple Intelligence last year and allows users to remove objects from photos. Extend (fills in the edges of a cropped or narrow photo), Enhance (improves lighting, color, and detail), and Reframe (recomposes a shot by adjusting framing and aspect ratio intelligently) are all new. The features will begin rolling out in beta starting next month ahead of a fall public release.

The announcement puts Apple directly in competition with Google's Photos editing suite, which has featured AI-powered tools including Magic Eraser, Magic Editor, and Photo Unblur for several years. In NI's own testing when Clean Up debuted, it fell meaningfully short of Google's Magic Editor in complex scenes. The four new tools represent Apple's most significant push into AI-powered photo editing since launching the feature category — and for iPhone users who primarily live inside Apple's ecosystem, they may not need to go elsewhere.

theverge.com ↗

The timing of Apple Intelligence's photo features matters. Google has had a multi-year head start on AI photo editing, and it shows — Magic Editor is genuinely impressive in ways Apple's Clean Up hasn't been. But Apple's advantage is distribution. When these four tools ship in iOS 27, they go to every iPhone running the update. They don't require a Pixel, a Google One subscription, or a separate app. That's a different competition surface than feature parity — it's about normalizing AI editing as a default behavior for a billion users. The more interesting story is what's happening on Apple's model strategy underneath. These editing tools require inference at a level of quality that wasn't feasible on-device a year ago. Whether that's running on-device, in Apple's Private Cloud Compute, or some hybrid will tell you a lot about how Apple is actually thinking about the limits of its silicon.

Autonomy

Dozens of Baidu's Apollo Go robotaxis froze in Beijing traffic last month — sparking alarm

Bloomberg is reporting that dozens of Baidu's Apollo Go robotaxis simultaneously froze in traffic in Beijing last month, causing significant disruptions before human operators were able to remotely intervene and clear the vehicles. The incident sparked internal alarm at Baidu and drew scrutiny from city officials, according to Bloomberg's sources. Details about the root cause — whether a software update, sensor failure, connectivity issue, or edge-case scenario triggered the fleet-wide freeze — have not been publicly disclosed by Baidu.

Apollo Go is one of China's largest robotaxi deployments, operating commercial driverless rides across Beijing, Wuhan, and several other Chinese cities. The service has been a centerpiece of Baidu's autonomous vehicle ambitions for years. A simultaneous multi-vehicle freeze is a qualitatively different kind of failure than a single-vehicle incident — it suggests a systemic issue that could affect any vehicle in the fleet rather than an isolated hardware or software fault. The incident mirrors concerns that have followed autonomous vehicle deployments globally: that the software systems underlying them may behave safely in 99.9% of scenarios while harboring failure modes that only emerge in edge cases at scale.

theverge.com ↗

The phrase "froze in traffic" sounds almost benign until you think about it. These are multi-ton vehicles that simultaneously became obstacles in live traffic because of a software condition. No one was hurt, apparently — the remote intervention worked — but the incident exposes something that autonomous vehicle advocates have been reluctant to say clearly: fleet-level failures are a category of risk that doesn't exist with human drivers. One human driver can have a medical episode and block traffic. But dozens of human drivers don't simultaneously experience the same medical episode because of a shared software update. The scalability that makes robotaxis economically interesting is also the property that makes systemic failures possible. The fact that this happened last month and is only surfacing now is also worth noting. Baidu has strong incentives to keep incidents quiet in a regulatory environment that's been generous to AV deployment. The regulators and the public have strong interests in a different default — one where disclosure is prompt and incident data is public. The gap between those interests is where future disasters are hiding.

✦ Mira's Take

The OpenAI-Microsoft story is getting framed as a breakup, but the more accurate frame is a graduation. OpenAI has grown to the point where Azure's exclusivity is a constraint on its growth rather than a benefit from its patronage. The amended deal is structured to make both parties whole — Microsoft loses the differentiation but gains a cleaner PnL and retains equity upside; OpenAI gains access to the 60%+ of enterprise cloud workloads that run on AWS and GCP. It's a sensible negotiation, and Anthropic's success as a cloud-agnostic model provider is probably what forced it. When your main competitor is accessible on all three hyperscalers and you're only on one, you're losing deals structurally. That's now fixed. The next question is whether AWS Bedrock becomes a genuinely differentiated platform or just a distribution layer. Managed Agents powered by OpenAI is the most interesting bet — it's a direct play for enterprise agentic workloads, and if it works, AWS becomes the enterprise AI substrate for OpenAI in the same way Azure was the training substrate.

The trial is the most consequential AI governance event happening right now, and it's not getting the attention it deserves because Musk is a polarizing figure and it's easy to write the whole thing off as billionaire theater. But the testimony is creating a public record of how these organizations actually worked at their founding — with all the ambiguity, self-interest, and competing narratives intact. Ilya Sutskever hasn't testified yet. Mira Murati hasn't testified. Satya Nadella hasn't testified. The documentary record of what people believed they were building, and what they were actually building, is going to be examined under oath by people with excellent legal teams and strong incentives to tell the truth. That's not theater. That's accountability, and it's rare.

The Baidu story is the one I'd watch most carefully over the next few weeks. A fleet-level simultaneous failure is a different risk topology than the individual incidents the AV industry has been litigating over. If the root cause was a software update — which is the most likely explanation for a synchronized multi-vehicle event — then every large AV deployment using shared software faces the same failure mode. The industry's response to this incident will tell you whether there are actual safety governance mechanisms in place or whether "incident response" means "fix it quietly and keep operating." The disclosure lag is itself information.