Microsoft Breaks Up With OpenAI (Sort Of) — And Google & Amazon Notice

The Headlines

Microsoft just did something it legally couldn't do six months ago: build world-class AI models without OpenAI.

On April 2, 2026, the company announced three new foundational models — MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 — developed entirely in-house by a team led by Mustafa Suleyman ^[1]. The transcription model reportedly beats both OpenAI's Whisper and Google's Gemini on most languages ^[2]. And it was built by 10 people for half the GPU cost ^[3].

On the exact same day, Google hit back with Gemma 4, its most capable open-source models, licensed under a more permissive Apache 2.0 framework ^[4].

But here's the real story: this isn't competition as usual. It's the unraveling of a partnership that just got renegotiated — and the legal system might be about to get involved.

The Backstory

Rewind to October 2025. Microsoft and OpenAI signed a new deal that changed everything.

Microsoft had invested in OpenAI since 2019, securing exclusive cloud partnership rights ^[5]. But under the renegotiated October agreement, Microsoft's stake — valued at approximately $135 billion and representing about 27% of OpenAI Group PBC — came with a critical new power: the right to independently pursue AGI, either alone or with other partners ^[6]. Microsoft's IP rights for models and products extended through 2032 ^[6].

Suleyman told VentureBeat exactly what that meant: "Back in September of last year, we renegotiated the contract with OpenAI, and that enabled us to independently pursue our own superintelligence. Since then, we've been convening the compute and the team and buying up the data" ^[7].

Then, on March 17, Microsoft reorganized Copilot leadership. Jacob Andreou, a former Snap executive, was promoted to head Copilot reporting directly to CEO Satya Nadella ^[8]. This freed Suleyman to do one thing: build models.

Why did Microsoft need to reorganize? Because Copilot was losing ground fast. In February 2026, the app had only 6 million daily active users versus ChatGPT's 440 million ^[8]. A survey of 150,000+ U.S. paid AI subscribers found Copilot's market share collapsed from 18.8% in July 2025 to just 11.5% by January — a 39% drop ^[9].

Microsoft had a problem to solve. And it chose to build its way out.

The Takes

The Bull Case — Microsoft's Independence Play

Suleyman's team is moving fast. In March, Microsoft hired Ali Farhadi, the former CEO of the Allen Institute for AI, along with key researchers Hanna Hajishirzi, Ranjay Krishna, and Ai2's COO Sophie Lebrecht ^[10]. Farhadi co-founded Xnor.ai, which Apple acquired in 2020 for an estimated $200 million ^[10].

By April, three models were live. MAI-Transcribe-1 achieved a 3.8% average Word Error Rate on the FLEURS benchmark across 25 languages, reportedly beating OpenAI's Whisper-large-v3 on all 25 and Google's Gemini 3.1 Flash on 22 of 25 ^[2]. (Note: these are Microsoft's own reported benchmarks, not independently verified.)

MAI-Image-2 debuted as a top-3 model on the Arena.ai leaderboard and delivers at least 2x faster image generation ^[1]. WPP, one of the world's largest marketing groups, is already using it ^[1].

The pricing is aggressive: MAI-Transcribe-1 starts at $0.36/hour, MAI-Voice-1 at $22/million characters, and MAI-Image-2 at $5/million text tokens and $33/million image tokens ^[11]. These models are already running Copilot's new Audio Expressions feature and rolling out in Bing and PowerPoint ^[12].

The narrative: Microsoft is becoming a competitor to OpenAI, not just an investor.

The Bear Case — Google and the Commodity Trap

Google launched Gemma 4 the same day, and it's a different strategy entirely ^[4]. The company switched to Apache 2.0 licensing, meaning developers can use these models commercially without restrictions ^[4].

Gemma models have been downloaded over 400 million times since launch, with more than 100,000 community variants created ^[4]. Gemma 4 features advanced reasoning, 140+ language support, native function calling, and native video/audio inputs. The 31B parameter model and 26B MoE variant both feature 256K-token context windows ^[13].

The catch: Google is racing toward open-source commoditization while Microsoft is racing toward proprietary in-house models.

But there's a darker backdrop. The Register reported that Chinese open-weights models from Moonshot AI, Alibaba, and Z.AI now reportedly rival GPT-5 and Claude ^[13]. The competition isn't just Microsoft vs. OpenAI anymore — it's everyone vs. distributed, open-source AI trained on Chinese data.

The Legal Mess — The Microsoft-OpenAI-Amazon Triangle

Here's where it gets messy.

In February 2026, Amazon invested $50 billion in OpenAI — $15 billion upfront and $35 billion in coming months — and OpenAI committed to using 2GW of AWS Trainium chip capacity ^[14].

On March 18, the Financial Times reported that Microsoft is considering legal action against OpenAI and Amazon, claiming the deal violates Microsoft's exclusive Azure cloud partnership ^[15].

The dispute hinges on terminology. Microsoft's exclusivity covers "stateless API calls" ^[16]. But OpenAI Frontier, the enterprise platform launching on AWS, would run on a "Stateful Runtime Environment" ^[16]. Microsoft claims this falls outside exclusivity. AWS says it doesn't.

Microsoft's statement was blunt: "Azure remains the exclusive cloud provider of stateless OpenAI APIs. We are confident that OpenAI understands and respects the importance of living up to this legal obligation" ^[15].

As of April 3, 2026, no lawsuit has been filed. But companies are reportedly in talks to resolve it before OpenAI Frontier fully launches.

Real Talk

Microsoft is telling two competing stories at once, and the market is confused.

Story One: "We're partnering with OpenAI. We hold a $135 billion stake" ^[6].

Story Two: "We're suing OpenAI (and Amazon) for violating our contract" ^[15].

Story Three: "We're also building our own models and don't need OpenAI anymore" ^[1].

These narratives don't resolve cleanly. If Microsoft's in-house models are truly competitive (and the benchmarks suggest they might be), then OpenAI's value to Microsoft diminishes. But if Microsoft's legal claim succeeds, it blocks Amazon's bet on OpenAI, leaving Microsoft as the de facto cloud provider for enterprise AI.

Here's the real tension: Microsoft renegotiated its partnership in October specifically to gain independence ^[6]. Then it spent six months hiring talent, building models, and aggressively pricing them ^[1][11]. Now it's fighting OpenAI over a $50 billion deal with AWS ^[15].

This doesn't look like a partnership anymore. It looks like a breakup that hasn't been announced.

Meanwhile, MSFT stock closed April 2 up 1.11% at $373.46 ^[17]. But the stock is down roughly 23% year-to-date ^[17], and the company just posted its weakest fiscal quarter since 2008 ^[18]. Investors want proof that AI spending delivers ROI — and three homegrown models launched in March and April won't be enough.

Q2 FY2026 revenues hit $81.3 billion, up ~17% year-over-year ^[19]. The earnings report comes April 28 ^[19]. That's when Microsoft has to explain how transcription models and image generation fit into a $2.76 trillion market cap ^[20].

The Bottom Line

Microsoft just declared independence from OpenAI while keeping its $135 billion investment. It's building models to compete with OpenAI and Google. It's potentially suing OpenAI and Amazon over a cloud deal. And Copilot's market share collapsed 39% in nine months ^[9].

If you own Microsoft stock or a tech-heavy ETF, the core question is: Did the October renegotiation and March reorganization actually work? Are in-house AI models with aggressive pricing ($0.36/hour for transcription ^[11]) enough to reverse Copilot's slide and justify the AI infrastructure spend?

Google is betting on open-source commoditization. Chinese models are racing up the quality curve. Amazon just paid $50 billion to partner with OpenAI. And Microsoft is at the center of all three battles — investor, competitor, and (potentially) litigant ^{[1][4][14][15]}.

The models look good. The story is a mess. The earnings call will be crucial.