There's a line you need to know about Microsoft and OpenAI. For years, the relationship was described as a marriage.
Microsoft pumped in over $13 billion, hosted OpenAI's models across every product from Word to Azure, and agreed not to build competing frontier AI. The deal was so tight that Microsoft was contractually capped on the size of the model it could train. There was even a FLOP limit. A mathematical ceiling on how smart Microsoft's own AI could get.
That contract was renegotiated in October 2025. And this week, we saw exactly what Microsoft was waiting to do.
On Thursday, Microsoft's MAI Superintelligence team, which was formed just six months ago, released its first three foundational AI models:
MAI-Transcribe-1: Speech-to-text across 25 languages. 2.5x faster than Microsoft's own existing offering. Priced at $0.36 per hour
MAI-Voice-1: Generates 60 seconds of natural audio in one second. Custom voice creation from short audio snippets
MAI-Image-2: Image generation that already ranks top three on the global Arena leaderboard, now rolling out inside Bing and PowerPoint
The full multimodal stack, built in-house and cheaper than Google's and OpenAI's, is available today.
The man leading this is Mustafa Suleyman, co-founder of Google DeepMind, then founder of Inflection AI, then acquired by Microsoft in 2024. He came in with a single explicit mandate: "true self-sufficiency" in AI.
Here's what makes this genuinely significant.
Microsoft has 15 million paying Copilot users. OpenAI has 50 million paying ChatGPT subscribers. OpenAI projects 220 million paying subscribers by 2030. Microsoft's own AI product is losing the consumer race to the company it funded into existence.
Building proprietary models is how Microsoft stops that from being a permanent problem. Every query that runs on an MAI model rather than a GPT model is a margin point that Microsoft keeps. Every enterprise that chooses Azure for AI workloads gets a Microsoft-native option rather than a third-party one. Every product from Word to Teams gets smarter on Microsoft's own research, not a competitor's roadmap.
The October 2025 renegotiation explicitly allows both companies to "independently pursue AGI alone or in partnership with third parties." That clause is the permission slip. These three models are the first use of it.
For years, Silicon Valley told a simple story: Microsoft built the pipes, OpenAI supplied the water. That story is over. Microsoft is now drilling its own well.
Meta is building ten natural gas power plants for one AI data center. The carbon math is brutal.
Meta's Hyperion AI data center in Louisiana will, when complete, draw as much electricity as the entire state of South Dakota. To power it, Meta committed to building these natural gas plants generating 7.5 gigawatts in total.
Those plants will emit 12.4 million metric tons of CO2 per year. That's 50% more than Meta's entire global carbon footprint in 2024.
The man building the AI that replaces jobs is telling you how to survive it.
Dario Amodei sat down for an unfiltered interview in Bangalore. His career advice was honest, not reassuring. The safest career path isn't doctor, engineer, or coder. It's becoming someone who can genuinely work with AI, like a collaborator.
He also warned about de-skilling: using AI to skip hard learning leaves you unable to do hard things at all.
A startup just raised $60M to let AI design the chips that power AI.
Cognichip wants AI models to design better, faster semiconductors, and then use those to run better, faster AI. Chip design currently takes years and hundreds of engineers. Cognichip is betting AI can compress that to months.
If it works, the AI hardware race stops being limited by how many human chip designers exist. It becomes limited only by how fast the AI can design the next version of itself.
FAST BREAK
A developer ported Claude Code's leaked architecture to Python in a single night. The repo hit 30,000 GitHub stars faster than any repository in history.
Anthropic spent years and hundreds of millions of dollars building what one developer could replicate in one sleepless night. And that too using an AI tool, in a different programming language, in a way that is completely legal.
The moat isn't the code. It was never the code. The moat is the trust, the distribution, and the name on the door.
Anthropic still has all three, but they had a very bad week.