Microsoft just built the off-ramp from OpenAI

THE AI BRIEF

Today's signal: Microsoft shipped its first frontier reasoning model trained without OpenAI. Plus Anthropic's S-1, Dell's 88%, GitHub Copilot's desktop app.

In today’s issue:

• Main story: Cognition is at $492M run-rate, not just $26B.

• Also worth knowing: Anthropic filed a confidential S-1 at a reported $965 billion valuation, Dell posted 88% year-over-year revenue growth on AI server demand, GitHub launched an agent-native Copilot desktop app at Build, and more.

THE READ

MAI-Thinking-1 is Microsoft's first reasoning model trained from scratch, with zero distillation, running on its own Maia 200 silicon. The interesting part is what it lets a buyer do next.

At Microsoft Build this morning, Mustafa Suleyman's team released MAI-Thinking-1, Microsoft AI's first in-house reasoning model. It is 35 billion active parameters with a 256K context window, trained from scratch on what Microsoft describes as "enterprise-grade, clean and commercially licensed data." Microsoft says independent raters prefer its outputs to Claude Sonnet 4.6 on a blind test, and that it matches Claude Opus 4.6 on SWE-Bench Pro for coding. It is in private preview on Foundry today and runs on Microsoft's Maia 200 chip, which the company says delivers 1.4x performance per watt over NVIDIA's GB200.

MAI-Thinking-1 is one of seven new first-party models Microsoft announced today. MAI-Image-2.5, MAI-Transcribe-1.5, MAI-Voice-2, and MAI-Code-1 round out the family, with the image model live in PowerPoint and the code model live in GitHub Copilot. Suleyman's framing is that each one outperforms a Gemini equivalent on its benchmark. The numbers are Microsoft's. They will get scrutinized fast.

The reason this matters for a buyer is not the benchmarks. It is the structure. Microsoft now operates a vertical stack: Maia silicon at the bottom, the MAI models above it, Foundry as the runtime, GitHub and Microsoft 365 as the distribution. A year ago, the question inside most enterprise IT shops was "how exposed are we if OpenAI prices change?" The honest answer was "completely." Today, the answer is more complicated. The same Foundry that serves OpenAI now serves MAI, plus the new general availability of Fireworks AI inside Foundry, plus Anthropic models, plus open weights from Mistral and others. Microsoft is not telling customers to switch. They are removing the reason a customer would have to switch providers to switch models.

AI READY PRO · FREE UNTIL FRIDAY

I will get you AI trained in 30 days for free

After several months of running it quietly with top AI operators and teams, I’m excited to launch AI Ready Pro today to our newsletter subscribers like you.

It’s a 30-day personalized AI training program that is personalized to you and where you are in your AI journey. It’s the culmination of 100’s of hours spent teaching AI to top Fortune 500 AI teams and operators (and working with Mark Cuban).

Each day, you receive a 10-15 minute exercise to complete to improve your AI skills.

This isn’t vague AI theory or a way to pitch you a tool. It’s 30 days of learning to use AI in your real work. So you come out the other side with the skills to use AI to actually improve you and your teams output.

Until this Friday (May 15), it’s free for newsletter readers who complete the extended assessment. This assessment will help us personalize the learning experience to you.

If you lead a team, AI Ready Team is open today too, also for readers first. It’s the same engine, but to train your whole organization. You’ll get a detailed view of:

• Your team’s AI readiness level + detailed strengths & weaknesses

• Ranked list of AI automation opportunities

• Get tactical advice on how to advance AI efforts in your org

Leading a team? Take the Team assessment instead.

What I keep hearing in these conversations is that the lock-in question has moved one level up the stack, from a model question to a platform question. If your Microsoft 365 contract is the surface where every agent lives, the model underneath the agent becomes a runtime decision rather than a vendor decision, which is a different sales conversation than the one most CIOs walked into 2026 expecting.

The bear case is straightforward and worth saying out loud. MAI-Thinking-1's benchmarks come from Microsoft, and the blind-rater preference over Sonnet 4.6 is the kind of claim that often gets walked back when the third-party evaluations land. Microsoft also has a history of declaring model parity that does not survive contact with production workloads, so the thing worth tracking over the next two quarters is whether enterprise agents actually deploy on Foundry start defaulting to MAI rather than to OpenAI, because that is the data point that tells you whether the vertical stack play is working in the field.

SPONSORED BY CLUTCH

Hire secure AI teammates that work 24/7.

Hire pre-built AI teammates. Give your engineers and operators a platform to ship their own AI apps. Stop losing sleep about what is running where.

Clutch is the platform behind both: pre-built agents for the workflows your ops team should automate first, plus the integration plane your team's vibe-coded apps and Claude Code projects plug into. One platform. Real production. Visible and safe by default.

Built for ops, engineering, and security teams that are tired of the shadow-AI surface area inside their own company.

ALSO WORTH KNOWING

Anthropic filed a confidential draft S-1 with the SEC. The filing is the first formal step toward an IPO, with the estimated valuation at the time of filing around $965 billion. Financial details remain sealed until Anthropic chooses to make the filing public, which under SEC rules can happen as late as 15 days before the roadshow.

Dell reported Q1 revenue of $43.8 billion, up 88% year over year. AI server demand is the primary driver, with non-GAAP earnings of $4.86 per share. HPE reported a $5 billion AI systems backlog the same day, sending the stock up 21%. The AI infrastructure trade now has two legacy hardware names confirming sustained demand.

GitHub launched an agent-native Copilot desktop app at Build. The app surfaces active coding sessions, pull requests, issues, and background automations in a single My Work view. It is GitHub's first desktop product built around agent workflows rather than around the repository.

Senator Bernie Sanders introduced the American AI Sovereign Wealth Fund Act. The bill would impose a one-time 50% stock tax on the largest American AI companies, including OpenAI and Anthropic, with the resulting equity held in a public fund. It is unlikely to pass in the current Congress, but it reframes the policy debate around public ownership of frontier labs.

DigitalOcean became an AI model provider on OpenRouter. The developer-focused hosting company now serves DeepSeek V3.2, Kimi K2.6, and DeepSeek V4 Flash. The signal is that the inference layer is broadening beyond hyperscalers and pure-play AI shops.

Nathan Lambert is leaving Ai2 after two and a half years. Lambert led post-training at the institute and built OLMo and Tulu, two of the most prominent open-source model releases of the last two years. His next move has not been announced.

WATCHING TOMORROW

Microsoft Build runs through Thursday. The Tuesday keynote leaned on the platform and model layer; the Wednesday and Thursday sessions are where Agent 365, Microsoft Execution Containers, and the Frontier Tuning private preview get a deeper walkthrough. The question worth watching across the rest of the week is which named enterprise customers Microsoft brings on stage for MAI-Thinking-1.

Back tomorrow,
Haroon