The OpenAI Leash Is Off — Has Been for Six Months

Since tracking how corporate America is souring on the 'buy every AI model' strategy, the Microsoft piece of that puzzle has come into focus.

Microsoft AI CEO Mustafa Suleyman told VentureBeat at Build 2026 that "We were only sort of set free from our contract with OpenAI about six months ago to formally pursue superintelligence." That means the separation began around December 2025 — not this week.

$13 Billion In, and Now Building Around It

Microsoft has poured a cumulative $13 billion into OpenAI, according to VentureBeat. That investment handed the company early model access and added hundreds of billions to its market cap. For three years, Microsoft's AI story WAS the OpenAI story.

Now Microsoft is building something parallel. Suleyman's team announced seven new in-house AI models under the MAI brand. The flagship — MAI-Thinking-1 — is a 35-billion-active-parameter reasoning model. Microsoft says it matches leading models in its weight class on software engineering benchmarks.

Suleyman keeps repeating: "We train our reasoning models from scratch. We don't distill from other labs." That's a shot across the bow at half the AI industry, which trains cheaper models by feeding them outputs from more expensive competitors.

What Microsoft Is Not Saying

Suleyman called this "very early days." That's honest. Seven new models and a bold mission statement don't make Microsoft an AI frontier lab overnight.

The Build keynote was full of agentic AI optimism — Microsoft announced Scout (a personal work agent), Microsoft IQ as a context layer across GitHub Copilot and Foundry, and Work IQ APIs dropping June 16. But the company has a developer credibility problem it's sidestepping.

GitHub Is Bleeding Trust — and Microsoft Knows It

Wired's interview with Microsoft VP Scott Hanselman surfaced the dark side of this week's announcements: GitHub, Microsoft's code repository subsidiary, has suffered significant downtimes. Longtime developers are complaining publicly. One Reddit post asked bluntly: "Has GitHub become a dumpster fire?"

Hanselman's explanation? Bots. He told Wired that GitHub traffic is now "as many bots as people," driven by AI coding agents hammering the platform at machine speed. He called it a "hiccup moment."

Dismissing developer frustration as a hiccup is the kind of corporate spin that can turn small problems into big ones. Steve Ballmer's "Developers! Developers! Developers!" wasn't a joke — it was Microsoft's survival formula. Losing developer trust while simultaneously canceling Claude Code licenses to push coders onto Microsoft's own Copilot is a risky combination.

Microsoft ended its Claude Code licenses to push developers toward Copilot, according to Wired. That's a forced migration, not an earned one.

The Copilot Adoption Problem

Microsoft's workplace AI products — all branded Copilot — have had disappointing enterprise uptake. Meanwhile Anthropic grabbed the agentic coding lead that Microsoft once held. These are ongoing failures that a seven-model announcement doesn't automatically solve.

Marco Casalaina, Microsoft's VP of Core AI, told VentureBeat that the Build announcements are about giving enterprises governance, memory, identity, and context — not just raw model power. That's the right diagnosis. Enterprises don't need smarter AI. They need AI that fits into how teams work.

The Real Problem: AI Agents Don't Share What They Learn

According to Asana Chief Product Officer Arnab Bose, 75% of knowledge workers are using AI — but only 5% of companies have reported actual productivity gains. That gap isn't a model quality problem. It's a memory problem.

When one team member corrects an AI agent, that correction vanishes when a colleague opens the same tool. Every user trains a different version of the same agent — and those versions never sync. In multi-agent enterprise workflows, agents can contradict each other.

Sriharsha Chintalapani, CTO of Collate, put it plainly: agents are sensitive to prompt quality, and without shared memory architecture, only people who are good at prompting get good results. Everyone else gets poor performance.

Microsoft's IQ stack — Microsoft IQ, Work IQ, Fabric IQ, Foundry IQ, Web IQ — aims directly at this problem. Whether the architecture delivers shared context at enterprise scale remains to be seen. It was announced this week. The Work IQ APIs don't launch until June 16.

What This Means for Regular People

If you work at a company that deployed AI tools, the situation is likely this: the tools are probably helping some people on your team and doing nothing for others. The reason is shared memory, not model intelligence. Your IT department likely has no strategy for this.

On the Microsoft stock front, the company's shares are down this year while competitors have soared, according to Wired. That's the market's take on promises versus results.

Microsoft has a real plan now. It also has real problems. Build 2026 addressed the plan. The problems are still there.