Satya Nadella Is Coding Again — And That Tells You Everything About How Badly Microsoft Needs to Win the AI Race

When a CEO of a $3 trillion company starts pulling all-nighters with his engineering team, showing up at whiteboard sessions unannounced, and building working prototypes himself using AI tools, it means one of two things. Either the culture is exceptionally healthy. Or the situation is exceptionally urgent.

At Microsoft, in the first months of 2026, it was emphatically the second. Nadella walked into Building 92 on Microsoft's Redmond campus in January and opened a laptop in front of a team of engineers working on a new AI personal assistant product. On it was a working application he had built himself — a system for coordinating multiple AI agents, which he called "Chain of Debate." He had vibe-coded it using AI tools. The engineers, according to those in the room, exchanged the kind of looks that said: the boss is serious. "That set the tone for how hard the team was going to push," said Jacob Andreou, the executive vice president now leading Microsoft's unified Copilot experience.

The urgency is real and the numbers explain it clearly.

HOW MICROSOFT LOST ITS EARLY LEAD

Two years ago, Microsoft looked like the early winner of the AI era. Its prescient investment in OpenAI — eventually totalling $13 billion — gave it exclusive access to the world's most powerful AI models. Azure was the only cloud where companies could access OpenAI's technology. The competitive advantage seemed insurmountable.

Then several things went wrong simultaneously.

The OpenAI partnership — the source of Microsoft's advantage — became its constraint. OpenAI's explosive growth after ChatGPT created friction on almost every dimension: computing capacity, intellectual property sharing, enterprise customers, and eventually equity in OpenAI's restructuring. The relationship that made Microsoft looked exposed as the relationship Microsoft could not fully control.

Meanwhile, the model landscape changed underneath Copilot. By early 2025, Anthropic's Claude was consistently outperforming GPT on complex tasks, and Google's Gemini was competitive on visual applications. But Microsoft's products were powered almost exclusively by GPT. What had been an engine became a constraint.

The consumer products also disappointed. A year into Mustafa Suleyman's tenure leading Microsoft's AI division, Copilot's weekly active users had flatlined at around 20 million — while ChatGPT raced to 900 million. GitHub Copilot, once the dominant AI coding tool, was supplanted first by Cursor and then by Anthropic's Claude Code. Less than 4.5% of Microsoft 365's 450 million customers pay for Copilot features. The stock fell 34% from its October 2025 high over five months — including a single-day wipe of $357 billion in market cap when Anthropic launched Claude Cowork, an agent that could autonomously use Microsoft's own productivity tools.

THE RESET: WHAT NADELLA IS ACTUALLY DOING

The response has involved several moves made in rapid succession. Nadella handed most of his commercial and operational duties to Judson Althoff, Microsoft's commercial CEO, freeing himself to focus exclusively on AI product development and infrastructure. He merged the consumer and enterprise Copilot teams — which had been operating separately and creating market confusion — under Andreou, who reports directly to Nadella. Suleyman was repositioned to lead a rebranded "Superintelligence" team focused on building Microsoft's own frontier models — a hedge against the day, in 2032, when the OpenAI access agreement expires.

Most importantly, Microsoft rewrote the rules of its OpenAI partnership. The old exclusivity arrangement — OpenAI only on Azure, Azure only for OpenAI — was abandoned. Microsoft secured a 27% equity stake in OpenAI for the reduced restrictions. And it immediately used that freedom to commit up to $5 billion to Anthropic and bring Claude models to Azure. Copilot Cowork — the enterprise agent product that can autonomously use software tools — runs on Claude. This is the same model that was threatening Microsoft weeks earlier.

The new strategy is model-agnostic by design. Microsoft's bet is that AI models will increasingly commoditise, and that the real enterprise value will accrue not in the AI brain but in the infrastructure, workflows, security, and data layer around it — all of which Microsoft already owns. "We don't think enterprises will swap out their information work platform every time a new model drops," Althoff says.

A new pricing model follows the logic: per-user licensing with token allowances, plus consumption-based charges beyond those limits — so that using third-party models does not destroy Microsoft's margins.

THE NUMBERS THAT SUGGEST IT IS WORKING

The early evidence is encouraging. Azure revenues are growing 40% year-over-year. Microsoft's total AI business is on pace for $37 billion in annual revenue — up 123% from the prior year. A quarter of the 20 million M365 Copilot paying subscribers were added in the first four months of 2026 alone, with Althoff describing adoption as accelerating.

But the overall numbers are still disappointing relative to the ambition. Less than 5% penetration of the 450 million Microsoft 365 base does not satisfy Wall Street. The capex commitment — $190 billion in 2026, more than triple 2024's level — creates a financial exposure that requires the revenue growth to actually materialise.

The broader competitive risk is also not resolved. OpenAI launched its Frontier enterprise platform in February — offering capabilities that directly compete with Microsoft's connective-tissue strategy. Anthropic is building Claude Managed Agents. The AI-native startups are not just making the brain — they are beginning to build the body parts that Microsoft thought it owned.

Microsoft's argument, as articulated by Bank of America analyst Tal Liani, is that it does not need to win the AI race — it just needs to not lose it. The bundling value, the enterprise relationships, the security reputation, and the deep product integration make it very difficult to displace even if a competitor has a better AI model on any given benchmark. Being good enough, across everything, has historically been one of Microsoft's greatest competitive advantages. The question is whether that principle holds when the pace of change is faster than Microsoft has ever moved.

Nadella's own answer to that question is visible in his behaviour. He is at the whiteboard with the engineers. He is attending weekly Copilot stand-up meetings. He described the current moment, at the Miami offsite for the Superintelligence team, as Microsoft "refounding the company." Whether that language reflects genuine conviction or manufactured urgency will be answered by the product roadmap of the next twelve months.

One product, at least, is already in build: a Microsoft version of the open-source autonomous agent tool OpenClaw — the always-on AI that works independently across systems and tasks. Nadella has personally tasked the Copilot team with building it with enterprise-grade security. If it lands, it would transform Copilot from an AI assistant into something considerably more consequential — a permanent, autonomous AI presence in every workplace that runs on Microsoft's infrastructure.

That, if it works, is what a comeback looks like.