On a sunny Seattle day in late September, about 50 executives representing large Microsoft customers got to ask Satya Nadella anything. One asked what Microsoft could do to help companies adapt to the AI age, and Nadella’s answer — delivered, per CNBC, while striding the stage of the Paccar Pavilion in a tight black crewneck — was “diffusion inside the enterprise,” meaning the tools need to be deployed widely, everywhere, by everyone. “Like all things, change is always bottom up, top down, and middle up,” he said. Which is a tidy summary of almost everything, and also of Microsoft’s pitch: the AI revolution will be sold as a site license.

The 59-year-old Nadella, who has spent more than half his life at Microsoft, is making this case at an awkward moment. Microsoft is worth almost $4 trillion, yet by common verdict it is not the AI leader. Nvidia owns the chips. Anthropic and OpenAI have the foremost models. Meta just shipped the first mass-market personal agent. Google’s Gemini services are popular with businesses and consumers. Microsoft was aggressively early — it first invested in OpenAI in 2019, and released GitHub Copilot with the lab in 2021, a year before ChatGPT — but the alliance frayed. Last year Microsoft lost its status as OpenAI’s exclusive cloud provider; this year its license to OpenAI’s intellectual property became nonexclusive. Microsoft said in April it remains OpenAI’s “primary cloud partner,” which is the sort of title you mention when the exclusive one is gone. OpenAI, for its part, is out raising at a valuation of about $1.4 trillion, so the breakup is treating at least one party well.

Wall Street has noticed. Even after a sharp third-quarter rally, Microsoft’s stock is up just 7% this year as of Friday’s close, trailing its megacap peers and the broader Nasdaq; it has not beaten the S&P 500 in a calendar year since 2023. “I think there needs to be a little bit more maturity of the product before we figure out actually what the best solution or what pricing is,” said Tiffany Wade, a co-portfolio manager at Microsoft shareholder Columbia Threadneedle, about Copilot — the centerpiece of the whole strategy.

Longtime investors give Nadella the benefit of this particular doubt because he has done the trick before. He became CEO in 2014, succeeding Steve Ballmer, under whom the stock had fallen roughly 33% over 14 years as Microsoft slid into the legacy-software bucket. Nadella ditched perpetual Office licenses for subscriptions, built Azure into a cloud powerhouse, let customers plug in products from bitter rivals like Oracle and Salesforce, and bought Mojang, LinkedIn and GitHub. The stock is up roughly 14-fold on his watch.

The new pivot, which Nadella has been acknowledging to investors in recent months, is a商业模式 change from subscriptions to consumption — Microsoft becoming a vendor of huge volumes of tokens, the units of computing usage that serve as the currency of the AI economy. Success, he said on the July earnings call, means “further expanding our TAM.” The growth story he preaches is software production: more code, written faster. “We went from the code assist to the agentic coding to software factory all within one year,” he told the Seattle crowd, and he wants customers opening software factories of their own — plus, for the nontechnical masses, a Copilot “super app” that his lieutenants demoed as a kind of virtual teammate. Sometime between now and Wednesday’s Windows and Surface event in San Francisco — billed as “a conversation on how local AI will shape the next chapter of the PC” — someone will explain all this to the person who still just wants Excel.

The encouraging precedent, from Microsoft’s point of view, is GitHub Copilot. After losing share to AI-native rivals — Cursor, now owned by SpaceX, and Anthropic’s Claude Code — Microsoft repriced the tool earlier this year to track computing demand. Customers balked at first. Then usage followed: 50 million users by July, up from 26 million last October, or one in five people on GitHub. Copilot sales growth accelerated 60% from the prior quarter, Nadella said in July, with “significant consumption revenue after the new model went into effect.” “GitHub Copilot went out and learned lessons,” said the unit’s operating chief, Kyle Daigle, and those lessons are now being handed to other Microsoft teams designing consumption pricing. Hayete Gallot, the security-products EVP, has internalized it as a doctrine: “I always say to my team, when customers consume, it means we add value. And for me, the truth starts with customers. If they don’t consume, it means we’re not giving value.”

Note what is being replaced here. The incumbent Microsoft model is the annuity: multi-year contracts for Microsoft 365, now more than 450 million commercial Office seats, with Copilot as a $30-per-person monthly add-on that has sold 30 million seats. Lovely, predictable revenue, whether or not anyone opens the app. The consumption model flips the risk: Microsoft gets paid only when people actually use the stuff. Gallot frames that as customer-first truth-telling, and it is — it is also Microsoft voluntarily giving up the coziest business arrangement in software history in exchange for a bigger TAM it has not yet captured.

The other uncertainty is Mustafa Suleyman, the DeepMind co-founder Nadella hired in 2024 to lead AI as the OpenAI relationship cooled. His group has shipped coding, image-generation and reasoning models, and this week he boasted on X that a new speech-to-text model was “the most accurate real time transcription model in the world... #1 !!!” Independent leaderboards still rank Anthropic and OpenAI ahead for raw digital intellect. Microsoft executives argue co-location helps — the Office apps group gets better transcription models at lower cost, and engineers get direct feedback. “The collaboration is much tighter with Mustafa,” said EVP Charles Lamanna. Though Suleyman’s remit has shifted: in March, with Copilot adoption sluggish, Microsoft merged its commercial and consumer Copilot engineering teams, promoted Jacob Andreou to EVP for Copilot experience, and told Suleyman to focus on model development.

Nadella, who has lately been posting on X about the need for “deliberate pacing” in model development and reassuring the All-In Summit about safety, has started downplaying whether Microsoft needs the world’s best model at all. This is either graceful positioning or the truth. If the really scarce thing turns out to be distribution — 450 million seats, an enterprise sales force, a cloud full of rented chips — then diffusion, bottom up and top down and middle up, is not a slogan. It is the moat.