Satya Nadella has made Microsoft 10 times more valuable in his decade as CEO.

Llama is the AI model, not the animal. It is a piece of open-source software created by Meta, the social media giant that has pivoted hard to AI and is competing with Microsoft and others to dominate the foundations of the emerging generative AI economy. “Both” is a reference to the fact that this Malaysian agriculture technology company—chosen to show off its use of Microsoft’s technology to Microsoft’s CEO— is using Meta’s rival AI model in addition to GPT-4, the large language model (LLM) created by Microsoft’s strategic partner, OpenAI. Nadella wants his Redmond, Wash.–based software giant to have the most capable, popular AI models on the market. 

“Um, yep, so we are, we are using Llama also,” says Adrian Lee, the chief technology officer at Agroz, the Malaysian startup, a hint of embarrassment in his reply. Agroz, which builds hydroponic farms, has created an AI chatbot to answer farmers’ questions about how best to tend their lettuce and bok choy.  

“What are you using Llama for?” Nadella asks pointedly, standing before Lee at Agroz’s exhibition kiosk.

Agroz, Lee explains, eventually wants to use humanoid robots for farming, and the robots may need to operate offline. Some versions of Meta’s Llama model are compact enough to be embedded in robots or phones, unlike the larger GPT-4. 

“Take a look at Phi,” Nadella says, switching from CEO-as-chief-market-researcher to CEO-as-chief-salesman. Phi is a family of small language models Microsoft has built in-house and offered, like Llama, as open-source software. Nadella apparently knows the parameter count of the smallest Phi model off the top of his head, and he rattles it off—just 3.8 billion parameters, making it small enough to run “on the edge.” 

Annice Lyn—Getty Images for Microsoft

Lee says Agroz will experiment with Phi. Meanwhile Agroz’s CEO Gerard Lim offers Nadella some Agroz-grown bok choy. “Oh, you want me to eat it?” Nadella says. He takes a small piece. “Mmmm,” he says, chewing vigorously, and, no doubt, chewing over the lesson about how perilously competitive the AI race is. 

It’s a telling exchange. Few companies have benefited as much from the generative AI boom as Microsoft. Investor fervor for the technology has helped make Nadella’s company a perennial contender for the title of the world’s most valuable corporation, with a market cap that consistently bobs somewhere in excess of $3 trillion. 

When Nadella took the reins at Microsoft in 2014, the company was floundering. Under his predecessor, Steve Ballmer, it had missed the smartphone revolution, was lagging on tablets, and was even losing market share in the PC operating system business that made it a household name. Microsoft’s shares had fallen more than 40% during Ballmer’s tenure. In 10-plus years as CEO, Nadella has reinvigorated the company, successfully steering it through two technological transformations: from PCs to the era of cloud computing, and now, to the age of AI. Nadella’s prescient and early bet on OpenAI and its technology, and the fruitful-if-sometimes-tense relationship that ensued, have given Microsoft as good a shot as any company at supremacy in this new era. The company arguably hasn’t been this powerful since it dominated the PC market in the 1990s’ “Wintel” era. 

But as Nadella begins his second decade at the helm, there’s no guarantee Microsoft will retain its lead. Regulators, hackers, and rivals each present threats serious enough to potentially undermine its industry leadership. Above all, it must reckon with its own leviathan size—and avoid becoming a victim of bureaucracy and bloat when AI’s protean nature calls for speed, agility and finesse.

Nadella’s approach to leadership reflects his acute awareness of these risks. Even when seemingly ahead of the pack on AI, he and his team are constantly listening, their corporate antennae sensitive to the tiniest pivots in users’ needs and preferences. Microsoft is continually investing in technologies and talent that might someday supplant OpenAI’s models—or even supplant generative AI altogether. In other sectors, such hypervigilance might seem like overkill,…