Mark Zuckerberg once poured tens of billions of dollars into a metaverse that few people wanted to inhabit. Now he is spending even more on Muse, Meta's personal AI agent—an assistant that, in Zuckerberg's case, coaches his MMA form between rounds, and, in Alexandr Wang's, hunts down obscure loose-leaf teas from farms in Taiwan and Japan.

Those are two of the favorite use cases cited by Zuckerberg and Wang, the 29-year-old MIT dropout and Scale AI cofounder whom Zuckerberg tapped last year to lead Meta Superintelligence Labs. The Muse app launched in the U.S. on Sept. 8 and quickly became the most popular free app on Apple's App Store, topping 2.5 million downloads in roughly two weeks, according to Sensor Tower data cited by CNN. On Monday, Meta followed up with Muse Wonder, the first model from the lab that Meta describes as frontier-class—a term the industry uses for models competing at the leading edge of capability.

“We're going to write this check for like, I don't know, like $100 billion or $200 billion,” Zuckerberg said of the discussions held late last year, appearing alongside Wang on the latest episode of Cheeky Pint, the podcast hosted by Stripe cofounder John Collison. “I'm like, Alex, you sure this is going to work?”

The figure was off the cuff, but the scale is real. Meta now expects capital expenditures of $130 billion to $145 billion in 2026, nearly double the $72.2 billion it spent last year—and investors have flinched at each increase. At that scale, the wager is as much a balance-sheet decision as a research one.

Wang said he heard a version of Zuckerberg's question from Meta CFO Susan Li at an executive gathering in Hawaii at the end of last year. “Susan pulled me aside many times,” Wang said. “I just want to tell you if this doesn't work, this is going to be really, really bad. It's going to be very, very bad in financial terms. I really need to impress on you that this model has to work. It has to be good. If it's good, it's all good. We're great. But if it's bad, it's really, really, really bad.”

The 44-minute call and the memo

Between sips of Guinness—a moment of apparent wincing by Zuckerberg that went viral—the pair recounted how their partnership began. In April 2025, shortly after Meta's Llama 4 models landed to a lukewarm reception, Zuckerberg started reaching out for advice; Fortune previously reported that the two began spending more time together that month. Wang was in a staff meeting when his phone pinged.

“I got the WhatsApp notification, like, 'Have time for a call?'” Wang said. “I was like, huh, interesting.”

That call ran about 44 minutes, by Zuckerberg's recollection. Two days later, Wang sent him a long memo laying out, as Zuckerberg put it, “here's what I think I would do if I were in your position.”

The first principle was talent density. “The very thing is that you don't need a lot of people, and talent density is the thing that matters,” Zuckerberg recalled. The second was a willingness to start from scratch. Zuckerberg said Meta was “not happy with the trajectory that we were on with Llama 4” and that he was prepared to rebuild the team. He needed someone who could recruit the best researchers and convince them to leave their labs.

Zuckerberg said he had been quietly impressed by Wang for years. “It always felt to me like he was the sharpest entrepreneur of his generation, of the ones who I had had a chance to meet,” he said.

The pick still raised eyebrows. Wang had built Scale AI into a data-labeling powerhouse—the unglamorous but essential work of preparing the data AI models are trained on—but as Collison pointed out on the show, he had never trained a frontier model. When Meta named him chief AI officer, Fortune noted that the company had chosen someone who is not a computer scientist to lead all of its AI efforts—a choice that might not sit well with its deep bench of AI PhDs.

Zuckerberg was not swayed. “I talked to like all the leaders of all the labs, like all the heads of research,” he said. “And I was like, no, Alex is the guy.”

The 'supernova' recruiting push

What followed was a recruiting drive Zuckerberg described as “supernova”—a frantic, months-long effort to assemble a small, elite team from inside Meta and poach top researchers from rivals. Meta paid $14.3 billion for a 49% stake in Scale AI to bring Wang aboard and reportedly dangled $100 million compensation offers to OpenAI researchers—figures that reflect a broader industry scramble for the small pool of researchers with experience at the frontier of AI. Zuckerberg cleared his calendar for two months. “I don't think either of us slept that much for the first couple of months,” he said.

The core research team Wang built inside Meta Superintelligence Labs numbers fewer than 200 people—“closer to 100,” Wang said—supported by larger infrastructure and product teams at a company with roughly 75,000 employees. It is the memo's first principle in practice: a handful of elite researchers rather than an army. Wang said he cultivated a culture of “insecure overachievers,” researchers who consistently under-promise and over-deliver.

That bet produced Muse Spark, the lab's first model, in April, followed by a string of updates—and now Muse Wonder, which Wang said is state-of-the-art “based on our measurement” for personal and professional agent tasks. Independent scorekeepers are a bit more measured. On Artificial Analysis's Intelligence Index, the flagship Muse Wonder Full lands roughly level with top models from OpenAI and Anthropic but a couple of points behind the leader, according to AI Weekly. As Meta ships further updates, those independent rankings—together with the download counts behind Muse's momentum—are the public yardsticks against which the bet will be measured.

Wang also addressed Zuckerberg's distinctive qualities as a leader. “Mark has fully internalized power laws in a way that many people have not,” Wang said. “Once he internalizes certain beliefs, the speed and magnitude of willing to change prior actions and pivot towards the worldview that is consistent with those beliefs is pretty unusual.”

This story was originally featured on Fortune.com.