Meta Paid $14.3 Billion to Put Scale AI's Alexandr Wang in Charge

Meta spent $14.3 billion for a 49% stake in Scale AI — and used the deal to put the data-labeling company's 29-year-old founder, Alexandr Wang, in charge of its entire artificial intelligence effort. The bet pairs one of the largest AI capital outlays in corporate history with a research team of fewer than 200 people.

The hire traces back to a 44-minute WhatsApp call between Wang and Mark Zuckerberg in April 2025, first reported by Fortune and detailed on a podcast hosted by Stripe co-founder John Collison. The call came shortly after Meta's Llama 4 models drew what the company acknowledged was a lukewarm response; Zuckerberg admitted the company was unhappy with their direction. Two days later, Wang sent Zuckerberg a memo. Its first principle, Zuckerberg recounted: "you don't need a lot of people, and talent density is the thing that matters." The second: a willingness to rebuild from scratch.

How small is the team behind the models?

Very small, by Big Tech standards. Wang said the core research team at Meta Superintelligence Labs is closer to 100 people; Zuckerberg put it under 200. Both figures sit inside a company of roughly 75,000 employees, with larger infrastructure and product teams supporting the lab.

The recruiting push that built it was anything but lean. Zuckerberg described it as a "supernova" effort that included reported $100 million compensation packages aimed at OpenAI researchers. He cleared his calendar for two months. Neither executive slept much, he said.

What has the lab actually shipped?

Three models in roughly seven months. The first, Muse Spark, arrived in April. On November 9, Meta released Muse Wonder, which the lab calls its first "frontier-class" model. Wang said it is state-of-the-art on his team's measurements for personal and professional agent tasks. On the Artificial Analysis Intelligence Index, the flagship version lands "roughly level with leading models from OpenAI and Anthropic but a couple of points behind the leader," according to AI Weekly.

The same day, Meta launched Muse Ember, a smaller model priced at $0.10 per million input tokens and $0.50 per million output tokens, positioned for high-volume, repetitive agent workflows. Meta's own benchmark tables show Ember leading on several agent and coding benchmarks against Anthropic's Claude Haiku 5.5 and OpenAI's GPT-6 Luna, though it trails on FrontierCode 2.0.

One caveat runs through all of it: the cited comparisons come from Meta and its chosen evaluation partner, Artificial Analysis. No independent third-party evaluation appears in the available reporting. Wang's "state-of-the-art" claim rests on his team's own measurements, and the Artificial Analysis Index places Wonder near, not at, the top.

Are consumers and cloud buyers biting?

Early traction is real. The Muse app, launched in the U.S. on September 8, became the most popular free app in Apple's App Store and topped 2.5 million downloads in about two weeks, per Sensor Tower data cited by CNN. On November 10, Google Cloud announced it would host Muse Ember, Muse Spark 1.3, and Muse Wonder via its Model Garden managed APIs. Google CEO Sundar Pichai replied to Zuckerberg's post: "Congrats and welcome to the model garden, excited to serve our Google Cloud customers!"

The financial pressure is explicit. Meta expects 2026 capital spending of $130 billion to $145 billion — nearly double 2025's $72.2 billion. That guidance is a forecast, not committed spend; the Scale AI stake, by contrast, is a closed deal. Wang said CFO Susan Li pulled him aside at an executive gathering in Hawaii to stress that "the model has to work, because failure would be very bad in financial terms."

What's the bet, exactly?

Zuckerberg is wagering that a small, elite team — led by an entrepreneur who built a data-infrastructure business but had never run frontier model training — can outpace larger, better-established labs. Wang is wagering that his management philosophy plus Meta's compute scale can justify the spending. Talent density is a stated operating principle, not a proven formula; three models in seven months is a data point, not a validation. Both bets remain unproven, and 2026's capex bill comes due regardless.