Meta’s $145 Billion AI Gamble Hinges on a 29-Year-Old Outsider
CFO Susan Li warned the 29-year-old Scale AI cofounder that failure would trigger severe financial fallout as capital expenditures approach $145 billion.
Susan Li pulled Alexandr Wang aside at an executive retreat in Hawaii late last year. Meta’s chief financial officer didn’t mince words. If the artificial intelligence models Wang was hired to build failed, the financial damage would be catastrophic.
Li wasn’t speaking hypothetically. Meta has guided capital expenditures between $130 billion and $145 billion for 2026, nearly doubling the $72.2 billion it burned through last year. That money is largely flowing into data centers and silicon required to train frontier models—a bet that dwarfs the tens of billions Mark Zuckerberg previously sunk into the metaverse, a venture that cratered investor confidence and dragged down the company’s stock before he pivoted.
The person absorbing that pressure isn’t a career machine-learning researcher. Wang, a 29-year-old MIT dropout who co-founded the data-labeling firm Scale AI, had never trained a frontier model before Zuckerberg tapped him to run Meta Superintelligence Labs. The appointment bypassed Meta’s deep bench of AI PhDs and ignored the conventional wisdom of Silicon Valley lab leadership.
Zuckerberg reached out via WhatsApp in April 2025. Meta’s Llama 4 release had just landed to a tepid reception from developers and enterprise clients, exposing cracks in the company’s open-source strategy against closed competitors. A 44-minute phone call followed. Two days later, Wang sent Zuckerberg a memo outlining how he would restructure the division if he were in charge. His first demand: shrink the team. Talent density mattered more than headcount.
Securing Wang cost Meta $14.3 billion for a 49% stake in Scale AI. Once inside, Wang initiated what Zuckerberg called a “supernova” recruiting drive. Meta dangled compensation packages reportedly reaching $100 million to poach researchers directly from OpenAI. Zuckerberg cleared his calendar for two months to help close hires.
The resulting core research team numbers fewer than 200 people—Wang estimates closer to 100—operating within a company of roughly 75,000 employees. They shipped Muse Spark in April and released Muse Wonder on Monday, which Meta classifies as its first frontier-grade model.
Wall Street remains skeptical of the timeline. While the consumer-facing Muse app topped 2.5 million downloads on Apple’s App Store within two weeks of its Sept. 8 U.S. launch, according to Sensor Tower data cited by CNN, download volume does not offset infrastructure depreciation. Independent benchmarks reflect that gap. On Artificial Analysis’s Intelligence Index, the flagship Muse Wonder Full model trades blows with top systems from OpenAI and Anthropic but trails the overall leader by several points, per AI Weekly.
The monetization path for personal AI agents—software that handles tasks ranging from sourcing loose-leaf tea from farms in Taiwan and Japan to coaching mixed martial arts form—remains unproven at scale. Enterprise software buyers have shown willingness to pay for workflow automation, but consumer agent economics are largely theoretical.
Zuckerberg acknowledged the sheer size of the wager during a recent appearance on Cheeky Pint, the podcast hosted by Stripe cofounder John Collison. Discussing internal budget talks from late last year, he recalled asking Wang directly whether spending $100 billion or $200 billion would actually work. Wang’s mandate relies on the premise that the models must be undeniably superior to justify the balance sheet strain.
“Mark has fully internalized power laws in a way that many people have not,” Wang said of his boss.
