Zuckerberg admits: Meta committed $100–200 billion to AI before any model was proven
The admission came on a podcast, just one day after Meta launched the models now meant to justify the bet. But the numbers meant to prove the models compete with OpenAI and Anthropic are Meta's own.
Mark Zuckerberg admits that Meta committed to AI infrastructure worth between $100 and $200 billion in late 2025 — before the company had succeeded with any models. That is the core of what he said on the Cheeky Pint podcast, which was posted on October 13, 2026, and which Bloomberg reported the same day.
"We're going to write this check for like, I don't know, $100 billion or $200 billion," Zuckerberg said on the podcast, according to Bloomberg. "You sure this is going to work?" he recalls asking Alexandr Wang, Meta's chief AI officer.
The irony of the timing
The remarks came in the same news cycle as something else: on October 12, 2026, Meta launched the models Muse Wonder and Muse Wonder Full — the very models the company now points to as proof that the bet was right. Bloomberg's headline sums up the admission: Zuckerberg admits Meta lacked certainty when the bet was made.
The backdrop is a capital commitment few companies have ever come close to. According to Crypto Briefing, citing Bloomberg, the decision was made in late 2025. That same year, Meta had bought 49 percent of Scale AI for $14.3 billion to bolster its AI push. The money is spent whether or not the models turn out well — the infrastructure has to be built in advance.
What the money has built
The bet goes to what is called gigawatt scale: according to Crypto Briefing, Meta is building data centers such as Prometheus in Ohio and Hyperion in Louisiana, and the company is said to have secured nuclear power contracts equivalent to 6.5 gigawatts through 2035. The capex trajectory is dramatic if the numbers hold: Crypto Briefing puts Meta's capital expenditure outlook for 2026 at $130–145 billion, versus roughly $72 billion in 2025 — nearly a doubling. These figures, however, are backed by only one source in this material and should be read with that caveat.
Muse Wonder: the answer only Meta has counted
The Muse Wonder models launched on October 12 are agentic models focused on coding. According to BigDATAwire, both models have a one-million-token context window and accept not just text but also images, video and audio as input.
On SWE-Marathon, a benchmark for large software development tasks, Meta says Muse Wonder Full scored 79 percent, versus 74.2 percent for Claude Opus 5.5. The standard Muse Wonder model is said to sit at 78.8 percent. BigDATAwire notes that all these figures are based on Meta's own published results — they have not been independently verified. Meta thus claims the models match or exceed competing offerings from OpenAI and Anthropic, but this is the company's own account.
The market was, for now, willing to be convinced. The stock rose more than 3 percent after the launch, according to Yahoo Finance/Barchart, amid a broader selloff in AI stocks.
Open questions
It remains to be seen whether Meta's own benchmark figures hold up in the real world, and how the returns on the infrastructure bet will actually materialize. Muse Wonder is days old, and Meta has yet to show how $100–200 billion will pay for itself. Zuckerberg himself has — at least in hindsight — admitted this was a bet made without certainty about the outcome.