Most chief executives wait for proof before betting the company. Mark Zuckerberg went the other way, saying that by the end of 2025 Meta was ready to commit somewhere between $100 billion and $200 billion before it had any real sense of what its new models could do.
Then came a twist few people saw coming. Just days after that admission, Zuckerberg publicly thanked Apple, the company whose rules cost Meta so much, for helping one of its new AI models run on Apple hardware. The two moves tell the same story about where Meta wants to stand in the AI era.
Muse Wonder is the first real test of the spending
The payoff Zuckerberg was waiting for arrived on October 12. Meta’s Superintelligence Lab released Muse Wonder and a stronger version called Muse Wonder Full, its first serious shot at a frontier model since Alexander Wang reorganized the company’s AI work in August 2025.
Also read: Mark Zuckerberg says Meta found a way to make more money
Meta is pitching memory as the standout feature.
The company says its own testing puts the models alongside leading rivals and highlights what it calls “active memory,” the ability to hold on to and combine information users share over time, as reported by TechCrunch. Both models will power Meta’s Muse agent, and they were trained at its Prometheus data center in central Ohio.
That training did not come cheap. Wang said building a frontier model now costs “tens of billions of dollars,” and Meta’s performance claims still need independent results to back them up, according to TheStreet. For now, investors are mostly taking the company’s word for it.
Early demand looks encouraging, though. Meta’s Muse personal assistant app logged 2.8 million downloads in its first 12 days, according to TechCrunch. Citi thinks the app could eventually add at least $27 billion in annual revenue by fiscal 2030.
The bill keeps getting bigger
The cost of owning everything is showing up in Meta’s books. When it reported second-quarter results on July 29, Meta raised the low end of its 2026 capital spending forecast by $5 billion, while free cash flow sank 91% from a year earlier to just $784 million. Some investors are already comparing the splurge to the costly metaverse push.
Zuckerberg used that same call to sketch a way out. He said businesses have offered to run their own AI projects on Meta’s capacity at a meaningful premium to what it cost to build, but added that “there will continue to be a significantly higher margin on selling intelligence rather than selling compute directly,” as reported by Reuters. In other words, renting out servers is a side gig, not the plan.
Hyperion shows how quickly the price tag can swell. In July, Meta said its Richland Parish campus in Louisiana would grow to 5 gigawatts of computing capacity from an original design of 2 gigawatts, pushing the cost above $50 billion from a prior estimate of $27 billion, Quartz reported. It will be Meta’s largest data center.
Investors have flinched more than once. On September 25, a Goldman Sachs note argued that the largest U.S. AI spenders need roughly $300 billion in annual AI revenue just to break even, and Meta shares slid about 4% that day. Zuckerberg’s own fortune dropped by billions in that single session.
Old Apple wounds shaped the go-it-alone strategy
The scars from Apple’s privacy change run deep. In February 2022, then-CFO Dave Wehner called the hit from App Tracking Transparency “a pretty significant headwind for our business,” admitted the $10 billion figure was only an estimate and watched the stock sink 23% in extended trading.
Wehner also took a pointed swipe at Apple’s search deal with Google, which he suggested was treated differently under the same privacy rules. “The incentive clearly is for this policy discrepancy to continue,” he said, as reported by CNBC. It was a reminder of how little say Meta had over the platform carrying its apps.
More Mark Zuckerberg:
- How much META stock does Facebook founder & CEO Mark Zuckerberg own?
- Mark Zuckerberg sends bold message to Meta investors after Muse
- Mark Zuckerberg says Meta found a way to make more money
The mood has softened lately. On October 15, Zuckerberg thanked Apple’s chief executive, John Ternus, for helping Meta’s Muse Spark Open model run on Mac Studio computers, in what appeared to be the first public partnership between the two companies in years. Neither side disclosed financial terms.
The power balance may be shifting, too. Bank of America analyst Wamsi Mohan warned on September 29 that AI agents could capture shopping, referrals and payments that now flow through Apple’s ecosystem. If agents become the new front door, owning the full stack could leave Meta holding a much stronger hand.
Wall Street wants answers on October 21
Analysts are mostly upbeat heading into Meta’s October 21 earnings report. Bank of America’s Justin Post wrote that Meta is “back on offense with AI” and said recent launches have strengthened confidence in Meta’s AI execution, though he flagged the risk that heavier AI investment could pressure 2027 earnings, according to Yahoo Finance.
JPMorgan has also come around. Analyst Doug Anmuth upgraded the stock to Overweight in September, arguing Meta is still early in rolling out frontier models and AI products beyond advertising, while projecting capital spending well above Wall Street’s consensus for the next two years. He also noted that the Muse app hit No. 3 in the U.S. App Store on its second day, TheStreet reported.
The harder question is what happens to the ad machine. Bank of America has listed competition from AI-native platforms for users’ attention as one of Meta’s key risks, noting that advertising still accounts for most of its revenue. Zuckerberg is betting that owning every layer gives Meta the best shot at controlling where those eyeballs go next.
Related: Mark Zuckerberg sends bold message to Meta investors after Muse
