Meta just blinked in the AI race, and it cost them $135 billion in credibility.

The Signal

Meta delayed Avocado, their next-generation AI model, after internal benchmarks showed it trailing Google's Gemini 3.0 and likely OpenAI's latest offerings. The company now faces a choice that would have seemed absurd 18 months ago: license Gemini from Google to power Facebook, Instagram, and Threads while they figure out what went wrong.

The numbers tell the story. Meta projected $135 billion in spending for 2026, nearly double last year's $72 billion, with most of that earmarked for AI infrastructure and talent. Mark Zuckerberg has publicly bet the company's future on owning the AI stack from chips to models. Now that bet looks shaky. Avocado beats their previous models and last year's Gemini 2.5, but that's table stakes. In foundation models, second place means irrelevance.

What's striking isn't the delay itself. Every AI lab has model releases that underperform. What matters is Meta's response. Discussing a Gemini license means their AI leadership doesn't trust the pipeline. A two-month delay to May suggests minor fixes. Floating a Google deal suggests they see a longer drought coming. The gap between those scenarios is massive, and Meta's clearly hedging because they don't know which world they're in.

Compare this to Apple's position. Apple never promised to build frontier models. They're integrating AI where it matters to users, licensing what they need, and spending a fraction of Meta's capital. If both companies end up running Gemini, one of them looks pragmatic and the other looks like they lit $60 billion on fire chasing a capability they couldn't deliver.

The Implication

Watch what Meta does in the next 90 days. If they ship Avocado in May, this was noise. If they announce a Google partnership, it's a strategic retreat that reframes the entire AI infrastructure build-out. For everyone building on the assumption that compute spending guarantees model leadership: Meta just showed you it doesn't. The agent economy needs reliable, improving models. If the biggest spenders can't deliver them in-house, the licensing model wins and the vertical integration thesis dies.

Source: Daring Fireball