Meta's AI Ambitions Are Running Into a Very Expensive Reality Check

For a company that has spent years telling the world that open-source AI is the future, Meta is now quietly asking whether it should borrow a competitor's closed model to keep up. That's not a small irony. That's a fundamental crack in the strategy.

The story: Meta's next major frontier model, internally codenamed Avocado, got pushed from a March launch to at least May after internal benchmarks showed it underperforming. It reportedly sits somewhere between Google's Gemini 2.5 and Gemini 3.0 on reasoning, coding, and writing tasks. Good enough to clear the bar from a year ago. Not good enough to compete with what's shipping right now. And in AI, six months behind is a long way behind.

This Isn't a Stumble. It's a Pattern.

Llama 4 launched in April 2025 to a lukewarm reception, falling well short of the bold predictions Zuckerberg had made just three months earlier. The flagship model in that series, Behemoth, reportedly hit a wall when engineers couldn't produce meaningful gains. By June, Zuckerberg had blown up the org structure, launched something called Meta Superintelligence Labs, pulled in Scale AI's Alexandr Wang through a $14.3 billion deal, and reframed the whole thing as the company's march toward personal superintelligence. In January 2026, he told investors to expect a "rapid trajectory" on new models.

Avocado shipping two-plus months late is not that trajectory.

None of this means Meta is cooked. It means frontier AI development is genuinely hard, and that throwing money and org charts at a benchmark gap doesn't always close it.

The Gemini Thing Is the Real Story

The more telling detail here isn't the delay. It's that Meta reportedly considered licensing Google's Gemini as a stopgap while Avocado catches up. No decision confirmed. Meta's spokesperson called it "speculative reporting about theoretical approaches." Fine.

But the fact that it was discussed at all? That's the story.

Meta's entire AI identity is built on openness. The Llama series exists specifically as the alternative to closed ecosystems at OpenAI and Google. Open weights, broad access, no vendor lock-in. Developers and enterprises bought into that pitch. If Meta ends up quietly running Gemini under the hood of its products while Avocado gets fixed, that pitch is dead. You can't be the open-source champion and the company licensing a rival's closed model at the same time. The two things don't coexist without one of them being a lie.

The Spending Math Is Getting Uncomfortable

Meta has committed $115 to $135 billion in capital expenditures for 2026. That number would be aggressive for any company. For Meta it's a specific kind of gamble, because unlike Microsoft, Google, or Amazon, there's no cloud business on the back end generating revenue from the infrastructure being built. Every dollar spent on AI compute has to pay off through ads, social, WhatsApp, and whatever consumer AI features people actually adopt.

That pressure is showing up in the workforce. Meta is reportedly planning layoffs hitting roughly 20% of its 79,000 employees, which would be its deepest cuts since the 2022-2023 downsizing. They're trying to fund an arms race while cutting operating costs at the same time. That works until it doesn't.

Avocado Might Be Fine. The Bigger Problem Won't Be.

When Avocado does ship, it might be excellent. The problem is that excellent-in-May competes against whatever the leaders are shipping in May, not whatever they were shipping when Avocado was originally planned. The target keeps moving. The window where any one company gets to call itself the clear frontier leader keeps getting shorter.

Meta will keep swinging. They have the money and the scale to stay in the fight. But the story Zuckerberg has been telling about open-source AI domination is getting harder to tell with a straight face, and licensing Gemini as a bridge would make it basically impossible. At some point the gap between the ambition and the reality has to close. Avocado is the next test of whether it actually can.