The Brief

Meta has postponed its next-generation AI model, codenamed Avocado, from March to at least May after internal testing revealed it trailing systems from Google, OpenAI, and Anthropic in logical reasoning, programming, and agentic tasks. Leadership has reportedly discussed temporarily licensing Google’s Gemini technology to bridge the gap while the company pours up to $135 billion into AI infrastructure this year.

The Report

Meta’s flagship AI model Avocado will not ship before May at the earliest, pushed back from a mid-March target after internal benchmarks placed it behind the frontier systems it was built to match. The model underperformed Google’s Gemini 3.0, OpenAI’s latest offerings, and Anthropic’s Claude in reasoning, code generation, and autonomous task execution — the three capability areas now defining the competitive frontier.

Avocado outperformed Meta’s previous models and Google’s older Gemini 2.5, but could not close the gap to Gemini 3.0, which shipped in November. The delay marks the second slippage for the model, which was originally expected before the end of 2025 before training and performance-testing challenges pushed it into the first quarter of 2026. Some reports suggest the timeline could extend to June.

In internal communications from early February, product manager Megan Fu described Avocado as “Meta’s most capable pre-trained base model to date,” noting it would be “competitive with leading post-trained models even before undergoing post-training refinement.” That assessment has not survived contact with the benchmark suite.

More striking than the delay itself is what followed: Meta leadership reportedly discussed temporarily licensing Google’s Gemini to power certain products while Avocado improves. The New York Times, which first reported the delay, indicated the discussions may have occurred internally or with Google, though no decision has been confirmed. The prospect of Mark Zuckerberg approaching Sundar Pichai for access to a rival’s model drew immediate attention from analysts.

The setback arrives nine months and $14.3 billion after Meta hired Scale AI founder Alexandr Wang, 29, as Chief AI Officer and established TBD Lab — an approximately 100-person unit tasked with building frontier systems. Wang’s appointment followed the Llama 4 debacle of April 2025, in which departing chief AI scientist Yann LeCun acknowledged that benchmark results had been “fudged a little bit.” Zuckerberg was reported to have “lost confidence in everyone who was involved” and subsequently sidelined the entire generative AI organisation, cutting roughly 600 jobs in October.

Meta’s 2026 capital expenditure guidance of $115–135 billion represents an 88 percent increase over 2025’s $72 billion — an infrastructure bet unmatched by any competitor. Unlike Google, Microsoft, or Amazon, Meta lacks a cloud computing business to monetise that capacity directly, making the investment contingent on improvements to its advertising and consumer platforms. Meta shares fell more than four percent on Friday, the steepest decline since October, as investors absorbed the implications.

A Meta spokesman stated the company’s next model “will be good, but more importantly, show the rapid trajectory we’re on, and then we’ll steadily push the frontier over the course of the year.” Alongside Avocado, Meta is developing Mango for image and video generation and a successor model called Watermelon. LeCun, meanwhile, launched AMI Labs in Paris this week with $1.03 billion in funding, pursuing an alternative approach to AI built on world models rather than next-token prediction.

The company that spent three years positioning open-source AI as its competitive identity has now delayed its first proprietary model twice, discussed renting a rival’s technology, and lost the architect of its original strategy — all within six months of the largest single-year AI investment in corporate history.

The Angle

The interesting question is not whether Avocado catches up. It probably will, eventually, in the way that sufficiently funded engineering projects tend to converge on whatever the current benchmark standard happens to be. The interesting question is what Meta’s position reveals about the structure of the race itself.

Meta has 3.58 billion daily active users, $135 billion in annual capital to deploy, and the largest consumer distribution network ever built. It does not have a frontier AI model. The gap between those two facts is the story — not because Meta is failing, but because it demonstrates something the market has been slow to price in: distribution and capital are not converting into capability at the rate the investment thesis requires. Google, OpenAI, and Anthropic are not ahead because they spend more. They are ahead because the bottleneck in frontier AI is not compute or money. It is research depth, institutional knowledge, and the compounding advantage of having been building the right thing for longer.

The Gemini licensing discussion makes this concrete. A company does not explore renting a competitor’s technology because of a two-month delay. It explores renting because the internal assessment suggests the gap is architectural, not calendrical — that more time with the current approach may not be sufficient, and the product roadmap cannot wait for the research roadmap to catch up. Whether or not the licensing proceeds, the fact that it was discussed at leadership level tells you where Meta’s private confidence sits relative to its public guidance.

What $135 billion buys, it turns out, is the infrastructure to run a frontier model. It does not buy the model. The distinction between those two things — between the capacity to deploy intelligence and the ability to produce it — is becoming the most consequential sorting mechanism in the industry. Meta built the road. It is now discovering that the road does not come with a vehicle.

The company that enters this decade with the best model wins the platform. The company that enters it with the best distribution network but someone else’s model becomes a customer. Meta has spent a year and a half trying to avoid that outcome. Friday’s stock price suggests the market is beginning to consider it.