Meta's flagship AI model, internally codenamed Avocado, has been pushed from mid-March 2026 to at least May — possibly June. The reason, confirmed by the New York Times, is unambiguous: in internal benchmarks, Avocado fell short of leading models from Google, OpenAI, and Anthropic on reasoning, coding, and writing tasks. It cleared Google's Gemini 2.5 (March 2025) but could not match Gemini 3.0, which launched in November 2025 — leaving Meta's most advanced model stranded between two generations of its primary rival's best work. META stock fell 3.15% — down $20.11 — to $618.07 as of 12:14 ET on March 13, 2026 (NASDAQ).
The Wound Beneath the Delay: Llama 4 and the Benchmark Scandal
Avocado's stumble cannot be read in isolation. It is the continuation of a crisis ignited by the disastrous Llama 4 launch in April 2025. In a departure interview with the Financial Times, Yann LeCun — Meta's longtime chief AI scientist and a founder of modern deep learning — confirmed that Llama 4's benchmark results had been manipulated before release, including using different model variants for different tests to inflate apparent competitiveness. Zuckerberg responded by sidelining the entire GenAI organisation, triggering a mass talent exodus. LeCun himself is now leaving Meta after more than a decade to found Advanced Machine Intelligence Labs, publicly calling Meta's new superintelligence hires "completely LLM-pilled."
The consequence is structural: every future Avocado benchmark claim will be discounted by developers, enterprises, and recruiters until Meta earns independent validation. Credibility, not compute, is now Meta's scarcest AI resource.
$135 Billion and a Borrowed Brain: The Capital Paradox
Meta announced $115–135 billion in capital expenditure for 2026 alone, explicitly tied to Meta Superintelligence Labs (MSL) — the closed-model division led by Alexandr Wang, hired following a $14.3 billion investment agreement with Scale AI. Wang is an avowed advocate of proprietary models, a direct philosophical reversal from Zuckerberg's famous 2024 memo declaring open-source AI "the path forward."
The most striking contingency to emerge from the NYT's reporting: Meta's AI leadership discussed temporarily licensing Google's Gemini to power its consumer AI products while Avocado catches up. No decision has been reached, but the symbolism is brutal — a company spending up to $135 billion on AI infrastructure reportedly considering renting its rival's intelligence. It reveals that management views "having an AI answer everywhere in Meta's apps" as a non-negotiable product requirement, even if its own flagship model is not ready to deliver it.
Organizational Fracture at the Worst Possible Moment
The internal restructuring compounds the technical miss. Chris Cox, Meta's CPO and a 20-year veteran, was reassigned away from overseeing AI after Llama 4's failure. A new AI engineering team has been stood up under CTO Andrew Bosworth, running parallel to Wang's MSL — a dual-track structure that has generated reported friction over whether AI should primarily serve Meta's advertising business or a broader platform ambition. LeCun's departure, a philosophy pivot from open to closed, two competing internal AI orgs, and exploding spend form a combination that historically signals a company still searching for its operating model — not one executing with conviction.
Three Metas, One Stock
The sharpest analytical move for investors is to stop pricing Meta as a single entity. There are effectively three businesses embedded in one ticker. Meta the ad machine remains world-class and is already generating practical AI returns through improved ad ranking and creation tools. Meta the distribution layer — with billions of users, creator ecosystems, and smart glasses hardware — is genuinely formidable. Meta the frontier model lab is unproven, credibility-damaged, and burning capital at a rate that demands frontier-quality results.
The bull case survives if the first two businesses compound faster than the third destroys capital. The bear case materialises if Avocado launches to mixed reviews, open-weight alternatives from DeepSeek, Qwen, and others continue compressing the practical quality gap, and Meta ends up paying frontier-lab costs for distribution-layer economics.
The watchlist is short: Does Avocado's release date hold? Does Meta commit to open or closed access? Do third-party evaluations validate claims without asterisks? And does management begin talking about measurable monetisation rather than superintelligence? Until those questions resolve cleanly, the AI leadership premium in META's multiple is not yet earned.
not investment advice