Meta Spent $135B on AI — And Is Now Considering Licensing Google's Instead

The $135 Billion Problem

Meta has committed more capital to AI infrastructure than any other company in 2026. Its capex guidance — $115 billion to $135 billion for the year — dwarfs the budgets of Anthropic, OpenAI, and Google DeepMind combined. The expectation was that this spending would produce a frontier AI model: Avocado, Meta's first proprietary closed-weight foundation model.

Internal benchmarks tell a different story. Avocado performs well above Meta's own previous models. It does not perform well enough to compete with Google's Gemini 3.0, which Google released in November 2025. In the areas that matter most for agents — logical reasoning, complex coding, and multi-step agentic behavior — Avocado falls short of the frontier.

What Avocado Is — and What It's Not

Avocado is a significant departure from Meta's previous AI strategy. Meta's Llama family — Llama 3.3, Llama 4 — were open-source, openly benchmarked, and celebrated by the developer community. Avocado is proprietary: closed weights, internal-only benchmarking, and designed to give Meta a competitive edge it can't share with competitors who also use Llama.

The shift to proprietary is driven in part by Alexandr Wang, Meta's Chief AI Officer, who joined the company after Meta acquired Scale AI for $14.3 billion. Wang's view is that meta-models — AI trained on vast, high-quality labeled data — require a competitive moat that open-source models undermine.

The Gemini License Discussion

The most striking detail in the New York Times' reporting is what Meta's AI leadership is considering as a stopgap: temporarily licensing Google's Gemini technology to power Meta AI across its products while Avocado is further developed.

That means the AI in your WhatsApp chat, Instagram DMs, and Facebook feed — currently powered by Meta's own models — could run on Google's technology. Meta has not confirmed this is happening. But the fact it is being discussed signals how significant the performance gap is internally.

If Meta does license Gemini, the EU AI Act creates a transparency obligation: users in Europe must be informed when AI responses are generated by a third-party system. Meta would be required to disclose that its AI is powered by Google.

Avocado Delay Timeline

• LATE 2025 (ORIGINAL TARGET)

  Avocado targeted for a late 2025 release, intended to coincide with Meta's Q4 earnings cycle and product announcements.
• EARLY 2026 (FIRST DELAY)

  Release slips to early 2026 after post-training complexity increases. Meta begins internal benchmarking against new competitive releases.
• MARCH 2026 (SECOND DELAY)

  Target moves to March 2026. Internal tests run against Gemini 3.0 (released November 2025) reveal Avocado is not at frontier performance.
• MARCH 12, 2026 (NYT REPORT)

  New York Times reports Avocado delayed to at least May 2026. Meta AI leadership discussion of Gemini license becomes public.
• MAY 2026 (CURRENT TARGET)

  Earliest possible Avocado launch. Meta states model will show a "rapid trajectory" of progress. Final decision on Gemini license still pending.

Why This Matters: The Platform Lock-In Risk

The Avocado story is a concrete example of why betting on a single AI provider is a structural risk. Meta's 3.3 billion users are dependent on whichever AI model Meta happens to be running this quarter. If Avocado is delayed, underperforms, or gets swapped out for Gemini, users have no control. The AI in their products changes without notice.

The same dynamic applies at a smaller scale for individuals and teams. If you use a tool locked to one AI model, your experience is at the mercy of that model's roadmap. If GPT-5.4 is the best today but Claude Opus 5 is the best in six months, a single-model tool leaves you behind.

Meta AI vs. the Competition: Where Avocado Needs to Land

Frequently Asked Questions