Meta has delayed the launch of its next-generation artificial intelligence model, codenamed Avocado, pushing the release from mid-March to at least May after internal testing revealed the system falls short of rival offerings from Google and OpenAI. The setback arrives as Meta prepares to spend up to $135 billion on AI infrastructure this year, raising questions about whether the company’s massive financial commitment is translating into competitive technology.
The delay was first reported by The New York Times, citing three people familiar with the matter. According to the report, Avocado underperformed in three critical areas during internal benchmarks: logical reasoning, programming, and writing capabilities. While the model exceeded the performance of Meta’s previous generation systems and outperformed Google’s Gemini 2.5, it failed to match the capabilities of Google’s newer Gemini 3.0 or leading models from OpenAI and Anthropic.
Meta Considered Licensing Google’s AI Technology
Perhaps the most striking detail to emerge from the report is that Meta’s leadership team discussed temporarily licensing Google’s Gemini technology to power certain Meta products while Avocado is brought up to competitive standards. No final decision on the licensing arrangement has been confirmed, but the consideration alone signals the gap between Meta’s AI ambitions and its current output.
The company has been positioning itself as a leader in AI development through its open-source Llama model family, which has been widely adopted by developers and startups. However, reports now suggest Meta may be shifting away from its open-source approach with Avocado, developing it instead as a closed, commercial-grade system designed to compete directly with proprietary models from its rivals.
A $135 Billion Bet With No Cloud Revenue to Show For It
Meta’s 2026 capital expenditure budget of $115 billion to $135 billion represents a near-doubling of its 2025 spending of $72.2 billion. The investment is funding new data centers, custom AI chips, and computing infrastructure, including a gigawatt-scale data center campus in Louisiana backed by a $27 billion joint venture with Blue Owl. The company also invested $14.3 billion in Scale AI, whose CEO Alexandr Wang now leads Meta’s frontier AI division known internally as TBD Lab.
Unlike Amazon, Microsoft, and Google, Meta does not operate a cloud computing business that can directly monetize its AI infrastructure investments. The company relies primarily on advertising revenue, using AI to improve ad targeting and content recommendations across Facebook, Instagram, and WhatsApp. That business model makes the return on Meta’s AI investment less straightforward than it is for cloud providers who can sell AI compute directly to enterprise customers.
What This Means for the AI Landscape
Meta shares fell approximately 2% on Friday morning following the report, as investors reassessed the company’s near-term AI execution timeline. The stock decline was modest, but analysts noted that the delay introduces uncertainty around Meta’s ability to deliver the AI capabilities it has been promising shareholders.
The Avocado delay does not mean Meta is abandoning its AI push. The company has additional models in development, including one codenamed Watermelon, which represents the next generation beyond Avocado, and an image and video generation system called Mango. Meta CEO Mark Zuckerberg has described the company’s goal as building “superintelligence” and has committed to spending tens of gigawatts worth of computing capacity over the coming decade.
For the broader startup ecosystem, Meta’s stumble highlights a widening gap in the AI race. Even with $135 billion in annual spending and access to top talent, building frontier AI models that compete with the best remains an extraordinarily difficult technical challenge. Startups building on top of AI platforms may want to consider how dependent their products are on any single provider’s model roadmap, as even the largest companies face unpredictable delays in delivering next-generation capabilities.
