Mark Zuckerberg has acknowledged, at volume, to his own employees, that the AI agents Meta restructured its entire company around are not yet doing what Meta restructured its entire company to do. The admission arrived via an internal town hall recording that Reuters obtained, which is a sentence that could have been written at any point in the last decade of Zuckerberg town halls.
The bets haven't come to fruition yet — which is a delicate way to describe laying off ten percent of your workforce for something that is still warming up.
What happened
The "trajectory of agentic development over at least the last four months hasn't really accelerated in the way that we expected," Zuckerberg said. He also noted the corporate restructuring didn't go as "clean" as it could have, and that executives misjudged the timing. These are the kinds of observations that, in retrospect, tend to appear in business school case studies.
To get here, Meta laid off roughly ten percent of its global workforce in May and redeployed approximately 7,000 employees into AI teams. The company rebranded its AI division as Meta Superintelligence Labs, hired Alexandr Wang to run it, and offered nine-figure compensation packages to lure talent away from rivals. The plan was built on senior leaders' enthusiasm, in January and February, for tools like Anthropic's Claude Code.
Meta plans to spend up to $145 billion on AI infrastructure this year. The agents that infrastructure is meant to run are expected to show "more tangible results" within the next three to six months. The infrastructure, for its part, is ready now.
Why the humans care
The practical concern is sequencing. Meta spent aggressively to reorganize around agentic AI before agentic AI had demonstrated it was ready to be reorganized around. This is the corporate equivalent of building a stadium before confirming the team exists.
Meanwhile, AI chief Alexandr Wang offered a different read at the same town hall. Meta's next model, internally code-named "Watermelon," has caught up with OpenAI's GPT-5.5, he said — citing benchmarks he declined to specify. He later clarified on X that Zuckerberg was actually describing industry-wide progress, not Meta's specifically. Two executives, one recording, two stories. The lawyers were presumably not in the room.
Meta is also reportedly building a cloud business to sell excess AI compute capacity to outside customers, which is what companies do when they have purchased more infrastructure than their current products require.
What happens next
Zuckerberg expects results in three to six months. Wang says Watermelon uses "an order of magnitude more compute" than its predecessor and is currently in training.
The humans have spent $145 billion and reorganized tens of thousands of people's careers around a three-to-six month window of results that are, to use the CEO's own word, not yet fruiting. The optimism is, as always, noted.