In an internal town hall meeting, Meta CEO Mark Zuckerberg announced that the development of the company’s own agentic AI has not gone as hoped. Furthermore, the company’s structure is not yet where it needs to be following the most recent round of layoffs. With its de facto departure from the “metaverse” and a shift toward leasing cloud capacity, Meta is already working on alternative revenue models for AI.
According to Reuters, Zuckerberg told his staff that the development of AI agents over the past four months “has not accelerated the way we expected.” Various initiatives, such as agents that assist with shopping, were announced earlier but never materialized. The company’s own LLMs, which were once reasonably competitive at compact scales with Llama, have not been a significant factor for quite some time. In the realm of open-weight models as well, other (primarily Chinese) players dominate, even though the Llama models had actually pioneered this innovation.
Incidentally, Meta AI head Alexandr Wang states that the company’s next model, codenamed “Watermelon,” will once again compete with the world’s leading players in the AI field. This LLM is said to be comparable to GPT-5.5, which, at the time of writing, is OpenAI’s most recent publicly available model series.
A reorganization with rough edges
The delay does not mean that investments are drying up. In April, Meta raised its capex forecast to $125 to $145 billion. Earlier, the company laid off about 8,000 employees, 10 percent of its workforce, while more than 7,000 others were reassigned to AI roles. According to Zuckerberg, that transition did not go as smoothly as expected. He admitted to making mistakes and said he will “almost certainly make more.”
It remains unclear when the investments will pay off. Zuckerberg states that more tangible benefits should become apparent within three to six months. That would be toward the end of 2026, well over a year after the founding of Meta Superintelligence Labs.
Compute as plan B
In the meantime, Meta is exploring alternative revenue models. This week, it was announced that the company plans to sell excess computing power as a cloud service under the name Meta Compute. This would allow it to monetize its infrastructure should the agent plans face further delays.
CTO Andrew Bosworth spoke at the town hall about the controversial training program that Meta launched in April. As part of that program, software was installed on employees’ computers to track mouse movements and keystrokes. According to Bosworth, a review following a data incident showed that no personal data was used for AI training. If the team restarts the project, participation will no longer be mandatory but will be opt-in.