Reports that Meta Platforms, Inc. (META) might sell computing capacity continue to weigh heavily on the AI hardware sector, with intense market discussion around potential "compute surplus" at the company.
An internal all-hands meeting was held on July 3rd. According to a Reuters report which obtained a recording of CEO Mark Zuckerberg's remarks, Zuckerberg candidly told employees that the company's large-scale restructuring implemented this year has "not yet borne fruit," and that the development trajectory of AI agents has "not accelerated as expected" for at least the past four months.
This admission fueled speculation that it confirms a "compute surplus," leading to circulating theories that this is the real reason behind Meta's potential move to sell capacity.
However, in the very same meeting, Meta's Head of Superhuman AI, Alexandr Wang, conveyed a starkly different and optimistic message to staff – the company's upcoming new model, codenamed "Watermelon," has already caught up to OpenAI's flagship GPT-5.5 model in benchmark tests.
Meanwhile, industry research firms are urging calm. In a report dated July 3rd, research firm SemiAnalysis argued that leasing out compute power externally does not equate to an industry inflection point or a compute glut, and that Meta's future data center and compute procurement will not slow down but instead accelerate further.
With conflicting information, the market remains uncertain about what is mere noise and what constitutes genuine insight.
Zuckerberg's Unusual Admission
According to a recording heard by Reuters, Zuckerberg made a rare self-critique at the July 3rd internal meeting.
He stated, "In hindsight, the trajectory for AI agents, at least for the last four months, has not accelerated in the way we expected." He added that the company's new organizational architecture "has not yet shown results."
The "new architecture" he referred to is the large-scale restructuring Meta implemented in May, which involved cutting roughly 10% of its global workforce and reassigning about 7,000 employees to AI-focused teams. This restructuring was predicated on the assumption that AI agent tools would mature rapidly, thereby helping the company improve efficiency.
Zuckerberg said that when planning the restructuring in January and February, the core leadership was "very worried that our pace of adjustment would not keep up with the industry," and there was "super optimism" about tools like Anthropic's Claude Code at the time.
The reality, however, is that the deployment speed of agent technology has been slower than anticipated. He admitted the restructuring was not executed "cleanly" and that executives misjudged the timing.
Despite this, Zuckerberg told employees he expects the company to see "more significant returns" from its AI investments "in the next three to six months."
AI Chief's Conflicting Assertion
According to two sources familiar with the matter who spoke to Business Insider, Alexandr Wang told the meeting that Meta's next-generation model, codenamed "Watermelon," has already matched OpenAI's flagship GPT-5.5 model in AI benchmarks.
He reportedly said at the meeting, "Watermelon is our next model after Avocado and is currently in training," and that "Watermelon uses an order of magnitude more compute than Avocado."
To clarify the codenames: Avocado is Meta's internal code for "Muse Spark" – the first in a series of models released in April, which performed well in benchmarks but was generally seen as not yet at the level of OpenAI or Anthropic's top-tier models. Watermelon is the next generation following Avocado.
GPT-5.5 is the model OpenAI released in April. It is worth noting that OpenAI launched an even newer generation, GPT-5.6, at the end of last month, but it has not been made fully public. Therefore, if Wang's claim is accurate, Meta has caught up to OpenAI's previous flagship, not its latest one.
Wang also posted on social media platform X, hinting at progress: he stated that an updated version of Muse Spark is coming soon with "major improvements" in coding and agent capabilities. When a user asked when a coding model comparable to Anthropic's Claude Opus would be available, he replied "soon" and said users would like what the company is "cooking."
Meta's current AI strategy is squarely aimed at industry leaders like OpenAI and Anthropic. Notably, OpenAI released the more powerful GPT-5.6 model last month, but it has not been publicly released yet due to U.S. government requirements.
Interpreting the Mixed Messages
This internal meeting occurred at a sensitive time.
Just the day before the meeting, media reported that Meta was developing plans to launch a cloud infrastructure business, selling AI compute capacity and model access to external customers. The news sent Meta's stock soaring 8.8% in a single day, but triggered a broad sell-off in global chip stocks – with the market fearing that Meta selling spare capacity signaled peak demand for AI chips.
Zuckerberg's subsequent admission at the meeting that AI agent progress was slower than expected partially corroborated the pessimistic "underutilized compute" interpretation. Meta's stock gave back most of its gains on Thursday, closing down 4.9%.
The Philadelphia Semiconductor Index fell 5.44% on Thursday, with all 30 components declining. Optical communication and memory concepts also fell sharply.
Regarding Meta's potential sale of compute, the market holds two interpretations: the optimistic view sees it as a new monetization channel to improve cash flow; the pessimistic view sees it as evidence that large model deployment is falling short of expectations, suggesting Meta is "retreating from frontier AI."
On the other hand, Wang's statement about Watermelon matching GPT-5.5, coupled with Meta's continuously increasing capital expenditure budget – raised from a previous forecast of $115-135 billion to $125-145 billion – indicates that Meta has not abandoned competition in frontier models.
Currently, neither Meta nor OpenAI has commented on the related content.
Massive Investment Underpins Strategy
These two starkly different internal assessments are both built upon Meta's unprecedented capital and human resource investments. To gain an edge in the AI race, Zuckerberg last year appointed Alexandr Wang to lead the renamed Meta Superhuman AI Lab and heavily recruited an elite AI research team. Previous reports indicated Meta offered compensation packages worth hundreds of millions of dollars to top AI talent.
Organizationally, Meta laid off about 10% of its global workforce in May and reassigned about 7,000 employees to teams focused on AI. This move sparked employee backlash and raised concerns about internal morale.
In terms of infrastructure, Meta's spending continues to balloon. The company has told investors it expects related capital expenditures to reach $125-145 billion this year, up from a prior forecast of $115-135 billion, due to rising component costs and increased data center spending. This also forms a significant part of the over $700 billion in total AI tech spending by large tech companies.
Internal Data Security Review
At the all-hands meeting, besides AI model progress, Meta's internal data management was also a focus. According to Reuters, Meta Chief Technology Officer Andrew Bosworth reported the results of a security review into a controversial employee mouse-tracking software.
The software was used to track employee mouse movements and digital activity for AI training. The project was paused last month after a security incident involving exposed sensitive data. Bosworth confirmed at the meeting that the review found no employee data was included in the AI training.
He also announced a policy shift. When Meta first installed the program on U.S. employees' computers in April, employees could not opt out. Bosworth stated that once the review is complete and the program is re-enabled, it will switch to an "opt-in" model. He emphasized that for employees unwilling to participate, it will no longer be mandatory.
Comments