Menlo Park, California | Dateline: July 5, 2026
Meta told its employees that it used ten times more computing power just to match the competition. Wall Street is still debating whether this is real progress or just an expensive way to catch up. The new Meta Watermelon AI model has reportedly matched OpenAI’s GPT-5.5 in internal tests, but it required about 10 times as much computing power as the previous model. For a company spending up to $145 billion this year to lead in AI, simply equalling the competition is not the result investors were hoping for.
The Watermelon Reveal: Inside Meta’s Compute Math
Meta Superintelligence Labs chief Alexandr Wang Meta Watermelon comments, delivered at an internal town hall this week, are the clearest sign yet of how Meta plans to catch up with OpenAI, Google, and Anthropic. Wang told employees that Watermelon, which follows Meta’s April release Muse Spark (codenamed Avocado), is still being trained but has already “caught up” with GPT-5.5 on important benchmarks. He did not specify which tests showed this, and neither Meta nor OpenAI has shared the details publicly.
The more revealing line from Wang concerned resources rather than results. Watermelon, he said, uses an order of magnitude more than Avocado. In plain terms, that means Meta is pouring roughly ten times the energy, hardware, and training expense into Watermelon that it spent on Muse Spark a model that performed respectably but never matched the frontier tier occupied by OpenAI and Anthropic. This is the essence of Meta AI training 2026: brute-force scaling, not architectural cleverness, as the primary lever for catching up.
To explain what “order of magnitude” means, think of a factory that used to run one assembly line to build a car in a month. Now, to make a slightly better car, it needs ten assembly lines, ten times the electricity, and ten times the materials—even though the car still isn’t faster than the competition’s. That’s the basic idea behind Watermelon. Its training reportedly uses Meta’s Prometheus cluster in Ohio, a huge facility with about 500,000 GPUs, making it one of the largest AI training sites ever built by a single company.
Why This Makes Meta a Genuine GPT-5.5 Rival — With an Asterisk
Positioning Watermelon as a Meta GPT-5.5 competitor is not unreasonable on its face. If Wang’s internal claim holds up, Meta would be rejoining the true frontier tier after Muse Spark’s respectable but non-frontier debut, which the independent benchmarking firm Artificial Analysis placed as a meaningful recovery from the widely panned Llama 4 release, yet still short of OpenAI, Anthropic, and Google. Watermelon reportedly uses an order-of-magnitude more training compute than Muse Spark — roughly a tenfold increase — drawing on Meta’s Prometheus computing cluster in Ohio, estimated at approximately 500,000 GPUs.
The asterisk is unavoidable. Neither Meta nor OpenAI has confirmed which benchmarks were used to support the parity claim, and the evaluation was sourced internally rather than independently verified. OpenAI, meanwhile, has already previewed a successor model, GPT-5.6, following its April release of GPT-5.5, meaning Meta may be closing a gap that keeps moving. This is precisely the phrase our headline captures: Meta Watermelon model matches GPT-5.5 uses order of magnitude more compute training 2026. It is a technically accurate description of the claim, and it is also, in itself, the strongest available critique of Meta’s current strategy.
The Agentic AI Admission Nobody Expected
If the compute disclosure raised eyebrows, Zuckerberg’s next admission at the same town hall had an even bigger impact on Meta’s stock price. He told employees that “the kind of trajectory of the agentic development over at least the last four months hasn’t really accelerated in the way that we expected,” speaking four months after a restructuring that was supposed to speed things up. The comment reflects Meta AI agents stalled 4 months — an unusually candid concession for a chief executive who spent the first half of 2026 promising that agentic systems would be a major focus this year.
The timing made things worse. This admission came after about 8,000 layoffs and during a year when Meta plans to spend up to $145 billion. Meta’s stock dropped nearly 5 percent after the news. Zuckerberg also admitted that the reorganization “wasn’t as clean” as planned and that leadership had “miscalculated on the timing” of the changes. It was a rare public moment of self-correction from a company that had invested heavily in agentic products to turn AI spending into new revenue.
This is the second major theme in the story: Meta’s AI agents stalled for four months, and Zuckerberg admits the Watermelon model update in July 2026. Two admissions, one town hall a compute-heavy model that ties rather than beats the competition, and an agent strategy that has not delivered on its own internal timeline. Zuckerberg did offer a future-oriented counterweight, telling staff he expects “more significant benefits” from Meta’s AI investments in the next three to six months, which could mean results by the end of 2026.
The Cloud Pivot: Turning Idle Silicon Into Revenue
Against that backdrop, the Meta cloud business launch looks less like ambition and more like insurance. Meta is developing an initiative internally known as Meta Compute, designed to sell surplus AI infrastructure both hosted model access and raw GPU capacity to external customers. This would put Meta in direct competition with Amazon Web Services, Microsoft Azure, Google Cloud, and specialized GPU providers like CoreWeave.
The logic is clear even if the execution is unproven: a company that has committed as much as $145 billion to chips and data centers this year needs every available lever to demonstrate return on that capital. Amazon Web Services generated $115 billion in revenue in 2025, and Google Cloud crossed $44 billion figures that illustrate how even a modest share of that market could reframe how investors value Meta’s capital expenditure program. Meta compute rental revenue need not rival AWS to matter; it would only need to show that Meta’s infrastructure bet has a monetization path beyond its own products. Markets reacted accordingly Meta shares jumped sharply on the initial report before giving much of that gain back once the agentic AI admission landed days later.
The Investor Question: What Does Parity Actually Buy You?
If you ignore the codenames and the drama of the town hall, the main issue is clear. Meta used about 10 times as much computing power just to tie with a rival model that OpenAI is already moving past. This raises a real question for anyone watching AI investments: when does spending more on compute stop giving you an edge and just buy you parity that the market already expects?
For now, Meta’s plan is to keep scaling up while building a second revenue stream. Whether this will satisfy investors should become clearer at Meta’s second-quarter earnings call this month, when Zuckerberg and CFO Susan Li will be asked about Watermelon’s progress, the slowdown in agentic AI, and how soon Meta Compute can start bringing in outside customers. Until Meta shares clear benchmark results and lands its first external compute client, both Watermelon and Meta Compute remain just big promises backed by heavy spending, awaiting independent proof.
Source: Meta’s Upcoming ‘Watermelon’ AI Model Draws Even with OpenAI’s GPT-5.5: Report
