Two Messages, One Room: Meta’s $145 Billion AI Bet Just Wobbled in Public

On July 2, inside a Meta town hall, two of the company's most senior figures gave employees two irreconcilable accounts of the same thing. One was candid about failure. The other was confident about triumph. Both were talking about Meta's AI. And sitting between them was the largest single-company AI spending commitment in the industry. gafam.ai reads what that contradiction actually tells us.

Message One — Zuckerberg's Rare Admission

The first message was unusually honest, and it came from the top.
According to an audio recording obtained by Reuters, Mark Zuckerberg told employees at the July 2 town hall that the trajectory of agentic development over at least the last four months "hasn't really accelerated in the way that we expected." He said the company's bets on its new structure "haven't come to fruition yet," acknowledged that the reorganisation was not as "clean" as it could have been, and admitted he had made mistakes in the workforce shift — adding that he would "almost certainly make more."

This is a striking admission from a CEO who spent the past year going all-in on AI. The restructuring he was apologising for is the same one behind roughly 8,000 layoffs earlier this year — cuts made in the name of AI efficiency. Zuckerberg told staff he expects more substantial returns within three to six months, which would put any payoff toward the end of 2026 — more than a year after Meta created its Superintelligence Labs under Alexandr Wang.

There is a pointed contrast in his own words. In January, Zuckerberg was telling the world that projects which used to need big teams could now be done by a single talented person. In July, he was telling his own staff that the bets haven't paid off. Six months separates the promise from the admission.

Message Two — Wang's Very Different Story

Minutes later, in the same room, Meta's AI chief painted a picture that could not have been more different.

Alexandr Wang told employees that Meta's next model — codenamed Watermelon, the successor to the Muse Spark model internally called Avocado — has "caught up" with OpenAI's GPT-5.5 on closely followed benchmarks. He added that Watermelon uses "an order of magnitude more compute" than Muse Spark, and that a Meta model to rival Anthropic's Claude Opus would arrive "pretty soon."

Here is where gafam.ai applies discipline, because this claim needs it. Wang's benchmark assertion is single-sourced, the specific benchmarks were not named, no methodology or evaluation data was published, and neither Meta nor OpenAI has confirmed it. This is an internal, unverified claim made in a motivational context — a leading indicator at best, not a verified result. It should be treated as provisional until Meta releases a model card or independent evaluation. An internal town-hall benchmark boast is marketing to your own staff, not evidence.

The Detail That Matters More Than the Claim

Even if we grant Wang's claim entirely, the more revealing number is the one about compute. Watermelon reportedly uses an order of magnitude — ten times — more compute than Muse Spark, merely to match GPT-5.5.

Sit with that. And then add the timing problem: OpenAI reportedly debuted a stronger GPT-5.6 late last month — available, as we reported, only to government-approved partners. So the best case Meta is presenting to its own staff is this: spend ten times the compute to catch up to a model OpenAI has already surpassed. Meta may have closed the gap to GPT-5.5 while the frontier moved to GPT-5.6.

This is the brute-force scaling strategy meeting its natural limit. When catching up requires an order of magnitude more compute and still leaves you a generation behind, the scaling lever is delivering diminishing returns. That is the real story inside the contradiction — not whether Watermelon matched a benchmark, but what it cost to try, and that the target had already moved.

The Money — And the Market's Verdict

The financial frame makes the stakes concrete. Meta plans to spend between $125 billion and $145 billion on AI infrastructure this year — a substantial slice of the more than $700 billion Big Tech is collectively pouring in. The town hall also surfaced that Meta is building a cloud business to sell excess AI compute to outside customers, echoing a move by Musk's xAI.

The market did not read the two messages as balanced. On the day, Meta shares closed down 4.9%, and the Philadelphia Semiconductor Index fell more than 5%, with memory-related stocks dropping sharply. Investors weighed Zuckerberg's candour more heavily than Wang's confidence — and the "excess compute for sale" detail was read by some not as a new revenue line but as a sign that Meta has capacity it cannot yet productively use.

The European Perspective