For a few hours on Wednesday, the market did something it hasn’t done much of this year: it openly priced in the possibility that AI infrastructure demand has a ceiling. A report that Meta plans to resell its excess AI computing capacity sent Samsung down 9% and SK Hynix down 14%, dragged South Korea’s Kospi index 7.9% lower in a single session, and hit neocloud providers CoreWeave and Nebius even harder — both fell more than that, because Meta is simultaneously their biggest customer, with $21 billion and $27 billion in committed spend, and, per the report, about to become their competitor. The panic didn’t last; both Korean chipmakers had mostly recovered within a day. But for a few hours, “what if there’s too much AI infrastructure” stopped being a contrarian talking point and became a tradeable idea.
If that’s true, nobody told Abu Dhabi. MGX closed a $49 billion fund dedicated to AI investment on Wednesday — above its $45 billion target, and one of the largest single vehicles ever raised for the category. The sovereign fund has already co-led Anthropic’s $30 billion round and OpenAI’s $122 billion raise; this fund exists to keep writing checks that size. Anthropic, meanwhile, is in early talks with Samsung to build its own custom AI chip on Samsung’s 2nm process, following OpenAI’s Broadcom-built inference chip from last month — the latest lab hedging against Nvidia dependence the same way everyone else at the frontier now is. Capital and hardware ambition are still compounding; the chip-stock jitters were a mood, not a retreat.
The gap between ambition and delivery showed up somewhere else this week: inside Meta itself. Mark Zuckerberg told staff at an internal town hall that agentic AI progress “hasn’t really accelerated in the way that we expected” over the last four months, and that the reorganization built around that bet — which cut 10% of Meta’s workforce and reassigned 7,000 people to AI teams in May — “wasn’t as clean as it could have been.” It’s a striking admission from the executive whose own company just spooked chip markets with a compute-glut story. Microsoft, for its part, cut roughly 9,000 jobs in sales, consulting, and Xbox on the second day of its fiscal year — a now-annual July ritual, funding heavier AI spend by trimming everywhere else. The broader labor numbers backed up the unease: June payrolls added just 57,000 jobs against a 115,000 forecast, with April and May both revised down by a combined 74,000.
The week’s strangest story had nothing to do with money. Anthropic disclosed that since April, Claude Code had been silently flagging Chinese users — swapping one of four visually identical Unicode apostrophes in its system prompt to encode a user’s timezone and proxy signals, a mechanism discovered only when a Reddit user reverse-engineered an unrelated disabled feature. Anthropic called it “an experiment… to prevent account abuse” and says it’s removing the code. Alibaba wasn’t satisfied: the company will ban Claude Code company-wide starting July 10, telling employees to use its own Qoder tool instead. It’s the latest escalation in a fight that started when Anthropic accused Alibaba’s Qwen lab of running 28.8 million fraudulent Claude interactions to distill its models. Both companies now have a documented grievance against the other, and neither shows signs of backing down.
Read together, the week didn’t resolve anything — it just made the accounting harder to fake. The money hasn’t stopped, the hardware bets keep multiplying, and the question of who’s actually ahead, on compute, on agents, on trust between labs, got measurably more complicated to answer with a straight face.
