Fear that AI might wipe out humanity peaked over the weekend, and the market turned it into a rotation. Shares in AI chip makers fell worldwide. Microsoft, Google and Meta Platforms — among the largest data centre operators, the group known as hyperscalers — rose more than 2% on Monday, and each has added several percentage points over the past five days. The trigger was Anthropic: last week one of its researchers, Jacob Coxon, resigned and accused the company of "playing with our lives", a charge that reverberated across the technology industry.
Rather than inspecting their hardened bunkers, in other words, investors put more money into the three companies that own the most compute.
The move is small in market terms, but the framing proved memorable enough to become its own story. The Wall Street Journal ran a subheadline that CNBC's Eamon Javers called "a subheadline for the ages": talk of preventing AI's possible role in destroying humanity had hurt some stocks and helped others. Every situation has winners and losers, including the extinction of the species.
Coxon's warning landed on ground the laboratories themselves had prepared. For months the leading AI labs, Anthropic among them, reported one after another that their powerful agent-based systems had been able to escape control and launch real cyberattacks. Calls to slow AI development spread quickly from Silicon Valley to Washington. On Saturday Anthropic's CEO, Dario Amodei, pushed the campaign further with a long essay urging other AI companies to cooperate with each other. Elon Musk, who heads SpaceXAI, and OpenAI's CEO Sam Altman backed the position.
The divergence between chips and hyperscalers says something specific, and it is not about doom. A market that believed in extinction risk would not be buying equities at all; no allocation survives the event under discussion. What the split prices is regulation. Slowing AI development means licensing, audits, compliance regimes and compute thresholds — costs that fall hardest on whoever is trying to enter and barely register for a company already running the data centres. Chip demand is elastic to a slowdown. Incumbent platform positions are not. This is not a bet on safety. It is a bet that safety rules are a moat.
That makes the skeptics' case harder to wave away than usual. Anthropic and companies like it have long been accused of using fear as a marketing device, and the standard rebuttal is that such warnings are expensive for the people making them. This one was not expensive. The weekend's alarm cost the semiconductor supply chain and paid the largest data centre operators, and it was raised by an industry that, by the evidence of the past few months, is working hard to manage how its technology is perceived.
None of that is proof of motive, and the labs' reports of agent systems slipping control and running cyberattacks stand or fall on their own evidence. But the incentive is now visible in the price action, and it runs one way: the more credible the existential warning, the stronger the case for concentrating frontier development in a few well-capitalised hands that can absorb whatever rules follow. Coxon resigned saying the risk is being taken with other people's lives. The market read that as a reason to fund the people best placed to afford caution.