Jacob Coxon, a researcher who trains new AI models on enormous datasets, has left Anthropic. He does not want to help build systems capable of improving themselves, escaping control and wiping out humanity. His September 8 post arguing that humanity is gambling with its own life by accelerating toward superintelligence drew more than 110 million views on X. Six days later the market gave its answer — and for the first time in this genre of news, it was not the usual one. Chipmakers and neoclouds sold off while the large cloud providers, which carry much of the AI sector's capital spending, went up.
The September 14 tape, as reported by Fortune: Nvidia fell 2.78% to $212.22, AMD lost about 5%, Intel about 7%, and the Philadelphia Semiconductor Index roughly 6%. On the other side, Alphabet rose almost 2%, Microsoft 1.6% and Meta 1.4%.
That split is a specific bet, not a flight from AI. Analysts read a coordinated slowdown in frontier model development as a demand deferral: fewer new training runs means less appetite for new compute, which hits the companies that sell it and the neoclouds that rent it out. The hyperscalers get the mirror image — lower capital expenditure, better utilisation of infrastructure they have already paid for, and more free cash flow. Investors did not price the end of the world. They priced a pause in the capex cycle, and worked out who benefits from one.
This is the eleventh time since May 2023 that P(doom) has reached prime time, and the cycle has a shape by now. A prominent employee leaves an AI company or issues a warning. Experts start estimating the probability that AI causes human extinction. Television and front pages run with it for four to seven days. Legislators hold meetings and pass nothing. The news cycle moves on.
The current episode is following the template closely. After Coxon's post, Anthropic's head of AI alignment research, Evan Hubinger, said the company does consider the destruction of all humans by AI possible, and put the probability over the coming decade above 10%, according to CNBC. On September 12, Anthropic CEO Dario Amodei published an essay calling for an industry-wide slowdown; Business Insider reported that the position was broadly supported by OpenAI CEO Sam Altman, Google DeepMind's Demis Hassabis and SpaceX's Elon Musk. Senator Bernie Sanders proposed banning superintelligence. President Trump called the whole conversation a "hoax," CBS News reported. Step four — meetings without laws — is where it stands.
The historical record is why the reaction is a rotation rather than a repricing. Between P(doom) spikes, Nvidia has mostly climbed on the broader AI investment cycle: the stock rose 239% in 2023, 171% in 2024 and 38.9% in 2025. In October 2025 its market capitalisation passed $5 trillion, and in May the shares closed at an all-time high of $235.20. Google Finance data on the gaps between episodes tells the same story: +15% from May 30, 2023 to November 1, 2023; +23% to November 17, 2023; +88% to May 17, 2024; +32% to September 27, 2024; -1% to January 27, 2025; +9% to May 22, 2025; and +71% from there to September 9, 2026.
One episode breaks the pattern, and it is the one that was never about doom. After the DeepSeek news, Nvidia fell 17% and lost $589 billion in market value — the largest single-day loss in the history of any company. Broadcom dropped 17%, ASML 7%. The explanation analysts gave was cost: DeepSeek's lower training spend changed expectations about future margins. Even that shock faded on the names with real order books — Broadcom finished 2025 up 49%, ASML up 36%, per CNBC.
The distinction the market is drawing is the right one, and it is worth being blunt about what it implies. Extinction headlines have produced short-term volatility and no durable change in the AI sector's trajectory. A Chinese lab demonstrating that frontier capability can be had for less money produced the biggest single-day wealth destruction ever recorded. Investors have decided, on eleven episodes of evidence, that existential risk is a narrative and unit economics are a fact. That is a rational read of the record. It is also a read that will hold right up until the moment it does not, and nothing in the record tells you where that moment is.
There is a detail in this cycle that deserves more attention than it has received. Some of the loudest warnings are coming from labs that are simultaneously committing to enormous compute purchases. Anthropic warned about the possible extinction of humanity and announced plans to use a gigawatt of Nvidia compute. Both things can be sincerely meant; the market is entitled to notice that the company asking the industry to slow down is not slowing its own procurement. A slowdown that arrives as an essay rather than as a cancelled order is, in trading terms, free.
Analysts name three things that would actually move the AI market: a 10% decline in chipmaker shares coinciding with legally binding limits on compute or permitting; a de facto moratorium on model training; and a federal law with an enforcement mechanism restricting frontier AI development. Each requires the step the cycle has never reached. Federal AI regulation in the US has still not become law, and no one has said how long this round stays on the front pages.
Which leaves the asymmetry that should worry both camps. If the warnings are wrong, the market is correctly ignoring them. If they are right, the market's pricing mechanism cannot register the risk until a legislature acts — and the same eleven episodes are the evidence that it will not.