NextFin News - The artificial-intelligence trade is selling off again, but this time the trigger did not come from a cheaper Chinese model, a valuation miss, or a macro shock. It came from inside the industry itself: over the weekend of Sept. 12-13, 2026, the chief executives of the world's most valuable AI companies called for the pace of model development to be slowed, and markets immediately began repricing the entire AI capital-expenditure cycle. By 04:46 ET on Monday, Nasdaq 100 futures were down 505.5 points, or 1.72%, Nvidia was falling more than 2% before the open, and chipmakers Intel, AMD and Marvell were each down 5% to 6%.
The selloff matters because it is the first time the AI boom has been attacked through the supply side of its own growth equation. Earlier routs questioned whether the chips would still be needed. This one questions whether the companies building them will be allowed - or will allow themselves - to keep shipping new models fast enough to consume them.
The Situation: A Selloff With an Unusual Trigger
The sequence began on Sept. 8, when Anthropic researcher Jacob Coxon resigned and warned on X that the industry is racing toward self-improving superintelligence while "gambling with our lives." His post, which drew wide attention, included a line that has since become the shorthand for the episode: "The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt."
Four days later, the warnings moved from a departing researcher to the executives who run the race. Anthropic CEO Dario Amodei published an essay calling on AI companies to "slow the pace at which we improve the capabilities of AI models," and proposed permanent access for independent safety evaluators plus greater coordination among the leading labs. Elon Musk and OpenAI CEO Sam Altman publicly agreed. Altman went further, saying in an interview that an initial public offering for OpenAI now would be an "ill-advised moment" and confirming the company would not go public in 2026.
The market's read was swift and directional. The stocks that have ridden the AI capex wave hardest were the ones punished. Nvidia fell more than 2% in premarket trading; its Magnificent Seven peers Meta and Amazon were each down more than 1%. Chipmakers Intel, AMD and Marvell fell nearly 6%, 5% and 6%. The S&P 500 and Dow futures also slipped, though the damage was concentrated in the AI complex.
What made the move unusual was not just the trigger but the cross-currents around it. The AI selloff arrived as Brent crude rose 3.6% to $108.31 a barrel and West Texas Intermediate climbed 3.3% to $103.35, levels not seen since May. Inflation data the prior week had come in hot, and traders were pricing in roughly an 89% chance of a Federal Reserve rate increase at the meeting later that week. A growth-sensitive, duration-heavy trade was selling off into a backdrop of rising oil, rising inflation expectations and a likely rate hike - the opposite of the easy-money environment in which the AI rally was built.
There was a second, quieter rotation inside the technology sector. While chipmakers and AI-infrastructure names fell, U.S. software stocks that had been battered by fears of AI disruption climbed: ServiceNow rose 3%, Adobe 2.5% and Workday 2.5%. The same news that hurt the companies selling the picks and shovels helped the companies that had been feared as the next victims of the machines.
Why This Selloff Is Different: Dissecting the Mechanism
A Capex Cycle Priced on a Cadence That Could Be Throttled
To understand why this selloff is different, start with what powered the rally. The AI bull market was not built on a single product cycle. It was built on a cadence: a predictable drumbeat of ever-more-capable models, each one requiring larger training runs, more GPUs, more networking, more datacenter power. Hyperscalers committed to capital-expenditure plans that assumed that cadence would continue, and the market capitalized the chip, utility, memory and networking suppliers as if the cadence were a law of physics rather than a business decision.
"I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to."
Evan Hubinger, Anthropic's alignment-science lead, said that after Coxon's resignation, adding that he personally put the probability of AI killing all humans at greater than 10% within the next decade. Samuel Marks, Anthropic's scalable-oversight lead, said: "In general, the more senior the employee, the more concerned they are."
Here is the mechanism. A demand shock - a cheaper model, a missed earnings quarter - reduces the number of chips buyers want at any given moment. That is what happened in January 2025, when the emergence of a low-cost Chinese model called DeepSeek sent Nvidia down nearly 17% in a single session, wiping out a record $593 billion of market value, while AI-power plays Vistra and Constellation Energy fell 28.3% and 20.8%. The DeepSeek rout asked: will buyers need fewer chips?
This selloff asks a different question: will the model-release cadence itself be slowed? If the leading labs deliberately stretch out the time between frontier releases - whether through a voluntary pact, regulatory pressure, or internal caution - then the training-run pipeline thins out not because demand disappeared, but because the pipeline is being deliberately throttled. The capex cycle does not break; it decelerates. And a deceleration is harder to underwrite than a pause, because it removes the one thing the market was really paying for: the certainty of the next order wave.
This is why the selloff is structurally different from every AI rout that came before it. Previous selloffs were about the size of the pie. This one is about the speed at which the pie is being eaten.
Cyclical Sentiment or Structural Regime Shift? The Call
The first-order read is cyclical. No binding agreement exists yet. The weekend statements were essays and social-media posts, not a treaty. Front-running a possible slowdown is not the same as a slowdown. If the "slowdown" remains rhetorical - if hyperscalers keep training runs on schedule because falling behind is commercially fatal - then Monday's move will look like a sentiment overshoot, the kind that has been bought aggressively throughout this bull market.
But the structural case is stronger than the market's initial reaction suggests, and it rests on three pieces of evidence. First, the warnings are coming from the insiders with the best information, and they are unusually specific: Coxon said executives couch their phrasing in public but express fear privately; Hubinger put a number on the existential risk; Marks tied seniority to concern. Second, the proposed remedy - independent safety evaluators with permanent access, coordinated release schedules - is a regulatory architecture, not a PR gesture. If even part of it becomes policy, it changes the rules of the industry permanently. Third, the macro window for absorbing that shock has closed: with oil above $100 and a rate hike priced at roughly 89%, there is no dovish cushion to reflate the multiple if the capex thesis softens.
So the judgment is this: the near-term leg of the move is cyclical sentiment front-running an unconfirmed pact, but the repricing is of a structural variable - the terminal growth rate of AI capital spending. A cyclical wave and a structural shift are both present, and they should not be blended. The cyclical leg can be bought; the structural leg cannot be ignored.
The Second-Order Trade Nobody Is Pricing: The Software Reversal
The most underappreciated consequence of an AI slowdown is not that chip stocks fall. Everyone prices that. The second-order effect is that the bear case for software stocks weakens.
For much of the past year, the dominant software thesis has been the opposite of the infrastructure thesis. If AI agents can autonomously execute complex workflows, the reasoning went, customers will no longer need full software applications; switching costs will fall, pricing power will erode, and the SaaS model will face a structural margin squeeze. That fear kept pressure on names like ServiceNow, Adobe and Workday even as the chip complex marched higher.
Monday's move inverted that relationship in a single session. ServiceNow, Adobe and Workday rose while the chip complex fell, because a slower AI-development path also means a slower arrival of the agentic capabilities that would disrupt them. The same news that cuts the numerator of the AI-capex story also cuts the denominator of the software-disruption story. Investors who had been long infrastructure and short software were forced to unwind both legs at once, and the rotation amplified the dispersion.
This is the second-order cross-asset transmission that the market is still working through: an AI slowdown is not uniformly bad for technology. It is bad for the supply chain of AI progress and good for the industries AI was supposed to displace. The trade is not "AI up, everything else down." It is a dispersion trade between the builders of intelligence and the built world intelligence was meant to replace.
The Adversarial Case - And What Would Prove It Wrong
The strongest argument against the structural bear case is the one Michael Burry made on X: the slowdown warnings are an attempt by major players to stifle smaller competitors. Under this reading, the calls for a pause are a cartel move - a way for incumbents to invite regulation that freezes the field while they consolidate their lead. If that is true, capex does not slow. Hyperscalers cannot afford to fall behind, and the GPU orders keep coming.
Brian Jacobsen, chief economic strategist at Annex Wealth Management, framed the skepticism more carefully: "The strongest arguments for caution are those grounded in evidence, not fear. We should be wary both of incumbent firms seeking to protect their position and of confident predictions about outcomes that no one can reliably quantify."
There is real evidence for the skeptical view. Nvidia's own response to the DeepSeek shock was that inference "requires significant numbers of NVIDIA GPUs and high-performance networking," and that new scaling laws - pre-training, post-training and test-time scaling - would sustain demand. The cloud vendors, in that framing, will buy every GPU they can get regardless of model efficiency, because the competitive race rewards whoever ships next. A slowdown pact among a few labs does not stop that race; it may even intensify it outside the pact.
So the bull case rests on a specific, observable condition: the capex cycle survives only if the slowdown stays voluntary and unenforced. The falsifying signal is equally specific. If a formal multinational AI development accord with enforcement mechanisms is announced before the next earnings cycle, the structural bear case is confirmed and the chip complex should not recover its prior highs. If, instead, no accord materializes by the Q4 2026 earnings season and hyperscaler capex guidance for fiscal 2027 comes in flat or higher, then Monday's move was a cyclical overreaction and the AI trade remains intact.
What Comes Next: Scenarios and the Forward Look
The practical implication is that the AI complex is no longer a single trade. It has split into three horizons, and each points in a different direction.
In the short term, sentiment and liquidity dominate. The move happened into a hostile macro setup - oil above $100, hot inflation, a priced-in rate hike - and that combination tends to produce overshoots in both directions. Volatility will stay elevated through the Fed meeting and the next wave of AI-policy headlines.
Over the medium term, fundamentals take over, and the key variable is capex guidance. The market will be watching hyperscaler spending plans for fiscal 2027, the order books of Nvidia, AMD and Marvell, and any sign of a formal slowdown accord. A flat or rising capex number in the face of safety rhetoric would be the clearest sign that the cyclical read was right. A downward revision paired with a policy announcement would confirm the structural one.
In the long term, the structural question is whether the AI industry can grow into a regulated utility-like model - slower, safer, more predictable - without losing the exponential cadence that justified its valuation. That is an open question with no clean answer, and it is why the terminal multiple on the entire complex is likely to compress even if earnings keep growing.
Three scenarios frame the path ahead. The base case is a negotiated deceleration: some coordination on safety evaluation and release timing, but no hard cap on spending, leaving the capex cycle slower but intact. The upside case is a clean rejection of the slowdown - no accord, capex guidance held firm, and the selloff fully retraced as a cyclical event. The downside case is a binding multinational pause with enforcement, which would break the cadence and force a deeper, more durable repricing of the chip, utility and memory complex.
What to watch, in order: first, whether the implied probability of a rate increase holds near 89% into the Fed meeting, because a hike would remove the last support for duration-heavy valuations; second, any formal announcement of an AI safety accord and its enforcement mechanism; third, fiscal 2027 capex guidance from the hyperscalers; and fourth, whether the software-versus-infrastructure dispersion that opened on Monday persists or collapses.
The market has spent two years treating every AI selloff as a buying opportunity, and for two years it has been right. This time the shock did not come from a competitor's product or a macro number. It came from the people who build the technology telling the world they are afraid of it. That is a different kind of selloff - and it deserves a different kind of answer.
Explore more exclusive insights at nextfin.ai.
