A rare show of consensus on "AI safety" over the weekend triggered a global selloff in chip stocks on Monday, September 14th. Anthropic CEO Dario Amodei published a lengthy essay titled "We Must Slow Down the Pace of Frontier AI Advancement," urging the industry to voluntarily decelerate the rate of model capability improvement—a call publicly endorsed by OpenAI CEO Sam Altman and Elon Musk, who also announced OpenAI would not go public in 2026. In response, Bernstein Research's semiconductor team lead and senior analyst Stacy Rasgon released a report titled with a question: "Can the AI genie be put back in the bottle?" His answer: no, and there's no need to try. The "slowdown narrative" does not equate to a "spending slowdown," and the fundamental drivers of AI semiconductor demand remain intact.

According to Bernstein's report, semiconductor sector sentiment had been steadily improving over the prior weeks—though still about 19% below June highs, it had rebounded roughly 13% from July lows, with year-to-date gains reaching 67%, as optimism around AI spending briefly outweighed sustainability concerns. This industry-wide "brake-pulling" has now pushed those latter worries back to the forefront. In his essay, Amodei proposed a three-step "pace framework": first, embedding independent third-party evaluation institutions with authority to verify and report on industry safety practice commitments (Anthropic has already unilaterally pledged to grant third-party assessors "permanent employee-level" system access); second, "democratic coordination," meaning US regulation targeting all American AI companies regardless of voluntary participation; and third, "global synchronized deceleration," involving some form of cooperation and coordination with China. The report notes this stance comes against a backdrop of multiple senior safety and alignment researchers publicly resigning from leading foundation model labs, with labs themselves acknowledging a "non-zero probability" of AI threatening humanity—Altman's phrasing being "unacceptable risk of human extinction."

Rasgon's team listed six possible interpretations of this call, ranked from best to worst: deflecting growing public hostility toward AI and data center construction; countering and restricting China's "distillation" practices; admitting an inability to secure all the compute power needed to maintain previous iteration speeds; attempting to slow down the competitive ecosystem (such as the open-source camp); recognizing declining returns on training new models and needing an excuse to "step down the ladder"; and the worst-case scenario—labs genuinely seeing something that frightens them, such as recursive self-improvement (RSI) beyond human control, colloquially known as the "Skynet" scenario. The report specifically highlights that the "China factor" likely carries significant weight. Amodei correctly argues that US restrictions cannot come at the cost of allowing China to overtake, recommending maintaining export bans on advanced AI chips and semiconductor equipment to China, cracking down on unauthorized distillation, and preventing model weight theft—the report references recent news of Chinese labs secretly using unauthorized Claude distillation to train models, with Amodei reiterating these points in weekend television appearances. "While the proposal wields RSI and safety concerns as a hammer, we suspect China is the primary driving force behind it," the report states.

The market's most pressing question: will "slowdown" impact AI capital expenditures? Bernstein's answer is "we don't think so," offering three layers of logic. First, Amodei's proposed pace shifts from "extremely fast" to "still fairly fast"—measured against current scaling speeds, this level "remains more than sufficient in our view," and he has not called for halting training. Second, AI semiconductor demand is increasingly driven by inference, especially with the rise of agentic use cases, and existing compute capacity is already far from sufficient to meet current models' inference needs, let alone future models. Third, this is not the first time calls for AI standards and regulation have emerged—Google DeepMind head Demis Hassabis proposed a similar framework back in July. Therefore, Bernstein judges that companies' recent AI revenue targets "should already reflect their spending plans," making substantial changes to those plans highly unlikely. The report also offers a longer-term defense: "Safer AI is more easily adopted AI." Safety mechanisms like third-party evaluations help alleviate political and social anxiety around AI (while also reducing doomsday probabilities, however small), thereby benefiting the industry's long-term penetration. "We suspect it's already too late to put the AI genie back in the bottle, so it's better to find a safer way to let it out."

Asian markets delivered the first response. SoftBank Group fell 10.7%, among the day's biggest decliners in Asian large-cap tech; Kioxia dropped 6.4%, Tokyo Electron fell about 1%, and the Nikkei 225 closed down roughly 0.8%. SK Hynix fell 6.4%, Samsung Electronics lost 4.1%, and the KOSPI index declined about 3.3%—by rough market cap estimates, the combined value erosion across just six companies (SoftBank, Kioxia, Tokyo Electron, SK Hynix, Samsung, and TSMC) reached the trillion-yuan level in a single day. As of Monday evening Beijing time, US markets had not yet opened, but pre-market data showed Nvidia down over 2.47% at one point, with AMD, Intel, Micron, and other chip stocks dropping around 5%, and Nasdaq futures down over 1%. Software stocks (Adobe, ServiceNow, etc.) showed relative strength.

This "slowdown consensus" arrives as Anthropic races toward an IPO valued at approximately $2 trillion, adding sudden uncertainty to its listing prospects. Media reports citing insiders indicate Anthropic has confirmed to some shareholders that Q3 adjusted operating profit will be positive for a second consecutive quarter, with gross margins exceeding 80% before distribution revenue sharing and training costs, and annualized revenue reaching $65 billion by end of July (up from just $9 billion at end of last year). However, reports suggest Anthropic may need to revise its confidentially submitted S-1 filing with the SEC, with underwriters potentially cutting the valuation or delaying the listing. Professional media commentary has been sharper, calling Amodei's initiative "far from sufficient"—if he genuinely believes in the risks, he should directly halt the most cutting-edge research like RSI. Skepticism about the "slowdown motives" extends beyond one party. D.A. Davidson analyst Louria believes Anthropic and OpenAI's actions increasingly resemble "pulling up the ladder," bordering on monopolistic behavior. OpenAI has already consulted members of Congress on whether industry-wide coordinated deceleration violates antitrust laws. Gartner analyst Chandrasekaran points out that if smaller competitors cannot afford frontier-level safety and evaluation investments, these standards effectively benefit Anthropic and OpenAI.

Bull-bear divergence among institutional players centers on "where the money goes after slowing down." Saxo Markets' Chanana cautions that AI and chip stock valuations are simultaneously built on strong demand and rapid technological progress—"even the mere possibility of delays is enough to trigger profit-taking." GlobalX's Billy Leung and Allspring's Gary Tan both argue that leading players calling for slowdown doesn't mean industry-wide synchronized deceleration; extended development cycles might instead push the industry from "spending on construction" toward monetizing existing assets. T. Rowe Price's Mallet cuts to the chase—investor focus has shifted from "how much compute AI needs" to "how much money these compute resources can ultimately generate."

The report reiterates a bullish outlook on AI buildout, with Nvidia Corp (NASDAQ: NVDA), Broadcom Inc (NASDAQ: AVGO), and semiconductor equipment stocks remaining the team's top picks in the sector. According to the report, Bernstein assigns Nvidia a $400 price target ("Outperform" rating, implying roughly 83% upside), Applied Materials (NASDAQ: AMAT) at $700 (implying about 53%), Broadcom at $575 (implying about 59%), KLA Corporation (NASDAQ: KLAC) at $250 (implying about 38%), and Lam Research (NASDAQ: LRCX) at $385 (implying about 29%). Advanced Micro Devices Inc (NASDAQ: AMD) (target $650, implying about 26%) receives an "Outperform" rating as "AI demand drives both CPU and GPU stories." Intel Corp (NASDAQ: INTC), Qualcomm Inc (NASDAQ: QCOM), Texas Instruments Inc (NASDAQ: TXN), and NXP Semiconductors NV (NASDAQ: NXPI) all carry Market Perform ratings, with Qualcomm being the only name whose target price sits below its current level ($165, implying approximately -9%).

In summary, this selloff triggered by "one of their own" changes expectations, not orders. If Bernstein's assessment holds—that what's slowing is the acceleration of frontier training rather than compute spending—then Monday's decline more closely resembles a concentrated emotional purge. The hard metrics truly worth tracking going forward are whether AI companies' capital expenditure guidance experiences any substantive downward revisions, and the final pricing of Anthropic's IPO.