In recent weeks, semiconductor sector sentiment has continued to recover, rising about 13% from its July low and gaining 67% year to date. The core logic supporting the rally is renewed market confidence that AI spending by large technology companies will continue.

However, Anthropic CEO Dario Amodei's sudden proposal to "pace the frontier" has brought a question that had already cooled somewhat back into investors' view: if leading AI labs voluntarily slow the pace of model capability improvement, will chip demand and data center construction also slow accordingly?

Bernstein's answer in its September 14 report is: not necessarily.

Its core judgment is that model iteration can slow down, but compute demand is unlikely to hit the brakes in sync. As the center of gravity in AI computing spreads from training to inference, especially as Agent applications gradually gain traction, semiconductor demand is no longer driven only by the next generation of frontier models.

Why are AI labs suddenly discussing "deceleration"?

Amodei proposed a three-stage "pacing framework."

The first step is to introduce independent third-party evaluators within AI labs to verify whether companies are complying with the safety commitments they have already made; the second step is to push for U.S. regulation to establish unified rules for all American AI companies; the third step is global coordination, including attempting to establish some kind of cooperation or constraint mechanism with China.

The initiative comes as the AI safety debate is heating up again. Recently, several safety and alignment staff at major foundation model labs have publicly resigned, criticizing the industry for failing to establish sufficiently reliable constraint mechanisms in the race toward self-improving superintelligence. Amodei's remarks were subsequently echoed by leading technology figures including Sam Altman, Elon Musk, and Satya Nadella, meaning that "whether frontier model development should be slowed" has moved from an internal lab debate into the public and regulatory agenda.

For semiconductor investors, what is truly sensitive is not the safety discussion itself, but the possibility that "pacing" could directly affect compute investment. Amodei mentioned that regulation could set limits around key areas such as training compute, the way training runs are conducted, and the use of AI to improve AI. Once constraints fall on cluster scale and training resources, the market will naturally worry that AI chip and data center spending could be affected.

Bernstein believes that this concern could weigh on semiconductor sector sentiment in the short term, but is not enough to change actual demand.

First, Amodei is not advocating stopping training, but rather slowing the pace of capability improvement for frontier models from "extremely fast" to "slightly fast." Considering the current pace of expansion in the AI industry, even if model iteration slows somewhat, the corresponding training investment behind it could still remain substantial.

More importantly, the drivers of AI chip demand have already changed.

Models can slow down, but inference demand will not stop

In the past, AI chip demand was mainly driven by large model training. Increasing model parameters, expanding training data, and lengthening training cycles directly drove the continued growth of GPU cluster scale. But as models enter the stage of large-scale application, inference is becoming the new center of compute consumption.

Especially in Agent scenarios, models no longer simply generate one answer to one question, but need to continuously plan tasks, call tools, read results, correct paths, and verify outputs. A single user request may trigger multiple rounds, or even dozens of rounds, of model calls, with inference consumption significantly higher than that of traditional chatbots.

This means that even if frontier models temporarily stop upgrading rapidly, the large-scale deployment of existing models will still continue to increase compute demand. Bernstein points out that the computing resources already put into use are even struggling to fully meet the needs of current models, let alone new models that may be launched in the future and more complex Agent applications.

At the same time, the AI capital expenditures of large technology companies have already entered the actual execution stage. The construction cycles for chip procurement, data centers, power systems, and network infrastructure are relatively long, and it is unlikely that they will be adjusted immediately because of a single industry initiative. Bernstein therefore judges that the near-term revenue targets of current AI companies have basically already reflected customers' spending plans, and the possibility of major changes in related plans is relatively low.

In other words, what an "AI slowdown" hits first may be valuations and market sentiment, rather than chip orders. As long as inference call volumes, enterprise deployment, and Agent applications are still growing, it will be difficult for AI infrastructure investment to truly stop.

Behind the Safety Initiative, There Is Also U.S.-China AI Competition

Bernstein believes that Amodei's initiative may not be entirely about safety issues, but also involves U.S.-China AI competition.

Amodei emphasized that U.S. restrictions on AI development cannot come at the cost of allowing China to gain a leading position. He advocates continuing to restrict exports of advanced AI chips and semiconductor equipment to China, while cracking down on unauthorized model distillation and strengthening model weight protection.

Based on this, Bernstein speculates that Chinese AI labs rapidly closing the model gap through distillation may be one of the important reasons driving this initiative. It should be noted here that this is Bernstein's judgment, not a policy motive directly acknowledged by Amodei.

For leading U.S. labs, unilateral deceleration carries obvious risks: if they themselves face stricter safety reviews and compute restrictions while open-source models or overseas competitors are not bound by the same rules, deceleration may instead weaken their leading position. Therefore, Amodei's proposal for unified U.S. regulation and global coordination is essentially also intended to extend the rules across the entire competitive environment.

From this perspective, so-called "AI safety" is increasingly intersecting with technological competition. Training compute, model weights, and distillation methods are not only safety issues, but may also become new industrial barriers. Related policies may change the regional structure of global chip sales, but they will not necessarily weaken the overall investment of U.S. tech companies.

Bernstein Continues to Bet on Nvidia, Broadcom, and Equipment Stocks

Bernstein believes that although stricter safety assessments will increase short-term regulatory pressure, they may benefit the long-term development of the AI industry. Auditable and constrained AI systems are more likely to gain acceptance from governments, enterprises, and the public, and also help reduce political resistance to data center construction and commercial deployment.

Therefore, the report did not downgrade its view on AI infrastructure because of "frontier pacing," but instead continued to list Nvidia, Broadcom, and semiconductor equipment companies as preferred directions.

Nvidia received an "Outperform" rating with a target price of $400. Bernstein believes that its data center market opportunity remains enormous, and both training and inference demand will jointly support GPU growth.

Broadcom also received an "Outperform" rating with a target price of $575. The report expects the company's AI growth trajectory in 2026 to accelerate further in 2027 and 2028, with custom AI chips and networking businesses remaining the main drivers.

On the equipment side, Applied Materials, Lam Research, and KLA all received "Outperform" ratings, with target prices of $700, $385, and $250, respectively. Bernstein is bullish on equipment demand driven by GAA, advanced packaging, HBM, and NAND upgrades. Among them, KLA has the basis for a premium valuation due to structural growth drivers, a solid competitive position, relatively low China substitution risk, and disciplined capital allocation.

Bernstein's final judgment is that the "AI slowdown" will become a new sentiment variable for the semiconductor sector, but is not yet sufficient to reverse the AI infrastructure cycle. Frontier model research and development may slow, but training has not stopped, inference demand is accelerating, and Agents are further raising the computational consumption per task.

AI labs may be starting to discuss tapping the brakes, but for the computing power industry, what truly determines demand is no longer just the pace of model iteration, but the scale and frequency of AI usage. As long as this variable is still growing, the demand logic for AI chips is not yet over.

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia