Bernstein 13 Sep 2026 U.S. Semiconductors and Semicap Equipment Can you put the AI genie back in the bottle

U.S. Semiconductors and Semiconductor Capital Equipment U.S. Semiconductors and Semicap Equipment - Can you put the AI genie back in the bottle?

Can you put the AI genie back in the bottle? Semi sentiment had been improving in Stacy A. Rasgon, Ph.D. recent weeks, with the group (while still down ~19% from June peaks) up ~13% from the July trough and up 67% YTD as optimism around AI spending had begun to overpower sustainability concerns. However, we have a feeling those latter concerns are going to move Arpad von Nemes back to the forefront on Monday with Anthropic CEO Dario Amodei penning an essay over the weekend proposing a plan to “pace the frontier” and slow the pace at which the industry improves the capabilities of AI models, following the public resignations of safety and Alrick Shaw alignment personnel at some the large foundational labs who were apparently unsatisfied with the efforts currently in place (and admission from the labs themselves that they see a non-zero chance that AI could, indeed, kill us all?). Dario’s comments were echoed on Twitter/X by Sam Altman, Elon Musk, Satya Nadella, and other industry luminaries as well, Eva Zhang and we would expect this to have potential to further depress sentiment on the group. Dario proposes a three-step “pacing framework” including 1) Embedded independent third-party evaluators equipped to verify and report adherence to practices and commitments (Anthropic is committing to this step now); 2) “Democratic coordination,” in other words US regulation that targets all US AI companies, whether they want it or not and 3) Global pacing, i.e. working out some form of cooperation/coordination with China.

We could see a number of possibilities here (from best to worst): This could be an attempt to redirect increasing ill-will that has been building toward AI and datacenter buildouts. It could be an effort to crack down on and limit Chinese distillation efforts. It could be an acknowledgment that they simply cannot obtain all the compute they need to continue model improvements at the prior pace. It could be an attempt to force a slowdown in competing ecosystems (such as open source). It could be a realization that the return on training new models is falling, hence an excuse to slow down (in other words, an off ramp). And of course (though we hope not) we suppose it is possible that the labs have really seen something they are afraid of, for example recursive self-improvement (“RSI”) where models improve themselves at an increasing pace, out of control of humans (i.e. “Skynet?”).

But does a “pacing slowdown” necessarily imply a spending slowdown? We don’t think so...Dario does talk about considering pacing based on “limiting the ingredients that go into frontier models such as training compute, the nature of training runs, or internal use of AI to improve AI.” As such, (putting Skynet aside) investors might worry that “pacing” will impact overall spending. We are not so sure. Clearly Dario is not talking about stopping training (he suggests slowing from “extremely fast” to “only somewhat fast,” probably still plenty in our opinion given the current rate of scaling). Increasingly though AI semi demand is being driven by inference especially as agentic use cases arise, and at the moment there is already not nearly enough compute to satisfy demand for even the current models, let alone new ones that might be coming down the pipeline. And this is not even the first call for AI standards and oversight (Demis Hassabis, the head of Google DeepMind, proposed something similar in July). Hence we suspect recent AI revenue targets from our companies should already be representative of spending plans, and we would be very surprised to see those plans changing materially. (Continued on next page)

See the Disclosure Appendix of this report for required disclosures, analyst certifications and other important information. Alternatively, visit our Global Research Disclosure Website. First Published: 13 Sep 2026 21:34 UTC Completion Date: 13 Sep 2026 21:34 UTC

Stacy A. Rasgon, Ph.D. September 2026

(Continued from previous page)

How much of this is actually about China? Dario notes (correctly) that…