Bernstein said in a report dated Sept. 14, 2026 that Anthropic CEO Dario Amodei’s newly proposed AI “pace framework” may be getting misread by the market as a signal of slowing AI spending. The firm’s view was straightforward: spending plans are unlikely to change because of it.
The report noted that the semiconductor sector is up 67% year to date, but still sits about 19% below its June high and roughly 13% above its July low. That backdrop has sharpened market attention on whether calls to slow AI development could feed into weaker capex expectations.
Amodei’s proposal centers on a three-step framework
According to the report, Amodei published an article over the weekend laying out a framework intended to slow the rate of improvement in AI model capabilities. The proposal has three parts.
- First, independent third-party evaluators would be embedded to verify and report whether companies are complying with their stated practices and commitments. Anthropic said it would implement this step now.
- Second, democratic coordination, meaning U.S. regulation would apply across all U.S. AI companies whether they opt in or not.
- Third, global pacing, which would involve some form of cooperation or coordination with China.
Bernstein analyst Stacy Rasgon framed the key question this way: does slower pacing mean slower spending? The firm’s answer was no.
The report said Amodei was talking about limits on training compute, the nature of training runs, and the use of internal AI to improve AI. He did not call for training to stop. Bernstein described the proposal as shifting from “extremely fast” to “somewhat fast,” which would still imply a very high rate of expansion by current standards.
Inference demand is already outstripping supply
Bernstein said AI semiconductor demand is increasingly being driven by inference, especially after the emergence of agent use cases. Those workloads consume multiple times the compute required for traditional inference, the report said, and current compute availability already falls well short of what existing models need, not to mention what future models may require.
On that basis, the firm said recent AI revenue targets likely already incorporate planned spending, making any material change to those plans unlikely. Bernstein also noted that this is not the first time major industry figures have called for AI standards and oversight. Google DeepMind head Demis Hassabis made similar suggestions in July.
Even so, Bernstein said those discussions are unlikely to alter real-world spending cadence because inference demand has already turned into concrete orders and capacity planning.
China concerns may be the bigger force behind the debate
The report said Amodei made clear in his article that U.S. restrictions cannot come at the cost of allowing China to pull ahead. He proposed maintaining the gap with China through measures that include continuing bans on sales of advanced AI chips or semiconductor equipment to China, cracking down on unauthorized distillation, and preventing model weight theft.
Bernstein said recent news reports indicated that Chinese labs had secretly used unauthorized Claude distillation to train models. Amodei repeated those points in an interview with CBS, according to the report.
Bernstein argued that while the proposal is publicly framed around fears tied to recursive self-improvement and safety alignment, concern over China may be the larger driver. In the firm’s view, U.S. AI companies want a regulatory framework that constrains rivals without slowing their own compute buildout.
Safer AI could support broader adoption
Bernstein said safer AI is easier to adopt. The firm added that it is not a committed believer in a “Skynet” scenario, but it does not want that outcome either.
The report said creating AI safeguard plans could help long-term industry adoption and ease political and social anxiety around AI. At this stage, the main item being discussed is third-party evaluation, which Bernstein described as a reasonable direction.
It also argued that it is too late to put the AI genie back in the bottle. A safer release path, in its view, is the more realistic course. Establishing safety frameworks could reduce policy risk and support long-term investment in AI infrastructure.
Nvidia, Broadcom and chip equipment remain preferred picks
Bernstein kept its constructive stance on the AI buildout and said Nvidia (NVDA), Broadcom (AVGO) and semiconductor equipment names remain its top picks.
- Nvidia: Outperform, $400 target price
- Broadcom: Outperform, $575 target price
- AMD: Outperform, $650 target price
- Applied Materials (AMAT): Outperform, $700 target price
- Lam Research (LRCX): Outperform, $385 target price
- KLA (KLAC): Outperform, $250 target price
- PDF Solutions (PDFS): Outperform, $65 target price
As of the Sept. 11, 2026 close, Nvidia traded at $218.29, Broadcom at $361.99, AMD at $378.78, Applied Materials at $456.49, Lam Research at $298.22, and KLA at $180.64.
Bernstein said the market’s Monday reaction would be only the first data point in judging whether Amodei’s proposal will alter AI spending plans or instead act as a catalyst for broader industry adoption.
Disclosure
This article is based on a整理 and interpretation by Chaoxiang Research of a third-party brokerage report from Bernstein dated Sept. 14, 2026, along with publicly available market information. The ratings, target prices, earnings forecasts and related judgments cited in the piece are the views of the brokerage analysts and represent only the position of their institution, not that of Chaoxiang Research, and they do not constitute investment advice.
Markets carry risk, and decisions should be made independently. This article should not be used as a basis for buying or selling any security.
