AI: Is It Time for Doubt?

“We must pace the frontier.” The statement comes not from a company being left behind by AI, but from Anthropic’s CEO himself. In a post, he calls for putting the brakes on the all-out development of frontier models, highlighting the safety and control issues they raise and advocating international political coordination. In a rare show of agreement, Elon Musk, Demis Hassabis of Google DeepMind, and even Sam Altman have followed suit. Altman also took the opportunity to officially announce that OpenAI’s IPO would be postponed. Has AI become so powerful that it is challenging the hitherto unquestioned pursuit of market-share gains and value creation in favor of greater regulation? Investors are entitled to ask the question.

The Chinese and US governments were quick to respond to this attempt at industry-led regulation. Beijing rejected calls for cooperation on sovereignty grounds. For his part, the President brushed aside the concerns of the standard-bearer of effective altruism in an angry tweet. Donald Trump argued that “the only control or ‘safeguards’ AI needs is a STRONG AND INTELLIGENT PRESIDENT (with a high IQ!), and the United States has that in abundance!” Hardly surprising given that the United States and China are locked in a technology race in which the former’s lead is visibly shrinking as Chinese models make increasing use of distillation* and mixture-of-experts architectures**.

Two days before his CEO’s post, Anthropic had released its Econ Scenario Explorer, a simulator modelling economic trajectories through 2030. These simulations offer a different take on the job apocalypse scenarios published so far. Anthropic’s researchers believe that the productivity gains generated by this technological revolution will be heavily skewed in favor of capital and therefore against labor, mechanically reducing the share of wages in the economy. As a result, sustained pressure on labor income could weigh on consumption and, ultimately, on growth. Governments are presumably mindful of this risk and will seek to offset it through higher corporate investment.

Beyond the geopolitical and macroeconomic considerations, it is difficult not to draw a connection between Dario Amodei’s call for restraint and his company’s planned IPO this year. Ahead of the listing, Anthropic will have to present prospective shareholders with an income statement and forecasts for future cash generation. Slowing the massive investment required to develop its most advanced models, such as Claude Fable 5.1, would probably make the term sheet more attractive. Beyond the obvious short-term financial impact, heavier regulation could also have the effect of “freezing the competitive landscape,” given that the largest players can absorb high compliance costs whereas smaller laboratories and open-source players cannot. As in the banking industry, tighter control mechanisms would also raise barriers to entry.

An analysis of the hyperscalers’ off-balance-sheet commitments does not point to any slowdown in infrastructure spending. The Wall Street Journal recently provided a timely overview, and the figures are staggering: $2.4 trillion. The bulk relates to data-center leases ($904 billion), with the remainder spread across purchases of chips, energy and services. Meta alone has $696 billion in off-balance-sheet commitments and, through Mark Zuckerberg, has firmly rejected calls to slow the pace of development, arguing that legal liability and market mechanisms are capable of providing adequate safeguards.

Higher control requirements could therefore result in some spending being redirected toward cybersecurity and model monitoring, through more extensive testing and automated supervision. In that scenario, infrastructure capex for computing capacity would remain unchanged. That is also the view of a majority of investors, according to Bank of America’s latest survey: 79% believe there will be no reduction in capital expenditure this year.

At a time when the industry is debating how to control this rapidly expanding technology, quantified use cases are beginning to emerge. One recent example came from cloud platform ServiceNow at a conference, where its CEO announced that the company had deployed twenty agents internally. The tasks performed by these agents were estimated to represent around 2.5 million hours of manual work, equivalent to $1 billion in value.

Despite the increasingly tangible evidence of returns on investment from AI projects, the Anthropic CEO’s initiative has acted as a wake-up call for some of the technology sector’s most consensual trades, particularly semiconductors. Sharp position unwinding drove the subsector down 6% on Monday, September 14***, while the AI loser basket rebounded, led by software stocks. These unexpected moves, together with the accompanying rise in volatility, are a reminder to investors that equity-risk diversification is far from obsolete…

* Distillation: a technique whereby knowledge is transferred from a large, complex AI model (the teacher) to a smaller, faster model (the student).

** Mixture of experts: tokens are processed by specialized subnetworks.

*** SOX Index.

Pierre Pincemaille, General Secretary of Investment at DNCA. This article was finalised in September 25th, 2026.

This promotional document is a simplified presentation and does not constitute a subscription offer or an investment recommendation. No part of this document may be reproduced, published or distributed without prior approval from the investment management company.

DNCA Investments is a trademark held by DNCA Finance