Why delaying an AI doomsday would benefit investors too

What price an apocalypse? Leaders of the biggest AI companies want to slow the pace of development so as to avoid creating “frontier models” capable of wreaking enormous damage on the global economy or worse. Fortunately, this is a goal their investors should be able to get behind too.

Anthropic chief Dario Amodei suggested over the weekend that AI advancements should be “paced”, with governments co-ordinating on standards where possible and companies giving access to third-party experts. Politicians in Washington seem disinclined to help. But even without regulatory coercion, AI hotshots such as Anthropic, OpenAI and SpaceX should be able to find a financial case for reining in the pace of progress.

One advantage of slowing down — besides avoiding the obvious reputational risks of causing a disaster — would be that it would reduce the enormous amounts of cash being lavished on training new models. True, Amodei and peers including OpenAI’s Sam Altman are talking about a slowdown, not a cessation. But that still helps investors. For example, OpenAI plans to spend $750bn on computing capacity by 2030. Every year that gets pushed back reduces the present value of that outlay by almost $100bn, Lex calculates.

The impact on revenue, meanwhile, is less clear-cut. While the best models carry the most pricing power and profitability — OpenAI’s GPT 6 Astra costs $50 per million “tokens” it outputs, more than double its GPT 5.6 Sol model — it’s likely that the bulk of sales already come from cheaper offerings, and AI labs have already been competing on price.

True, one risk of “pacing” is that also-rans will have a chance to catch up and snatch market share. Open-weight models that users can download, run and modify are just a few months behind the leading edge, and the gap has been narrowing, according to the AI Security Institute. China is a formidable competitor on that front, though Amodei suggests keeping cutting-edge AI chips and technologies out of its hands could buy three to five years of time.

But there is plenty of hay to be made from the AI models that already exist, and slowing down would create extra incentive to make it. SpaceX, in its initial public offering, suggested there was a $22.7tn market for “enterprise applications”, meaning software and services that package up AI for real-world use. Anthropic may, in its own IPO filings, hint at a market worth $30tn, the Wall Street Journal has reported.

Those numbers may be pie in the sky, but even more modest estimates leave plenty to play for. Deutsche Bank reckons that if 60 per cent of knowledge workers are using AI-related apps by 2031, they would equate to a $400bn revenue pool. Anthropic’s recent tie-up with Salesforce, embedding the reasoning powers of its Claude bot into the software company’s products, suggests there are many ways to carve up that market.

Any investor in Anthropic — and the list includes such financial titans as BlackRock, Singapore’s GIC, Coatue and Sequoia Capital — should have another pressing reason to support a co-ordinated pacing effort. After all, the extinction of humanity, which would presumably include all of their clients, is a risk for which there is no known hedge.

[email protected]