Public transcript

Anthropic C.E.O.: Massive A.I. Spending Could Haunt Some Companies

By: The New York Times
Length: 36 min
Language: English
Text from: the video's captions
Transcribed: 12 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

In this on-stage interview with Andrew Ross Sorkin, Anthropic CEO Dario Amodei discusses whether the AI industry is in an economic bubble. He says he is confident in the technology but worried that some players are taking unwise financial risks with massive compute spending. He also covers Anthropic's enterprise focus, competition with Google and OpenAI, chip export policy toward China, regulation, and the impact of AI on jobs.

Key points

Amodei is confident in the technology but concerned some players are taking unwise financial risks with compute spending.
Anthropic's revenue has grown 10x a year for three years and is expected to end this year between $8 and $10 billion.
He describes a cone of uncertainty around future revenue and plans conservatively for the lower end.
He says Anthropic's enterprise focus gives it better margins and a privileged position away from the consumer model wars.
He believes scaling, with small modifications, will get to AGI-level capability and rejects AGI as a single privileged point.
He maintains that advanced chips should not be sold to China because powerful AI is a national security capability.
He rejects accusations of regulatory capture, noting the bills Anthropic supports exempt small startups.
He outlines three levels of response to AI job losses: company-level value creation, government involvement, and long-term restructuring of society.

Questions it answers

00:35Looking back from 2014

What did Dario Amodei expect about AI's trajectory back in 2014?

He says he expected AI to become real and central to the economy, national security, and scientific research. What surprised him was leading one of the companies in the space and the strange lingo and financialization that developed around it.

Amodei is not surprised AI became central to the economy, national security, and science.
He did not expect to be leading one of the companies in the space.

03:48Are we in a bubble?

Is the massive AI spending justified, or is the industry overspending?

Amodei separates the technology, which he is bullish on, from the economics, where he sees genuine uncertainty. He describes a cone of uncertainty around future revenue and the long lag time in building data centers, and says some players are YOLOing and taking unwise risks.

Anthropic's revenue has grown roughly 10x a year for three years.
He plans conservatively for the lower end of a wide cone of uncertainty.
Buying too much or too little compute both carry serious risks.

11:59Circular deals and chip depreciation

Do the vendor-financing style deals between chipmakers and AI companies make sense?

Amodei walks through a stylized deal where an investor funds part of a roughly $50 billion data center build, and says such deals can make sense in principle. The risk comes from stacking them at huge scale, and he argues the key question is how quickly new chips make old chips less valuable.

A gigawatt of compute costs roughly $50 billion in capital expense.
Anthropic assumes aggressive continuation of the chip efficiency curve.
Stacking such deals to huge amounts can lead to overextension.

17:03Model competition and moats

How does Anthropic compete with Google and OpenAI, and is there a moat?

Amodei says Anthropic is on a different path focused on enterprise, while Google and OpenAI fight over consumer. He argues models specialize even at high capability, and enterprise relationships and switching costs create a durable business.

Anthropic's enterprise focus gives it a privileged position away from consumer model wars.
Companies find it hard to switch models because of prompting differences and downstream customers.
He believes scaling, with small modifications, will get to AGI-level capability.

20:26Scaling to AGI

Can scaling alone reach AGI, and when?

Amodei says scaling will get us there with occasional small modifications. He dislikes the term AGI and says there is no single point where models change; they will just keep getting more capable on an exponential, like Moore's law.

He rejects the term AGI as a single privileged point.
Anthropic staff now let Claude write first drafts of code and only edit them.
Models are progressing from high school to college math olympiads.

22:54Chips, China, and national security

Why does Amodei oppose selling the most advanced Nvidia chips to China?

He argues that as models approach a country of geniuses in a data center, which country hosts that capability matters enormously. He says selling chips to China makes it more likely they get there first, and that democracies need to get there first.

He views powerful AI as a singular national security capability, not like the internet or 5G.
Selling advanced chips to China makes it more likely they get there first.
Democracies should use AI aggressively except in ways that make them like their adversaries.

27:23Regulation debate

How does Amodei respond to accusations of regulatory capture and fearmongering?

He says the accusations don't match the bills Anthropic has supported, which exempt small players. He argues those closest to the technology are both excited and worried, and that refusing to regulate for 10 years is like ripping out a car's steering wheel.

He denies Anthropic pursues regulatory capture, noting SB 53 exempts startups under $500 million in revenue.
He compares a 10-year regulation moratorium to ripping out a car's steering wheel.
He considers himself highly optimistic about AI's upsides, including possibly extending lifespan to 150.

31:27Jobs and restructuring society

What should be done about AI-driven job losses?

Amodei outlines three levels: companies favoring value creation over pure efficiency, government involvement such as retraining and possibly tax policy, and long-term restructuring of society so work is about fulfillment rather than survival.

Companies should favor creating new value with AI over pure efficiency-driven job cuts.
He sees a role for government, possibly through tax policy, given large productivity gains.
He invokes Keynes's idea of technological unemployment and a society where work is about fulfillment.

Notes

Background

Dario Amodei, CEO and co-founder of Anthropic, is interviewed by Andrew Ross Sorkin.
He previously worked at Baidu, Google, and OpenAI, where he led development of ChatGPT-2 and 3.
Anthropic is backed by Amazon, Microsoft, and Google.

The AI bubble question

Amodei is bullish on the technology but concerned about the economics.
He and his co-founders were the first to document the scaling laws of AI.
Anthropic's revenue has grown 10x a year for three years: $0 to $100 million in 2023, $100 million to $1 billion in 2024, and an expected $8-10 billion this year.
He describes a "cone of uncertainty" for future revenue, planning conservatively for the lower end.
Data centers take one to two years to build, forcing compute decisions now for revenue in 2027.
Risks: buying too little compute means turning customers away; buying too much could mean bankruptcy.
He says some players are "YOLOing" and taking unwise risks, but declines to name them.

Circular deals and depreciation

A gigawatt of compute costs roughly $50 billion in capital expense with maybe a 5-year useful lifetime.
Vendor-financing style deals can make sense in principle but become dangerous when stacked to huge amounts.
The depreciation question hinges on new chips making old chips less valuable, not chip lifetime.
Anthropic assumes aggressive continuation of the chip efficiency curve.

Competition and moats

Anthropic focuses on enterprise, while Google and OpenAI fight over consumer.
Opus 4.5 is described as hands down the best model for coding.
Enterprise switching costs are high because models have different personalities and prompting patterns.
He believes scaling, with small modifications, will reach AGI-level capability; he dislikes the term AGI and sees only an exponential.

Policy

He maintains his view that advanced chips should not be sold to China, calling it a national security issue.
He rejects the regulatory capture accusation, noting SB 53 exempts startups under $500 million in revenue.
He compares a 10-year regulation moratorium to ripping out a car's steering wheel.

Jobs

Three levels of response: companies favoring value creation over pure efficiency; government involvement including retraining and possibly tax policy; long-term restructuring of society.
He cites Keynes's idea of technological unemployment and a possible 15-20 hour work week.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.