IGI Logo

AI honchos are suddenly asking for restraint

The last few days have seen a remarkable change in tone from the people building frontier AI. Researchers from Anthropic have publicly discussed the possibility of human extinction, Dario Amodei has called for frontier development to be deliberately paced, Elon Musk said, “Dario is right,” Sam Altman said, “we need to pace the frontier,” and Demis Hassabis backed the same direction. The stated concern is that autonomous systems, cyber capabilities and recursive self-improvement may advance faster than human ability to control them. For the first time, the biggest AI labs themselves are openly discussing whether the race needs to slow.

But there may also be a very simple economic reason

The economics of frontier AI are deteriorating rapidly because intelligence is becoming a commodity. Chinese models are priced roughly 80% below Western models, while AI token prices have already fallen sharply. Our calculation suggests that at a 5% hurdle rate (which is minimum needed as yields are hardening), another frontier model stops clearing economically if Chinese models gain only around 3.2 percentage points of market share annually; at an 8% hurdle, the threshold falls to 0.6 point, while at 10% and above the incremental investment fails even without further Chinese share gain. The AI arms race may therefore be approaching the same problem that destroyed returns in telecom fibre: spectacular demand growth but collapsing unit economics.

The Chinese threat is strongest at the frontier

The median Chinese model costs around US$0.83 per million blended tokens versus US$4.50 for Western models, an 81.5% discount; at the extremes, the gap can exceed 200x. Importantly, however, the discount is concentrated against premium Western frontier models rather than Western AI generally. GPT-5.6 Luna and Gemini Flash already compete aggressively with Chinese pricing, implying that Western labs are building a barbell strategy: defend commodity inference on price while charging heavily where frontier capability remains differentiated.

Even after adjusting for quality and speed, China remains cheaper

Price alone exaggerates the Chinese advantage because slower or weaker models impose latency and supervision costs. Our Delivered Intelligence Cost (DIC) metric adjusts token cost for speed and quality. On this measure, the Chinese advantage narrows from 81.5% to roughly 34.9%, but it does not disappear. DeepSeek V4 Flash remains the genuine outlier at around US$0.142 of DIC, roughly one-third of the Western frontier level, driven by both low price and very high generation speed. If Chinese models continue closing the intelligence gap while retaining this cost advantage, Western frontier economics become increasingly difficult to defend.

The real variable is revenue share – not rhetoric

The thesis is not yet proven. Chinese models still trail badly on revenue, regulated Western customers may refuse to use Chinese APIs, export controls could restrict Chinese compute access, and the quality gap at the top remains meaningful. The single most important variable to watch is therefore Chinese lab revenue share, not Chinese model pricing. If cheap prices start translating into sustained revenue and the intelligence gap narrows, the economics of spending tens of billions on every new frontier generation deteriorate very quickly. If they do not, the Western barbell survives and the current slowdown rhetoric may prove premature.

Above views are of the author and not of the website kindly read disclaimer