Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
The price of using artificial intelligence is falling. In China, entry-level access to AI models has been marketed for about $3 a month by the country’s hottest AI group Zhipu. Widely seen as one of China’s closest local equivalents to OpenAI, it develops large language models to compete with US systems.
In a market where paid plans from US peers are closer to $20 a month depending on tier, and enterprise contracts run into the millions annually, this move by Zhipu amounts to a price war.
That gap goes beyond product pricing. Since its Hong Kong debut in January, shares of Knowledge Atlas Technology Joint Stock, better known as Zhipu AI, have gained 300 per cent. Yet even after that rally, its market value of $29bn remains a fraction of top US AI groups. OpenAI’s private market valuation stands at $500bn and Anthropic was valued at about $350bn in its latest funding round.
Markets are pricing a world in which US AI groups maintain outsized control over global AI revenue and dominate the highest margin segments, while global users continue to accept higher price points. But how sustainable is that assumption?
There are good reasons US AI developers attract higher valuations. They have larger existing revenue bases than Chinese peers and benefit from established enterprise relationships and easier access to capital. OpenAI and Anthropic are backed by US cloud providers, giving them scale.
Geopolitics also matters. Data security concerns and regulatory scrutiny would probably limit the use of Chinese AI models in the US and parts of Europe, particularly for government workloads. That constraint should help preserve pricing power for US providers in some markets.
Yet the technological gap between US and Chinese AI models is narrowing. On standardised industry benchmarks, leading Chinese LLMs now operate close to their US counterparts across reasoning, coding and general knowledge tasks. In areas such as mathematical problem solving, the gap has narrowed to low single-digit percentage differences, according to Stanford’s 2025 AI Index. China also accounts for a large chunk of global AI research output and highly cited machine learning papers.
The biggest perceived constraint on China’s AI ambitions has been access to advanced AI chips and whether it could create a viable domestic workaround to US export controls. Zhipu has trained its models on Huawei’s Ascend AI chips. While Huawei chips do not match the cutting-edge performance of Nvidia’s latest chips, its use suggests that export restrictions have not eliminated China’s ability to build credible rivals to US models, especially if improvements in chip design and software help offset lower computing power.
If technology is no longer the binding constraint, competition then shifts to price. OpenAI’s pricing for developers of its flagship GPT-5.2 model is about $1.75 per million input tokens, units used to measure text processed and generated by AI models, and $14 per million output tokens. On Zhipu’s domestic developer platform, its latest flagship language model GLM-5 is priced at about $0.58 per million input tokens and $2.60 per million output tokens. The difference is most striking in output pricing, where it undercuts most US models.
Even in markets where US AI companies benefit from brand recognition and distribution advantages, a sustained price gap can shift behaviour over time. As AI becomes part of everyday work, this is particularly true for students, independent developers, small businesses and cost-sensitive start-ups.
That means Zhipu’s current valuation may reflect overly pessimistic expectations. Global consumer spending on generative AI alone is forecast to approach $700bn by 2030, according to Counterpoint Research, before accounting for enterprise and cloud infrastructure revenue. If markets more receptive to Chinese AI providers were to capture a third of that demand and Zhipu took just 10 per cent of that market, revenues would surpass $23bn by the end of the decade. Even using conservative revenue multiples and discount rates, that would support a valuation far above today’s level.
Much of the optimism behind US AI stocks rests on the assumption that users will keep paying for incremental performance gains. But as differences narrow, the price of AI may soon be set by the cheapest model that is good enough.
