The author is a Reuters Breakingviews columnist. The opinions expressed are her own.

By Robyn Mak

HONG KONG, April 16 (Reuters Breakingviews) - To understand the seismic shift in China's artificial intelligence industry, look no further than a public notice posted in March by the National Committee for Terminology in Science and Technology. The agency proposed an official Chinese term for "token" - the unit of text or data, usually about four English characters long, that large language models process and produce. Its choice of "ciyuan", literally "word currency", is fitting: Chinese firms from Alibaba 9988.HK to MiniMax 0100.HK are betting that tokens will emerge as the most important commodity for the technology. The trouble is, they're only as good as the models they run on. In that regard, China risks falling behind.

The country's token rush kicked off earlier this year, thanks to the viral popularity of digital assistants. Tools like OpenClaw can help manage emails, calendars and other tasks by exchanging text instructions between apps and AI models. They're incredibly token-hungry. Unlike chatbots, which rely on single-prompt queries, the digital assistants can constantly go back and forth with models to complete tasks. OpenClaw can consume orders of magnitude more tokens a day than chatbot sessions.

Since users can decide which models to run their agents on, this trend has delivered a windfall for low-cost Chinese ones. In turn, it has also contributed to a strategy among the country's technology companies: flood the market with cheap access to tokens. China's models are, on average, one-sixth the price per token of U.S. offerings from OpenAI, Anthropic and others, Jefferies analysts reckon. Indeed, daily token consumption in the People's Republic has rocketed to 140 trillion as of March, from 100 trillion at the end of 2025, according to one government estimate. Overseas demand is on the rise too, with officials touting so-called token exports as a potential source of growth.

All of this is underpinned by China's formidable cost advantages in inference, which refers to the moment when customers use an AI model. The People's Republic boasts cheap and plentiful electricity for data centres, helping to compensate for the less powerful and less energy efficient domestic chips. Combined with efficiency breakthroughs in software and algorithms - a necessity for Chinese AI labs handicapped by Washington's tech restrictions - it makes for a potent recipe. Companies like $40 billion MiniMax have seen inference costs fall even as their models have improved in quality, nearing parity with Western rivals. Its 36-year-old CEO, Yan Junjie, predicted last year that inference costs for top models could fall by another order of magnitude within one or two years.

Little wonder Chinese companies have embraced tokens. Last month e-commerce titan Alibaba separated its AI business from its cloud computing arm and rebranded it to the Token Hub Business Group. The appeal is that charging for tokens rather than selling subscriptions is a much simpler business model. MiniMax describes its February release, which costs $1 to run continuously for an hour at a rate of 100 tokens per second, as "the first frontier model where users do not need to worry about cost, delivering on the promise of intelligence too cheap to meter". That makes tokens sound like infrastructure or utility products, akin to electricity. It's one way China could stay competitive even as its labs suffer from U.S. chip-sale bans: focus on efficiency and volume, and trust that the benefits of low costs will win out over time, echoing elements of the country's traditional industrial playbook.

Still, there are problems with this approach. One is that AI risks becoming a utility-like commodity prone to destructive price wars and overcapacity. Another is that too-cheap-to-meter AI is only applicable to a certain extent. For enterprises, model quality will matter just as much, if not more, than cost efficiency as they outsource more complex and high-value work to digital agents.

Chinese models from Alibaba, DeepSeek, Moonshot, Zhipu 2513.HK and others have narrowed the performance gap with the West in benchmarks like coding, reasoning and math. But they still trail in overall frontier capabilities, a broader measure of aggregate performance, versatility and reliability. Take Anthropic's latest Mythos system, which the company claims is so powerful it will only be initially available to vetted firms including JPMorgan JPM.N , Amazon.com AMZN.O and Microsoft MSFT.O. One million tokens of that model are worth more than the same number processed through an inferior one. In other words, tokens are not fungible.

The bet may be that Chinese labs will eventually move up the value chain, similar to the trajectory of the country's manufacturers in electronics, cars and solar panels. But that seems a long shot in AI, given Washington's restrictions on advanced chips and chipmaking equipment. If anything, U.S. labs are likely to widen their lead in overall frontier capabilities, thanks to Nvidia's NVDA.O highly-anticipated Vera Rubin generation of graphic processing units due later this year. These new semiconductor systems promise 10 times the performance per watt of their Grace Blackwell predecessor, and American export rules mean they probably won't be available to the People's Republic.

Against Nvidia's advances, domestic chipmakers like Huawei are struggling even to match the performance of the U.S. titan's older H200 processors. Those are the most powerful chips legally available for training models in China, and they're already two generations behind Rubin. Earlier this year, Justin Lin, who formerly led Alibaba's open-source models division, lamented a "massive amount of OpenAI’s compute is dedicated to next-generation research, whereas we are stretched thin — just meeting delivery demands consumes most of our resources".

Investors are pessimistic. Alibaba's market capitalisation of about $315 billion, for example, is below the private price tag of $852 billion for OpenAI and $380 billion for Anthropic. It's notable because the Chinese group also includes a giant e-commerce platform, which JPMorgan analysts reckon will generate 196 billion yuan ($29 billion) of earnings in 2027. Apply a conservative 10 times multiple to that unit, and the implication is that investors ascribe almost no value to the cloud and AI businesses, which are targeting $100 billion of combined annual revenue within five years.

Low-cost tokens play to China's industrial strengths and will speed up adoption of new technologies across the economy. But the U.S. remains comfortably ahead in computing power and frontier models. That makes China's token boom less decisive than it looks.

Follow Robyn Mak on X.