The Word Factories of Ningxia: Why China’s Telecoms Took a Profit Hit to Chase AI Compute

China’s state telecoms posted first-half earnings this month that told two contradictory stories at once: China Unicom’s net profit fell 34.6% even as its computing-power revenue jumped 13%, and China Telecom’s core telecom business grew modestly while its intelligent-computing revenue surged 95%.

Call it the barrel-of-oil moment for artificial intelligence. Chinese state media has begun describing the country’s cheap, abundant electricity as a raw commodity that gets refined, not into gasoline, but into AI tokens, and shipped out at a fraction of what US model providers charge.

The Pivot Hits the Books First

China Unicom’s interim results, filed August 18, show operating revenue essentially flat at RMB201.4 billion, up 0.6% year on year.

Net profit, meanwhile, dropped to RMB9.47 billion. The company attributes the shortfall to a VAT reclassification that pushed mobile data and broadband services from a 6% to a 9% tax rate, plus a 19.3% jump in labor costs. Computing-power revenue, the one line item management wanted the market to notice, rose 13% to RMB41.9 billion, and the share of capital expenditure devoted to compute climbed to 37%, an increase of more than 80% year on year.

China Telecom told a version of the same story with a better ending: intelligent-computing revenue up 95%, intelligent-business revenue up 7.1% to 31.1 billion yuan, cloud revenue up 7.8% to 61.8 billion yuan. The company’s Xingchen TokenHub platform, which now hosts 142 large language models and more than 420 industry-specific applications, is the retail face of the same repositioning: bandwidth was the old product, tokens are the new one.

Neither earnings call disclosed matching figures from China Mobile, the country’s largest carrier, though it appears elsewhere in this story as an infrastructure operator rather than a reporting company.

Electricity, Rebottled

It’s not about just power generation anymore; the focus is on manufacturing tokens with that electricity.

That is the entire thesis running underneath an essay published by China’s Economic Daily, which argues that a kilowatt-hour of cheap Ningxia solar power, once converted into AI inference, can be “magnified tens to hundreds of times in value” and sent anywhere in the world at the speed of light.

The piece cites estimates that domestic AI inference costs one-tenth to one-sixth of overseas equivalents, and frames the resulting token exports as China’s answer to being called “the world’s factory:” not a nation that ships hardware, but one that ships intelligence.

Whether a market exists to actually price that intelligence the way the essay assumes is a separate question. And not one it answers.

Where the Tokens Actually Get Made

Rows of servers. Half of all tasks answered within six seconds, 90% within ten, response times held within a 20% band. That is Beijing’s first token factory, the Beijing Yihao Token Factory, built by the IT services firm iSoftStone and opened in the Beijing Economic-Technological Development Area in June.

Phase one produces 1.4 trillion tokens a day; the stated long-term target is 10 trillion.

iSoftStone also open-sourced a benchmark, LoadGen 2.0, meant to standardize how the industry measures a computing cluster’s real service quality under the “agentic,” multi-step workloads that today’s AI systems actually run, as opposed to the single-shot chatbot exchanges the industry used to benchmark against.

The factory’s stated plan to eventually draw power from green bases in Zhangjiakou and Ulanqab is aspirational, not yet built. The infrastructure that already exists sits nearly a thousand miles west, in Zhongwei, Ningxia, where China Mobile runs a data center now drawing more than 80% of its power directly from an adjacent 500-megawatt solar farm, with a 1.5-gigawatt wind farm under construction nearby.

Zhongwei is one of ten national data-center clusters under China’s “East Data, West Computing” program, launched in 2022 to move compute toward cheap land and renewable power instead of the other way around. Six of China’s ten largest internet data-center operators now have a presence there.

The Catch

Power is the whole appeal of Zhongwei. Power is also the thing experts say China hasn’t solved.

Data centers are, by industry consensus, a poor match for renewable providers compared with traditional heavy industry, because their peak demand is difficult to forecast and, once GPUs are purchased, operators want to run them constantly rather than flexibly.

“They do not appear to be very flexible,” Pei Shanpeng, a director at China’s State Power Investment Corp, said at a Beijing industry conference, adding that the push for green power in data centers is driven more by emissions targets than cost savings.

China’s data-center power demand is projected to grow by 300 to 500 billion kilowatt-hours between 2026 and 2030, a range whose low end alone would roughly match the United Kingdom’s total annual consumption, according to figures Pei cited. Grid operators, for their part, are wary of dedicated green-power links that could leave them holding the cost of transmission infrastructure if data-center demand ever slows.

The Bigger Ledger

For scale: mobile technology and services, GSMA Intelligence estimates in its Mobile Economy China 2026 report, contributed $1.5 trillion to China’s economy in 2025, or 7.2% of GDP, and are projected to reach $2.1 trillion, or 8.4% of GDP, by 2030.

The same report documents China Mobile’s Zhongwei buildout as a green-infrastructure case study in its own right, describing the identical solar-and-wind project cited above in the vocabulary of national accounts rather than digital exports.

The two framings, one measuring GDP contribution and the other measuring geopolitical leverage, describe the same facility without ever quite talking to each other.