Inner Mongolia is where China’s AI ambitions are turning into concrete. In the grasslands around Ulanqab, farmers are lifting potatoes next to rows of huge rectangular buildings that will run the country’s AI platforms. A BBC team visited this week, just as Xi Jinping travelled to Washington for a summit with Donald Trump at which AI safety is on the agenda. What it found was a construction boom that local workers describe, only half joking, as “China speed”.
The story is bigger than one city. Inner Mongolia combines cold air, cheap electricity, large wind and solar resources and a short fibre run to Beijing. That mix has drawn DeepSeek, Huawei, Alibaba, Apple and an energy company building what it calls the world’s largest single AI data centre building. It also raises harder questions about who benefits, what powers the machines and how far China can get without Nvidia’s best chips. We covered the chip side of that race in Huawei’s plan for a new AI chip in early 2027.
This article sets out what the BBC saw, the numbers behind Ulanqab’s rise, the projects being built, the power arithmetic that makes Inner Mongolia attractive, and what it means for businesses outside China that depend on AI compute.
Table of contents
- What the BBC Found in Inner Mongolia
- Ulanqab: Inner Mongolia’s “Capital of Token”
- Why Inner Mongolia Makes Cheap Compute
- Who Is Building in Inner Mongolia
- Inner Mongolia and the Chip Question
- The East Data, West Computing Map Beyond Inner Mongolia
- Who Benefits on the Ground in Inner Mongolia
- Talent: The Other Half of China’s AI Build-Out
- What Inner Mongolia Means for Businesses Outside China
- Frequently Asked Questions About Inner Mongolia’s AI Data Centres
- References
What the BBC Found in Inner Mongolia
The BBC report, published on 22 September 2026, describes Ulanqab as “an unlikely digital frontier”. It is a city of steppes and extinct volcanoes that has long called itself China’s official “potato capital”.
Potato fields and construction cranes
Hundreds of workers in yellow helmets carry equipment into building sites while cranes lift rooftops into place. The data centres sit beside working farmland, and the BBC describes potato pickers working next to them. The scene captures how fast the build-out in Inner Mongolia has moved from plan to steel.
“This is China speed”
“This is China speed,” one worker told the BBC, calling his country “an infrastructure maniac”. An electrical engineer explained the phasing: “There is one phase every year. This is the second phase. There will be many more phases afterwards.” Nobody on site expected the building to slow down soon.
Why Beijing treats it as national security
To the Chinese Communist Party, these buildings are critical to economic and national security. Once finished, they will support Beijing’s aim to embed AI across 90% of industries and society by 2030. The computing power inside Inner Mongolia’s data centres is, in the BBC’s words, the key to China’s contest with the United States.
The summit backdrop
The report landed as Xi arrived in Washington for a White House summit on Thursday 24 September. “Whoever wins AI, wins,” Trump said the week before. Anthropic chief executive Dario Amodei has called for keeping the US ban on selling advanced chips to China. China’s state-run Global Times dismissed such US fears as part of a “Cold War playbook”.
Ulanqab: Inner Mongolia's "Capital of Token"
Ulanqab’s local government has set itself the goal of becoming the “Capital of Token”, a reference to the units of text that AI models read and write. The figures reported by 36Kr, a Chinese business publication, show why.
100 projects and 500 billion yuan
In about two years, Ulanqab has signed 100 data centre projects with total planned investment of more than 500 billion yuan. Investors include Huawei, Alibaba, Apple, Kuaishou, DeepSeek and China Data. The city’s entire GDP is about 126 billion yuan, so the planned investment is roughly four times the size of the local economy.
Compute scale and national share
Ulanqab’s operating computing capacity is about 172,000 P, meaning petaflops of processing power. It is expected to pass 200,000 P by the end of 2026, a year-on-year increase of more than 112%. That already represents 8.7% of China’s national computing capacity, concentrated in one prefecture of Inner Mongolia.
Revenue and electricity
In the first half of 2026, Ulanqab’s computing industry earned more than 3.66 billion yuan and used 3.267 billion kWh of electricity, up 89.52% on a year earlier. The city’s target for the full year is more than 8 billion yuan of revenue.
| Measure | Ulanqab figure | Context |
|---|---|---|
| Data centre projects signed | 100 | In roughly two years |
| Planned investment | Over 500 billion yuan | City GDP about 126 billion yuan |
| Operating compute | 172,000 P | 8.7% of China’s national total |
| Compute by end of 2026 | Over 200,000 P | Up more than 112% year on year |
| H1 2026 industry revenue | Over 3.66 billion yuan | Full-year target over 8 billion yuan |
| H1 2026 electricity use | 3.267 billion kWh | Up 89.52% year on year |
The build-out is still accelerating. The chart compares today’s operating compute with the end-2026 expectation.
Why Inner Mongolia Makes Cheap Compute
Cold weather is the obvious explanation, and it is only part of the story. Inner Mongolia wins on three things at once: climate, distance and, above all, the price of electricity.
Cold air and free cooling
Ulanqab sits at about 42 degrees north with an average annual temperature of 4.3°C. Data centres can use outside air for cooling for nearly ten months a year. Operators say that can bring power usage effectiveness below 1.2, meaning less than 20% of the electricity goes on anything other than computing. Huawei’s Ulanqab site reported an annual figure of 1.32, falling to about 1.22 in months of full natural cooling.
Close enough to Beijing
Ulanqab is roughly 300 to 350 kilometres from Beijing. A dedicated fibre link gives a round-trip network delay of about 4.2 milliseconds. That lets technology companies in Beijing run their computing in Inner Mongolia with almost no noticeable lag, which cold cities much further away cannot offer.
The grid accident that made power cheap
The deciding factor is price. Ulanqab sits inside the West Inner Mongolia Power Grid, an independent regional grid that was left outside State Grid when China split up its power system in 2002. For years it generated more electricity than it could export, so prices fell. Today data centres there pay about 0.33 to 0.36 yuan per kWh, compared with about 0.8 yuan in eastern Inner Mongolia and 0.6 to 0.7 yuan in Beijing.
Electricity is the largest running cost of an AI data centre, and Ulanqab’s price is roughly half the going rate. The chart uses the top of each published range.
Wind, solar and the green-power rule
Inner Mongolia also has large wind and solar resources. China’s National Development and Reform Commission requires new intelligent computing centres to source more than 80% of their power from green electricity. In July 2025 China Data connected a green-powered computing base in Ulanqab that combines 200 MW of wind, 100 MW of solar and 45 MW of storage, designed to produce about 848 million kWh a year for 25,000 server cabinets.
Who Is Building in Inner Mongolia
The list of companies in Ulanqab reads like a directory of Chinese AI, plus one of the world’s best-known American brands.
DeepSeek’s gigawatt campus
DeepSeek plans at least 1 GW of computing capacity in Ulanqab, partly built itself and partly leased from other operators. Some capacity could come online by late 2027 or early 2028. Bloomberg reported on 4 September that DeepSeek plans to install at least 160,000 of Huawei’s Ascend 950DT accelerators there. At full use, a gigawatt campus would draw enough energy for roughly 750,000 homes.
Envision’s Galaxy Campus
On 6 August, energy technology company Envision commissioned its Galaxy Campus in Ulanqab. It is designed to scale beyond 2 GW, powered directly by renewable energy. Its central building covers 120,000 square metres, about 20 football pitches, and is designed for up to one million AI accelerators and one million PFLOPS of computing. Envision calls it the world’s largest single AI data centre building.
Huawei, Alibaba and Apple
Huawei, Alibaba and Apple already operate data centre infrastructure in the area. Apple’s presence stands out, because it shows that Inner Mongolia’s appeal is not limited to Chinese AI labs. Local rules on data storage mean international companies serving Chinese customers also need capacity inside China.
| Project | Scale | Status |
|---|---|---|
| DeepSeek campus | At least 1 GW; 160,000 Ascend 950DT chips planned | Partial operation late 2027 or early 2028 |
| Envision Galaxy Campus | Designed to exceed 2 GW; up to 1 million accelerators | Commissioned 6 August 2026 |
| China Data green-power base | 200 MW wind, 100 MW solar, 45 MW storage; 25,000 cabinets | Connected July 2025 |
| Huawei Ulanqab data centre | Annual PUE 1.32, about 1.22 in free-cooling months | Operating |
Why the projects cluster
More than a dozen companies have received approvals for projects in Ulanqab, so new entrants join an existing cluster rather than starting in empty grassland. Shared fibre, substations and skilled contractors make each new Inner Mongolia project cheaper and faster than the one before.
Inner Mongolia and the Chip Question
Buildings and power are only half of a data centre. The other half is the chips, and that is where US export controls bite.
Building around Huawei
DeepSeek intends to use the Ascend 950DT chips for running its models, not for training them, according to Bloomberg’s sources. Huawei’s output of the 950DT will be limited to the low hundreds of thousands this year by shortages of components such as high-end memory, so DeepSeek’s order could take more than a year to fill. The 950DT is widely seen as comparable to Nvidia’s previous Hopper generation.
How far China has come
Six months ago, a Shenzhen newspaper reported the opening of China’s first 10,000-chip computing cluster using Huawei’s Ascend processors. A 160,000-chip installation in Inner Mongolia would be sixteen times that size. Food delivery giant Meituan said in June that it trained a model on 50,000 domestic chips, and Z.AI completed a 1 GW data centre in July that it plans to fill with Chinese chips. Our piece on GLM-5.3-Flash running on Chinese chips shows what that hardware can already do.
The Nvidia allegation
Trump administration officials have alleged that DeepSeek obtained Nvidia’s top Blackwell chips and installed them in Inner Mongolia. A senior US official told Reuters in February that DeepSeek had trained an advanced model on Blackwell processors despite export restrictions. Bloomberg has not independently verified the claim, and DeepSeek has not confirmed how any such chips were obtained.
Why inference comes first
Running models, known as inference, is where most of the day-to-day computing demand sits once a model is in use. By putting Huawei chips into inference first, Chinese labs can serve large numbers of users on domestic hardware while still relying on Nvidia for training. Inner Mongolia’s cheap power matters even more for inference, because it runs around the clock.
The East Data, West Computing Map Beyond Inner Mongolia
Ulanqab is part of a national plan, known as East Data, West Computing, to move computing from crowded eastern cities to western and northern regions with more land and power.
The other hubs
China is building hundreds of data centres for companies including Huawei, ByteDance and DeepSeek in thinly populated provinces. The BBC names Ulanqab, Ningxia and Gansu in the north and Guizhou in the south-west. Inner Mongolia’s advantage over the others is its combination of cold climate and proximity to Beijing.
Renewables, land and water
These regions already have large wind and solar capacity, cheaper land and cooler weather, which reduces the water needed for cooling. That contrasts with the US, where water and power use are now political issues. California’s new rules, covered in California tightens rules on AI data centre energy and water, are one example.
Coal is still in the grid
China wants most of its data centres to run on green energy by 2030, but some of its grid still depends on coal power. Water supply also remains a concern in the arid parts of Inner Mongolia. The green-power share of any given facility depends on whether it has dedicated renewable supply or draws from the wider grid.
| Factor | Inner Mongolia advantage | Open risk |
|---|---|---|
| Electricity | About half the Beijing price | Part of the grid is still coal-fired |
| Cooling | Nearly ten months of free cooling | Water scarcity in arid areas |
| Network | About 4.2 ms round trip to Beijing | Distant from southern China |
| Chips | Large Huawei orders being placed | Huawei output capped by memory shortages |
| Land | Cheap and plentiful | Village relocations and local opposition |
Who Benefits on the Ground in Inner Mongolia
The BBC’s most striking material came from the people who live next to the data centres, and it is much less upbeat than the investment figures.
Relocated villages
Entire villages have been relocated to make way for the build-out, although the BBC found it hard to establish how many, or what compensation residents were offered. Few people will discuss it, because voicing concerns too openly can be dangerous.
“What benefits has it brought us?”
One local, speaking anonymously, said the data centre boom had “absolutely nothing to do with ordinary people. What benefits has it brought us?” He said few locals had found jobs there, and others feared losing their land. A sheep farmer told the BBC he had no use for AI and barely used the internet. His phone screen was smashed, and his wifi came from a pocket dongle costing just over 40 yuan a month.
The optimists
Others were more positive. A man named Zhang, standing in a potato field near the data centres, said his children had found jobs there. “China is now at the forefront of the world,” he said. “It’s not like the 1960s, when it was relatively backward.”
Where the money goes
36Kr’s analysis explains the gap between the investment and local benefit. A large data centre employs only 200 to 500 permanent staff once built. Of the 500 billion yuan being invested, it estimates that 40% to 50% goes to chip, memory and optical suppliers, 15% to 20% to server makers, and only 10% to 15% stays in Ulanqab as electricity bills, land rent and construction work.
Most of the investment leaves Inner Mongolia before a single model runs. The chart shows the midpoint of each estimated range.
The resource-city trap
36Kr compares Ulanqab with Ordos, another Inner Mongolia city built on energy. Ordos had China’s highest GDP per person in 2025, at 272,700 yuan, which was 4.3 times the 63,946 yuan disposable income of its urban residents. The risk is that compute repeats the coal pattern: high output, with little of the profit reaching local people.
Talent: The Other Half of China's AI Build-Out
The BBC report makes the point that China is not only building data centres in Inner Mongolia. It is also building a pool of people to use them.
The returnees
“Fewer students and researchers are heading to the US, and more are returning from the US to China. That is very clear,” said Jack Zhang of the University of Science and Technology of China. He was recruited back after studying at Yale and Cambridge, and said he returned partly for the resources and government support for AI in medicine.
Five million graduates a year
China produces around five million science and technology graduates each year, compared with around half a million in the US. As more of them stay, they find a growing ecosystem. Dang Ronghao, a 27-year-old AI developer in Shanghai, described research combining AI with nuclear fusion, astrophysics, new materials and drug discovery.
“Not as open as I wish”
Not everyone is convinced. Professor Liu Jun of Tsinghua University, who has taught at Harvard and Stanford, said Chinese education still rewards conformity. “I keep encouraging students here to ask more questions, to challenge more, but that’s hard to change,” he told the BBC.
Open-source models as strategy
Xi has encouraged Chinese AI firms to develop cheaper, open-source models that others can adapt, in contrast with the closed approach of OpenAI and Anthropic. Sharing the technology has helped Chinese firms move fast despite limited access to advanced chips. The compute in Inner Mongolia is what lets those open models be trained and served at scale.
What Inner Mongolia Means for Businesses Outside China
A UK business will never rent a rack in Ulanqab, but what happens there still affects the price and availability of AI.
Power is the real constraint
Inner Mongolia shows that the winners in AI infrastructure will be those with cheap, plentiful and politically secure electricity. Western data centre projects face grid connection queues and local opposition, as we saw in X’s claim about a Chinese bot farm posting anti-data-centre messages. Our data centre operations team sees the same pressure in UK colocation pricing.
Cheap Chinese models will keep coming
Cheap power and domestic chips mean Chinese labs can keep offering low-cost, open models. That puts downward pressure on AI prices globally, as DeepSeek’s earlier API price cuts showed. It is good news for buyers, provided they check where their data goes.
Data residency and compliance
Models served from Inner Mongolia run under Chinese law. Any business using a hosted Chinese model API should treat it as a cross-border data transfer and check it against UK GDPR and its own cybersecurity policies. Downloading open weights and running them on your own infrastructure avoids that particular risk.
Treat efficiency claims with care
DeepSeek became famous for doing more with less computing power. Its gigawatt plans in Inner Mongolia show that efficiency gains lead to more demand, not less. Businesses planning AI budgets should expect usage to grow as prices fall, and set limits accordingly.
Frequently Asked Questions About Inner Mongolia's AI Data Centres
Why is China building AI data centres in Inner Mongolia?
Cheap electricity, a cold climate that allows free cooling for about ten months a year, large wind and solar resources, cheap land and a low-latency fibre link to Beijing.
What is Ulanqab?
A city in central Inner Mongolia, long known as China’s potato capital, which has signed 100 data centre projects worth more than 500 billion yuan and now holds about 8.7% of China’s computing capacity.
Is DeepSeek building in Inner Mongolia?
Yes. DeepSeek plans at least 1 GW of capacity in Ulanqab and, according to Bloomberg, at least 160,000 Huawei Ascend 950DT chips for running its models.
How cheap is electricity there?
About 0.33 to 0.36 yuan per kWh for data centres, roughly half the 0.6 to 0.7 yuan paid in Beijing.
Do local people benefit?
Some find jobs, but a large data centre employs only 200 to 500 permanent staff, and an estimated 10% to 15% of the investment stays in the local economy. Villages have also been relocated.
Does this affect businesses in the UK?
Indirectly. It keeps Chinese AI models cheap and competitive, and it adds to pressure on power and data centre costs worldwide.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
