Ulanqab and the Villages Under the Server Halls
In Ulanqab, a city on the Mongolian Plateau in Inner Mongolia, the sky is wide and the soil is thin. Potato farmers there once timed planting to the last frost and sold tubers to processors in Hohhot and Beijing. Now, bulldozers flatten fields for rectangular data centres that train large language models and run inference for companies in Shenzhen and Shanghai. Village committees sign land-transfer contracts. Households receive compensation in RMB, sometimes a new apartment in a county town, sometimes a one-time payment. The money can look generous against farm income. It does not restore access to the same loess soil, the same windbreaks, or the graves on the ridge. Hukou, the household registration system, still ties many families to a place even after that place is fenced off. People become what some researchers call digital refugees: not fleeing war, but displaced by servers, cooling towers, and fibre-optic trenches.
Compensation and the Limits of a Cash Settlement
Compensation formulas count hectares, crops, and structures. They rarely count the knowledge that passes between generations: which slope holds moisture, when to rotate oats with potatoes, how to read a late frost in the grass. A former farmer in Chahar Right Front Banner can buy cooking oil with the settlement, but he cannot buy back the social role of feeding a village. Younger people often leave for construction jobs or delivery work in Baotou. Older people stay in resettlement blocks with no fields to walk. The data centre campus offers a few security jobs and cleaning contracts. Those jobs go to people with connections or younger bodies. The rest wait. The local government counts the investment as growth. The farmers count the loss in seasons they no longer plant.
Water in a Dry Grassland
AI hardware produces heat. Evaporative cooling removes that heat by letting water absorb it and escape as vapour. The metric WUE, water usage effectiveness, measures litres per kilowatt-hour of IT load. In humid coastal cities, cooling is easier. In arid Inner Mongolia, water is scarce and evaporation is fast. Some campuses draw from groundwater; others rely on municipal supplies or diverted river water. A single large training cluster can consume millions of litres per day in peak summer. That water does not return to the aquifer. It leaves the local hydrological cycle. Herders who move with sheep and goats already compete for wells. Data centres add a new buyer with long-term contracts and political backing. The result is not a metaphor. It is a pump, a pipe, and a water truck queued at dawn.
Green AI and the Coal Behind the Wire
Wind farms and solar arrays do rise near Ulanqab. The region has strong gusts and high irradiance. Officials point to those turbines when they call the data centres green. The claim needs a wider boundary. A server rack contains GPUs, memory, printed circuit boards, and power supplies. Making one GPU requires ultrapure water, rare gases, and electricity often generated from coal in Taiwan, South Korea, or mainland China. Shipping and assembly add emissions. When the wind stops, the grid still leans on coal plants in Inner Mongolia and Shanxi. Batteries can smooth minutes, not weeks. PUE, power usage effectiveness, can look efficient on a dashboard while absolute consumption climbs because the cluster is enormous. Efficiency per unit does not cancel scale. A campus that adds 500 megawatts of load can be cleaner per search and still drain more total water and coal than the villages it replaced.
Cool Air, Hot Chips, and the Rebound Effect
Ulanqab's average temperatures are low, so free-air cooling can run for much of the year. That reduces water and power for cooling. It does not reduce the heat that chips produce under load. When a model trains for weeks, the halls still need mechanical cooling in summer and humidity control in winter. Retrofitting older halls to liquid cooling helps PUE but raises capital costs and creates new maintenance hazards. The rebound effect appears: cheaper cooling makes it possible to pack more GPUs into the same shell. Total demand rises. A campus that once drew 100 megawatts can grow to 500 without a new headline about water. The arithmetic is plain. Cool climate is an advantage, not a solution, when the workload grows faster than the efficiency gain.
Ladders, Overtime, and China Speed
Construction crews work around the clock to meet commissioning dates. Subcontractors hire migrant men from Henan, Sichuan, and Gansu. They weld cable trays, pour concrete, and climb ladders in winter wind that cuts through padded jackets. Many are paid by the day or by the square metre. Overtime is expected. Safety harnesses appear in photographs; on some sites they stay in the tool container because they slow the work. Injuries are handled with cash and a bus ticket home. The men joke about infrastructure mania, but the joke hides a calculation: finish before the inspection, get the bonus, leave before the next crew arrives. Turnover is high. Labour brokers take a cut. The speed that impresses analysts in Beijing is built from twelve-hour shifts, temporary dormitories, and bodies that wear out before the data centre opens.
Who Gets the Compute?
Open-source model weights have changed the AI divide. A developer in Nairobi or Jakarta can download a model and run it on a rented GPU. That lowers the barrier to entry. It does not remove the barrier of compute. Training frontier models still requires thousands of chips, high-bandwidth memory, and data centres with reliable power and cooling. Those resources sit in a few provinces and a few companies. When inference moves to the edge, some power decentralises. When training stays centralised, control over the largest models stays centralised too. The Global South may get open weights but not the grid, the water, or the capital to build a rival campus. The digital divide becomes a divide in energy and land, not only in software licences.
The Rectangular Giants and the People in Their Shadow
Drive past a data centre at night and you see a wall of grey, a fence, and a few blue lights. Behind it, fans roar. In front of it, a village tries to sleep. Dust from construction settles on grazing land. Diesel generators test on Tuesdays. Property values rise for some, rents rise for others. The local tax base grows, but the promised jobs are few and specialised. A herder who once moved across the steppe now negotiates with security guards about a lost goat. A former potato farmer watches a water truck pass his resettlement block. These are not abstractions in a US-China tech race. They are the local price of inference. The question is not whether China can build fast. It clearly can. The question is who absorbs the heat, the thirst, and the displacement when the servers hum.
