The Water Ledger in Inner Mongolia
In Ulanqab, Inner Mongolia, the same arid ground that grows potatoes now hosts rows of server halls. Evaporative cooling, the cheapest way to keep AI accelerators from overheating, consumes water at a rate measured by water usage effectiveness, or WUE. A typical facility may cycle millions of litres per day. Local reports describe household wells dropping and small reservoirs shrinking. The green energy claim often focuses on electricity: wind turbines on the steppe, solar arrays in the desert. Water is a different ledger. Wind and solar cannot cool a GPU. The region receives less than 300 millimetres of rain a year. Aquifer recharge is slow. When data centre operators drill deep wells, they compete with farmers who cannot drill deeper. The 'East Data, West Computing' strategy, known in Chinese as 东数西算, sends computation to western provinces for cheaper land and power. It also sends water demand to places that already ration irrigation. A PUE of 1.2 looks efficient on a corporate sustainability report. It says nothing about WUE. That omission matters where water, not electricity, is the binding constraint.
Cooling technology choices are not neutral. Chilled water systems use less water but more electricity. Direct evaporative cooling uses less electricity but more water. In a dry climate, the trade-off tilts toward water. Some operators report WUE below 1.0 litres per kilowatt-hour, but those figures often exclude water consumed in coal mining and power generation upstream. A life-cycle accounting changes the sum. Inner Mongolia's coal mines also require water for dust suppression and washing. The same watershed supplies data centres, farms, and mines. When the water table falls, the first wells to fail are usually the shallowest, which belong to villagers. The green energy claim does not address this competition.
When the Grid Bends Toward Coal
Rapid construction strains local grids. A single hyperscale campus can demand hundreds of megawatts, comparable to a small city. Wind and solar farms are built alongside, but their output varies. Grid operators must keep coal-fired baseload running to prevent voltage collapse. In Zhangjiakou, Hebei, curtailment of renewable power has been documented even as new data centres connect. The result is a paradox: a facility marketed as green may increase coal generation at the margin because stable power requires spinning reserves. Ultra-high voltage, or UHV, transmission lines move electricity from west to east. They do not eliminate local balancing needs. Coal plants cycle up and down, which raises emissions of sulfur dioxide, SO2, and fine particulate matter, PM2.5. The NDRC, China's National Development and Reform Commission, has issued quotas for data centre efficiency. Local governments compete for investment and sometimes relax enforcement. The friction is not between China and the United States. It is between a server hall's cooling load and a provincial grid's ability to absorb intermittent renewables without coal. Every new AI cluster makes that balancing act tighter.
Grid stability has a physics that marketing cannot bypass. Alternating current must match supply and demand within milliseconds. Batteries and pumped hydro can help, but capacity is limited. In several western provinces, new data centres have signed contracts with coal plants for firm capacity. Those contracts may be labelled 'transitional.' They still burn coal. The emissions accounting often assigns them to the power sector, not the data centre. That accounting choice lets an AI company claim a low carbon footprint while the local air quality worsens.
Potato Fields to Server Halls
Villages have been relocated to make room for industrial zones. In several cases, potato farmers lost prime agricultural land with irrigation access. Compensation is often a one-time payment. It does not restore a livelihood. Food security is local as well as national. When a county converts its best soil into concrete pads for transformers and backup generators, it imports more food from elsewhere. Rural households lose a source of income and a source of calories. The national framing calls this 'economic security.' For a farmer in Ningxia or Gansu, it looks like livelihood insecurity. New jobs in data centres require electrician certificates, security clearances, or software skills. A 60-year-old potato grower does not qualify. Younger workers may leave for cities. The village tax base changes. Local reports on land use show a shift from crop rotation to industrial zoning. The tension is not abstract. It appears in the price of potatoes at a county market, in the number of children who stay in school, in the maintenance of irrigation canals now serving cooling towers. The infrastructure of AI ambition has a footprint measured in hectares, not just in FLOPS.
Displacement is not only physical. It is administrative. Household registration, or hukou, ties access to schooling and healthcare to a locality. When a village is dissolved, residents may be reclassified as urban or moved to a resettlement site. The paperwork can take years. Meanwhile, the land is rezoned. Provincial officials report GDP growth from construction. The same officials may not report the loss of a seed potato cooperative. Food security plans at the national level can coexist with local food deserts. The two scales do not automatically align.
Open Source, Centralized Control
China's open-source push is often described as a strategic bypass for chip sanctions. Projects such as OpenEuler, PaddlePaddle, MindSpore, and RISC-V-based designs are released under permissive licenses. That part is real. The framing stops there. Open source does not guarantee decentralized power. A foundation can be guided by state ministries. Telemetry, update channels, and compliance audits remain points of control. When a government promotes domestic open-source stacks for data centres, it also creates dependencies on domestic hardware, domestic cloud operators, and domestic standards. Global data sovereignty becomes complicated. A country that adopts these stacks may gain alternatives to US vendors. It may also accept a new set of chokepoints. The code is open. The supply chain is not. The governance model is not. Decentralized technology can be governed centrally through procurement rules, security reviews, and export controls. The question is not whether China's open-source model beats the closed model. The question is who holds the keys to updates, identity systems, and data pipelines. That answer shapes who can audit an AI model and who cannot.
Data sovereignty is a term with many definitions. For some governments, it means data stored on national territory. For others, it means control over the software stack. China's open-source model offers a third path: code that is globally readable but state-aligned in practice. A foreign ministry can require that public sector AI systems use certified repositories. That requirement is not in the license. It is in the procurement rule. Decentralized development can thus produce centralized deployment. The risk is not that open source fails. The risk is that it succeeds under conditions that concentrate power over updates, audits, and identity.
Who Pays for the Infrastructure of the Race
The usual story is a race. Washington and Beijing compete for AI leadership. The Global Times and US officials use the same vocabulary. That vocabulary hides the bill. The bill is paid in water from Inner Mongolian aquifers. It is paid in coal emissions when grids need stability. It is paid by potato farmers who lose land and receive compensation that does not last. It is paid by local governments that borrow to build power lines and roads for campuses that may become obsolete. Cross-referencing claims of unprecedented growth with provincial land-use records and water withdrawal permits gives a different picture. The picture is not a winner and a loser. It is a set of uneven costs. Some costs are delayed. Some are displaced. Some are counted as progress. A data centre is not a village. A server hall does not grow food. An AI model does not drink water. The infrastructure around them does. That is where the investigation should begin. Who pays for the race? Not the algorithms. Not the investors. The answer is in the aquifer, the coal plant, and the converted field.
Local resistance is rarely visible in English-language coverage. It appears in Chinese social media posts about well water, in provincial petitions about land compensation, and in environmental impact assessments that are briefly posted then removed. Researchers who cross-reference these sources find patterns: water stress precedes data centre construction; coal capacity is reserved after renewable promises; farmland conversion outpaces job creation. The patterns do not fit a simple race narrative. They fit an infrastructure narrative. Infrastructure distributes benefits and harms. The harms are often quieter. They accumulate in aquifers, lungs, and soil. That is the price of 'China speed.'
