The short answer

Ulanqab, a city of 1.7 million in Inner Mongolia, consumed 68 TWh of power in 2023 — mostly for AI data centers, including DeepSeek's planned 1 GW site. The advantages are real: cold air (PUE under 1.2), cheap wind power, and China's Eastern Data Western Compute policy. But viral claims (1,000× Colossus, 5M built racks) are inflated — this corrected edition checks every figure against primary data.

On a cold night across the Ulanqab steppe, wind cuts through rows of white turbines, producing a sound like a lone flute echoing across the vast Inner Mongolian landscape. This place was once known for a story of smallness: He Pingping, a man just 74 cm tall, born in Huade County within Ulanqab, once held the title of the shortest man on Earth. He traveled the world, making people look at that tiny body and wonder: how could a human survive in such a harsh place?

Now that same land is doing China's biggest thing: feeding artificial intelligence.

1. The Tiny and the Colossal

In 2023, this city of 1.7 million people consumed 68 Terawatt-hours of electricity. Nearly 0.7 thousandths of all the power used by a country of 1.4 billion people. That figure roughly tracks the 1% mark when you include the first half of 2024, when demand for computing power exploded. To put it in perspective: every year, a city the size of Nghe An province uses as much electricity as dozens of Vietnamese provinces combined.

The question starts here: where does the power go, when residential consumers in Ulanqab account for just 1.9% of total consumption? It goes into the machine mountains. Into windowless buildings that run day and night without sleep. Into AI.

The Chinese call Ulanqab "the nation's token capital" — where every character you type into DeepSeek, every question you ask an AI, might be computed on this 4.3°C cold steppe.

Why a steppe, and not Shanghai or Beijing? Think of Sapa in winter. Servers fear heat as much as people do. In Beijing, you have to run high-capacity air conditioning year-round to cool a rack row. In Ulanqab, you just open the wind. An average temperature of 4.3°C helps push the PUE ratio below 1.2 — among the best in the world. Then there is the abundance of on-site wind and solar power, especially after China aggressively pushed the "Eastern Data Western Compute" policy starting February 2022. That policy selected 8 hub nodes and 10 computing clusters, pushing heavy workloads westward — where land is cheap, the climate is cold, and population density is low. Ulanqab hit every factor. From here to Beijing is just 4.2 milliseconds of transmission — in the blink of an eye, data reaches the capital.

And so the land rush began. Over 5 million standard racks were contracted, with total committed capital exceeding 5,000 billion CNY. Apple, Alibaba, Huawei, Kuaishou all came. DeepSeek finalized a 1 Gigawatt project — enough electricity for a city the size of Vinh. In the first half of this year, power consumption for computing here surged 89.5%. Some claim Ulanqab is 1,000 times larger than Elon Musk's Colossus supercomputer in the US. That is an exaggeration. Colossus has roughly 200,000 GPUs, about 5,000 rack cabinets. Ulanqab is around 25 to 50 times larger by equivalent output capacity — already colossal, no need to inflate.

But there is a number you cannot inflate: actual disbursed capital stands at only about 8%. The rest is still paperwork, land with stakes, power lines drawn but machines not yet arrived. Currently operational capacity is 172,000 P, up 112% — fast, but the road ahead is long.

2. Behind the Buzzy Numbers

Amid the excitement, some figures have been inflated beyond recognition. A viral social media video once claimed each Ulanqab household consumed an average of 105,000 kWh per year — roughly 10 times the US average. That number sounds impressive, but it is a statistical deception.

Residential electricity in Ulanqab accounts for only about 1.9% of total consumption. The remaining 92% is industrial and computing load. Dividing total industrial load by the number of households to compare with the US is a methodological error: you cannot take all the power of an industrial city, divide it by household count, and compare it to a country where electricity is primarily residential. In reality, each Ulanqab household uses only about 1,635 kWh per year — lower than the US average of 10,791 kWh per household per year, by a factor of roughly 6.6. The inflated portion is factories and data centers, not families.

Then there is the 5 million rack figure. Many reports write "Ulanqab builds 5 million racks" as if they already exist. In reality, this is the contracted number based on standard rack units, each at just 2.5 kW, mainly used for storage and cloud. High-density AI racks can reach 40–135 kW per unit, converting to many more standard racks. The actual utilization rate for racks already built is about 66% — meaning 330,000 out of 500,000. Contracted and operational are two different stories.

Similarly, comparing Ulanqab to xAI's Colossus requires care. Colossus has about 200,000 GPUs, estimated at 250–300 MW of power. Ulanqab has a total planned IT capacity of approximately 12.5 GW — about 25 to 50 times larger than Colossus in actual capacity, not 1,000 times as some articles mistakenly wrote.

US analysts also do not agree on the level of "threat." In its 2028 scenario, Anthropic noted that Chinese labs have world-class talent, but the main bottleneck remains compute and chips. The concern has strategic basis, but that does not mean China has surpassed cost-per-dollar AI efficiency if the chip embargo remains unresolved.

The broader implication for the global AI infrastructure race is scale versus utilization. The US leads in actual deployed GPU compute and software ecosystem maturity. China leads in raw construction speed and centralized resource allocation. Ulanqab demonstrates both the ambition and the Achilles heel: you can build the world's largest data center shell in 18 months, but filling it with functional silicon and running workloads profitably is a different engineering problem entirely.

3. The 1 Gigawatt Gamble and the Future Among Machine Mountains

DeepSeek did not arrive in Ulanqab by chance. The company is transforming from a language model startup into a computing infrastructure giant. In 2026, it raised approximately $7.4 billion at a valuation of $52–59 billion, with plans to IPO on the Shanghai STAR Market by the end of 2026 and a formal listing in Q2 2027. The pricing strategy behind that ambition is what we analyzed in DeepSeek's permanent 75% price cut.

The heart of that gamble is the 1 Gigawatt project in Ulanqab. A number equal to the entire electricity consumption of a city the size of Vinh. If completed, this will be one of the largest data centers in the world by power consumption. Reports suggest the first phase could come online in late 2027 or early 2028. But the chip question remains open.

DeepSeek plans to use Nvidia Blackwell chips — a product line currently banned from sale to China under US orders. Nvidia claims it has seen no evidence; the US alleges violations. No independent audit has confirmed either way. The market depends on smuggled or domestic chips like Huawei's Ascend. This is the biggest double-edged sword.

Three risks need close monitoring. First, chips and embargoed technology. Second, the power grid and cooling infrastructure at Gigawatt scale. Third, high-level talent: the local government is pushing vocational training, but top engineers are still reluctant to move to a remote steppe.

So is Ulanqab a real miracle or a paper bubble? It is currently the largest testbed of the Eastern Data Western Compute strategy, with contracted scale exceeding some US clusters in power and rack count by standard units. But a serious assessment must distinguish "contracted" from "operational," "standard rack" from "AI rack," and avoid household-dividing tricks that assign industrial load to families. After verification, the core thesis stands: Ulanqab is one of the most worth watching places in the world for AI infrastructure — if you get the numbers right.

For Vietnamese SMBs watching this story, the signal is not about megawatt race. It is about what happens when computing becomes abundant and cheap: the moat shifts from owning hardware to orchestrating it — a comparison we unpacked in DeepSeek vs GPT-5 vs Claude. The lesson from Ulanqab's paper-rack gap is that infrastructure without utilization is just real estate. Utilization without talent is just electricity bill. And talent without software is just a very expensive science project.

Correction Summary

  • Ulanqab population: 1,706,328 (2020 census), correct.
  • 2023 societal electricity: 68.11 TWh, correct.
  • Share of national electricity: ~0.69–0.74%, near 1%, approximately correct.
  • Per household 105,000 kWh: WRONG in essence, actual ~1,635 kWh/household.
  • Total contracted investment: >5,000 billion CNY, correct but disbursed only ~8%.
  • 5 million racks: contracted number, not necessarily built; needs power-unit conversion.
  • Colossus vs Ulanqab: 25–50 times larger, not 1,000 times.
  • DeepSeek 1 GW: report-backed, partial operation late 2027/early 2028.
  • Blackwell chips banned: unconfirmed.
  • IPO late 2026/Q2 2027: analyst reports note this.

Sources

  1. Wikipedia — Ulanqab
  2. Science and Technology Daily, 2026-08-20
  3. xAI Colossus official
  4. Guangming Daily, 2026-01-26
  5. ChinaTalk — Eastern Data Western Compute is Fake
  6. ChinaTalk — How much AI does 1 dollar get you in China
  7. Data Center Dynamics — DeepSeek planning 1GW data center in Inner Mongolia
  8. Anthropic — 2028 AI leadership
  9. NEA China — 2024 power consumption statistics
  10. EIA — US household electricity consumption