The ledger doesn't lie, but it can be misleading. Ulanqab, a city in Inner Mongolia, has committed to 12.5 gigawatts of data center capacity—a figure that dwarfs OpenAI's Stargate project. The only problem? Actual operational capacity sits at 1.2GW. That's a 10x gap between promise and reality.
The public sees the spark—a headline-grabbing capacity commitment that positions China at the forefront of the global AI infrastructure race. I track the fuel lines: the engineering timelines, the capital requirements, the chip supply constraints, and the uncomfortable truth that 70% of these commitments were made in the last twelve months alone.
This isn't a story about data centers. It's a story about the widening chasm between strategic ambition and operational reality—a chasm that could swallow billions in capital if the AI demand curve doesn't bend exactly as projected.
The Context: A City Positioned at the Intersection of Policy and Ambition
Ulanqab sits roughly 300 kilometers from Beijing, connected by fiber optic links that deliver sub-5ms latency—a critical threshold that separates "cold storage backup" from "core compute capable." The city's climate, with average annual temperatures around 4°C, provides natural cooling that drives Power Usage Effectiveness (PUE) ratios down to an estimated 1.2-1.3, compared to the 1.5+ typical in warmer regions. Combined with some of China's lowest industrial electricity rates and abundant land, the physical fundamentals are genuinely compelling.
This is the "Eastern Data, Western Computing" (东数西算) strategy in action—China's national initiative to route compute-intensive workloads from coastal megacities to western provinces with surplus energy and cooler climates. Ulanqab is one of the designated hub nodes, and it's leveraging that designation aggressively.
The demand side reads like a who's who of Chinese AI and internet infrastructure: DeepSeek has committed to 1GW, Xiaohongshu (Little Red Book) has reserved 600MW, and both ByteDance and Alibaba have staked claims. These aren't speculative startups—they're the operators of China's most compute-hungry AI models and largest consumer platforms.
But here's where the forensic analysis gets interesting: the gap between what's been promised and what's actually running.
The Core: A Systematic Teardown of the 12.5GW Commitment
The Numbers Don't Add Up—Yet
Let me be precise about the arithmetic. The current operational capacity is 1.2GW. The committed capacity is 12.5GW. That's a 10.4GW delta that must be designed, financed, permitted, constructed, and commissioned. Based on my experience auditing infrastructure projects, let me walk through what that actually entails.
Power infrastructure: 10.4GW of additional capacity requires dedicated substations, transmission lines, and grid interconnection agreements. In China's current grid environment, where renewable energy integration is already straining transmission infrastructure, securing grid access for 10GW+ of new load is a multi-year endeavor. The typical timeline for a 500MW data center campus from site selection to operational status is 24-36 months under optimal conditions. Scaling to 10GW+ implies a 5-7 year horizon, assuming no regulatory or supply chain disruptions.
Chip supply constraints: This is the elephant in the room that most analyses gloss over. The U.S. export controls on advanced semiconductors—specifically NVIDIA's H100, H200, and the upcoming B-series GPUs—directly constrain what can be deployed in Chinese data centers. Domestic alternatives from Huawei (Ascend 910B) and others exist but offer different performance characteristics and software ecosystems. The 12.5GW commitment implicitly assumes access to cutting-edge silicon. If that access remains restricted, the capacity will be built but underutilized—a stranded asset scenario that would make the 2022 crypto mining exodus look trivial by comparison.
Cooling and mechanical systems: At 12.5GW scale, the cooling requirements are unprecedented. Ulanqab's cold climate helps, but AI workloads with 30-50kW per rack densities require liquid cooling solutions that are still maturing in the Chinese supply chain. The transition from air-cooled to liquid-cooled infrastructure at this scale represents a technology leap, not an incremental step.
Construction logistics: The sheer volume of materials—concrete, steel, copper, fiber—required for 10GW+ of data center space would strain regional supply chains. Ulanqab is not a major industrial hub; it's a city of approximately 1.7 million people with limited construction workforce capacity. Importing labor and materials at this scale introduces cost overruns and schedule slippage.
The Business Model: Real Estate Meets AI Hype
The underlying economics deserve scrutiny. Data centers are, at their core, a real estate play with a technology wrapper. The revenue model is straightforward: lease space and power to tenants, earn a margin over operating costs, and amortize capital expenditure over 10-15 years.
Ulanqab's cost advantage is real. Electricity at roughly ¥0.30-0.40/kWh (compared to ¥0.80-1.00/kWh in Beijing) and land costs that are a fraction of Tier-1 city prices create a genuine margin opportunity. At 1.2GW operational capacity, the economics work. The question is whether they work at 12.5GW.
Here's the problem: the demand side is concentrated in a handful of hyperscale tenants. DeepSeek, ByteDance, Alibaba, Xiaohongshu—these are sophisticated buyers with significant negotiating leverage. They know that Ulanqab is competing with Zhangjiakou, Qingyang, Zhongwei, and other hub nodes for the same workloads. The "low latency to Beijing" advantage is real, but it's not exclusive—Zhangjiakou offers similar proximity.
The pricing power that Ulanqab's operators might expect from their cost advantage is likely to be competed away. When you have 12.5GW of supply chasing a finite pool of AI workloads, the tenants set the price, not the landlords.
The Growth Quality Problem
Seventy percent of the committed capacity was announced in the past year. That's not organic growth driven by confirmed demand; that's speculative positioning driven by the AI narrative. Companies are reserving capacity they may never use, locking in land and power allocations at today's prices, and hedging against the possibility that AI compute demand explodes faster than expected.
This is rational behavior for the tenants—it's an option play. But it's dangerous behavior for the operators and the city, which are making firm commitments based on what may be soft reservations.
I've seen this pattern before. In 2017, ICO projects committed to token utilities they never delivered. In 2021, NFT projects promised decentralized storage they never implemented. The pattern is consistent: during hype cycles, commitments outpace reality by an order of magnitude, and the correction is brutal.
The Contrarian Angle: What the Bulls Get Right
I'm not a permabear. Let me steelman the case for Ulanqab's 12.5GW ambition.
First, the latency advantage is genuinely strategic. Sub-5ms connectivity to Beijing means Ulanqab can host latency-sensitive workloads—AI inference, search, recommendation systems—not just batch processing and training. This positions it as a "compute suburb" of Beijing, not a remote backup site. That's a fundamentally different value proposition than most western data center hubs.
Second, the customer quality is exceptional. DeepSeek is one of the most technically sophisticated AI labs in China. ByteDance operates the world's largest short-video platform with massive recommendation engine demands. Alibaba's cloud division is a top-tier global player. These aren't speculative startups; they're operators with proven demand and the balance sheets to back it up.
Third, the policy tailwind is real. "Eastern Data, Western Computing" is a national strategy with central government backing. This means preferential grid access, tax incentives, and streamlined permitting. When the state wants something built, it gets built—the question is only the timeline.
Fourth, the green energy angle is underappreciated. Inner Mongolia has abundant wind and solar resources. A data center powered by 100% renewable energy is not just an environmental statement; it's a competitive advantage for serving international customers with carbon neutrality requirements. This could be a genuine differentiator.
Fifth, the ecosystem effect is already emerging. When DeepSeek and ByteDance commit to a location, their supply chains follow. Server manufacturers, network equipment providers, and operations specialists will cluster around Ulanqab, creating a self-reinforcing ecosystem that makes it harder for competitors to dislodge.
These are real factors. They don't negate the risks, but they explain why rational actors are making these commitments.
The Takeaway: What to Watch, Not What to Believe
The ledger doesn't lie, but it also doesn't predict. The 12.5GW commitment is a statement of intent, not a statement of fact. The 1.2GW operational capacity is the only number that represents actual, revenue-generating infrastructure.
Here's what I'll be tracking over the next 12-24 months:
Operational capacity growth: If Ulanqab can double operational capacity to 2.5GW within 12 months, the commitments have substance. If it stagnates below 2GW, the gap between promise and reality is structural, not temporal.
Capital expenditure disclosures: DeepSeek, ByteDance, and Alibaba's earnings calls will reveal whether they're actually spending on Ulanqab capacity or just reserving options. Real CapEx is the only reliable signal.
Chip deployment: Whether Ulanqab data centers are deploying latest-generation GPUs—domestic or imported—will indicate whether the compute quality matches the capacity claims.
Competitive dynamics: Watch Zhangjiakou and other competing hubs. If they start undercutting on price, Ulanqab's margin advantage erodes.
The public sees a 12.5GW commitment and thinks China is winning the AI infrastructure race. I see a 10x gap between promise and operation and wonder who's going to eat the cost when the hype cycle normalizes.
The structure of this deal—massive commitments, concentrated demand, policy-driven supply—dictates the outcome. Either AI compute demand grows at an unprecedented rate, or Ulanqab becomes a monument to overcommitment.
The data will tell us which. It always does.