The 12.5GW Mirage: Ulanqab's Promised AI Empire and the Liquidity Illusion

Liquidity doesn't lie. But it does exaggerate. In the past twelve months, a city in Inner Mongolia with a population smaller than most Beijing districts has promised more compute capacity than the entire OpenAI Stargate project. The number is 12.5 gigawatts. The reality is 1.2 gigawatts. That gap isn't a typo. It's the market's collective fantasy about AI infrastructure, written in land grabs and power purchase agreements. I've spent the last decade watching capital flow into digital assets, and the pattern is always the same: the promise of future utility gets priced today, and the physical bottleneck is ignored until it's too late. Ulanqab's data center plan is not a story about compute. It's a story about liquidity velocity, institutional commitment, and the difference between a signed LOI and a fully-loaded GPU rack. Let me break down what's actually happening behind this headline, because the real story is not about 12.5 gigawatts. It's about the 11.3 gigawatts that don't exist yet. Here's the context. Ulanqab is located in the Ulanqab region of Inner Mongolia, a flat, cold plateau about 200 miles from Beijing. Its entire economic strategy for the past decade has been to sell its cold weather and cheap coal power to data centers. The 2024 AI boom, specifically the Chinese iteration of it, turned this into a gold rush. DeepSeek promised 1 gigawatt. Xiaohongshu promised 600 megawatts. ByteDance and Alibaba signed letters of intent. Goldman Sachs, the source of this data, calls it a 'game-changer.' This is not a small infrastructure project. To put 12.5GW in perspective, the entire global installed base of Bitcoin miners is estimated around 2-3GW of ASIC power. The largest single data center campus in the US, under construction by the Stargate consortium, is projected at around 2-3GW. Ulanqab alone is proposing a capacity equivalent to four Stargate projects, all crammed into a region that currently operates just over 1.2GW. This is the macro context: we are witnessing the first major test of whether the AI compute narrative can turn into physical reality. It's a test of the entire infrastructure supply chain, from transformers to liquid cooling to grid interconnects. The core analysis here is about liquidity and latency. My background in analyzing token flows and macro M2 tells me that when you see a 10x jump between operational and committed capacity, you are looking at a supply shock that is going to be priced in three years ahead of schedule. In crypto, this is exactly what happens when a DEX lists a token with a 10% circulating supply. The price reflects the promised total supply, but the liquidity that exists is only for the tiny operational fraction. That's a classic "liquidity vacuum" signal. Technically, Ulanqab has advantages. The PUE (Power Usage Effectiveness) is low, around 1.2 or 1.3, because of the cold climate. The latency to Beijing is under 5ms, which is a killer feature for AI inference. That means it's not just a backup data center; it's a legitimate extension of Beijing's compute grid. The cost of electricity is significantly lower than the national average, and land is cheap. These are the physical foundations. But the technical architecture for 12.5GW requires a level of engineering coordination that is unprecedented in this region. You need substations at the 500kV level, direct power lines, and liquid cooling systems for the density of GPU racks that AI requires. The current infrastructure supports 1.2GW. The gap is not just in data center space; it's in power generation and distribution. My audit of crypto mining sites taught me that a power agreement is not a power supply. And a signed PPA is not a plug. Now, let's get to the liquidity-first skepticism. We are in a bull market for AI infrastructure, just like we were in a bull market for DeFi in 2020. In 2020, the "liquidity" was yield farming, and the TVL (total value locked) went up 4000% in six months. But a huge percentage of that was wash trading and double counting, and it collapsed when the underlying yield (which was just token inflation) evaporated. This Ulanqab case is analogous. The 70% of commitments made in the last year are the "yield farming" of the compute world. They are announcements made to signal strategic alignment with the AI narrative, often in exchange for local government subsidies or land options. They are not capital expenditures. They are marketing expenses. Skepticism isn't a resistance to progress; it's a filter for the difference between a promise and a payment. Let's follow the money. DeepSeek's 1GW commitment, for example, is real. They are running inference for a popular model, and they need the power. ByteDance's 3GW might be real, given they are running massive recommendation engines. But the key is the velocity of capital. Are these companies actually allocating the billions of dollars in CAPEX needed to fill this space? In the crypto world, we saw that a commitment from a VC doesn't mean anything until the treasury votes to release the funds. Same here. Let me talk about the unit economics of the Ulanqab. The average data center contract is long-term, 5 to 10 years, with a fixed price. If I'm a wholesale data center operator in Ulanqab, I'm signing a contract to deliver power and cooling. I need to build the shell, install the transformers, and run the pipes. The CAPEX for 1GW of data center capacity is roughly $3 billion to $4 billion, excluding the GPU clusters. That's the capital cost. So, to go from 1.2 to 12.5GW, you need to spend roughly $30 billion in construction costs alone, before you even buy the GPUs. The GPUs are the killer. The latest H200s are $40,000 each. To fill a 1GW facility with H200s, you need over 100,000 GPUs, which is $4 billion just for the chips. So we are looking at a $10 billion price tag per GW if you include the GPU and the land. For 12.5GW, that's a $125 billion commitment. That is a massive liquidity requirement. This is not a regional project; this is a global liquidity event. Now, the crucial contrarian angle. Everyone is talking about the "decoupling" of China's AI compute from the US. But I think the bigger story is not decoupling; it's the supply chain bottleneck. The US export controls restrict the sale of H100/H200 to China. That means Ulanqab cannot deploy the world's most advanced GPUs. They have to rely on domestic chips, like Huawei's Ascend 910B or 920. These chips are available, but the yield and the performance are still behind. This creates a specific dynamic: Ulanqab's 12.5GW capacity is not 12.5GW of Nvidia compute. It's 12.5GW of domestic compute, which is maybe 70% of the performance of Nvidia. So the "promised" capacity is actually a "promised" capacity at a discount. When you model the liquidity of this AI, you realize that the unit economics are not as rosy as the headlines suggest. Furthermore, I have to challenge the "Macro Watcher" crowd. The general consensus is that 12.5GW means China is winning the AI race. But it's a race to the bottom if it's built on unfulfilled demand. Look at the current supply of actual operational capacity. 1.2GW is operational. The top AI players in China - DeepSeek, ByteDance, Alibaba, Xiaohongshu - they need maybe 2.5GW of incremental compute in the next year, if they all double their usage. They don't need 12.5GW. They need maybe 2-3. So the 12.5GW is a promise built on a speculative demand curve, not a real one. The demand curve is currently a flat line, and the supply curve is a rocket ship. That's a classic macro liquidity mismatch. I'll give you a concrete historical analogy. In 2017, I audited ICO whitepapers. 80% of them claimed they would have a network live in 6 months. 99% of them had no liquidity model. The token prices were based on the total supply, not the circulating supply. The team held 30% of the supply, but the token price was based on 100% of the supply. So when the team sold tokens, the price collapsed. It was a liquidity vacuum. Ulanqab is the same: the "committed capacity" (12.5GW) is the total supply, and the "operational capacity" (1.2GW) is the circulating supply. The market is pricing the total supply, but the liquidity (actual power and cooling) is only for the circulating supply. When the operational capacity doesn't catch up to the committed capacity, the "price" of the compute (i.e., the rental rate) will collapse. It's a supply overshoot. The other blind spot is the "energy" side. The power grid in Inner Mongolia is a coal and wind mix. To get 12.5GW of data center power, you need to have a dedicated power plant or a massive connection to the national grid. The local grid can't handle a 10x jump in load. You need to build new transmission lines. That takes 4-5 years. And the electricity is not free; the cost of electricity is a core input. As the AI supply increases, the marginal cost of power will also increase. In a power market, when demand jumps 10x, the spot price doesn't stay flat. It spikes. So the original "low power cost" advantage will erode as the supply grows. This is the "scarcity premium" that is often ignored. So what does this mean for the institutional convergence? I see this as a classic case of "institutional capture" without "institutional commitment." The local government and the banks are offering subsidies. The AI companies are signing MOUs to get the subsidies. But when the subsidies expire, the actual power purchase agreements will be renegotiated. I see this as a "duplicate" of the Ethereum staking bubble, where the promised yield is far higher than the actual issuance of the network. The staking yield was based on a high number of validators, but the transaction volume didn't justify it. The yield dropped. The same thing will happen here. The rental yield for AI compute in China is going to drop significantly as supply outpaces demand in the next 12-18 months. Now, let me talk about the takeaway. I'm not saying that Ulanqab is a complete bubble. I'm saying it's a macro experiment. It's a bet on the velocity of AI adoption. If the demand for AI inference grows 10x in the next 3 years, then the 12.5GW is just enough. But if it grows 3x, then the capacity is an oversupply. The key indicator is not the promised capacity but the "operational capacity" growth rate. I will track the monthly increase in the actual MW consumed. If it's rising at 15% a month, the future is real. If it's rising at 5% a month, the promise is a lie. Let's do a quick scenario planning. In the bull case, the Chinese government makes a strategic effort, they use domestic GPUs, they build the grid, they get the power, and they fill it with inference for the national "AI+" initiative. The PUE is 1.2, and they use it as a "green data center" which is a world benchmark. That's the bull case. In the bear case, the demand for domestic AI models, which are still not as good as GPT-5 or Claude 4, fails to attract enough paying customers. The supply is there, but the demand is not. The data centers sit half empty, and the rental prices crash. The debt for the CAPEX is high. The project is a zombie infrastructure. There is also the "AI-agent scenario" I've been thinking about. In 2026, we have AI agents that will be trading, chatting, and booking trips. These agents need compute. They don't need "traditional" web data centers; they need micro-transactional, high-frequency, low-latency compute. Ulanqab's 5ms latency to Beijing is perfect for agent-based compute. If the AI-agent economy takes off, the demand curve might be exponential, not linear. This is the only way that 12.5GW is justified. But this is a speculative technology. It's not a proven market. So I'm using the institutional convergence model to say: the current price is a discount for a speculative future. Skepticism isn't about ignoring the potential. It's about measuring the distance between the potential and the price. The price of land and power in Ulanqab is cheap, but the price of capital and the price of supply chain is not. The 12.5GW is not a capacity; it's a liability. It's a promise to the grid, to the government, to the bank. The government wants the taxes. The bank wants the interest. The grid wants the load. But who is the end user? If the end user is a speculative AI startup with no revenue, the whole thing collapses. If the end user is a profitable company like ByteDance or Alibaba, it's a real foundation. So, here is my macro takeaway. The Ulanqab story is the ultimate test of the "liquidity illusion." The liquidity of the capacity is not real until it's filled with electricity and GPUs. The 12.5GW is a "phantom" asset. The only way to profit from this is to be a supplier of the physical assets (the cables, the cooling, the land), not the owner of the capacity. I recommend that my readers ignore the headline number and focus on the actual "power purchase agreement" signed with actual load. When a real company signs a PPA with a guarantee of a 1MW load, that's a signal. When they sign an MOU, that's just a marketing press release. And here's the final contrarian twist. The fact that this is happening in China, not in the US, might be a better bet. The US is building in Texas, with a 4ms latency to the grid, but the labor costs are higher and the regulatory environment is less flexible. In China, the state can move faster. The government can order the grid to build the transmission lines in 24 months. So the 12.5GW could actually happen, not because of market demand, but because of state capacity. But this is exactly the double-edged sword. When the state builds something, it doesn't care about the return on investment. It cares about the strategic position. So the capacity will be built, but the profitability will be poor. The Chinese banks will get stuck with the bad debt. But the AI industry gets the capacity. So the question isn't "can they build 12.5GW?" The question is "can they build 12.5GW without destroying the local economy?" The answer is probably yes, they will build it, but the real estate and the equipment will be stranded assets in the next 3 years. I'll leave you with this. The global AI infrastructure is a race. But it's a race to build a bridge to the future. The issue is that the bridge is being built with the assumption that the future is a certain. I've seen this in crypto. The bridge to the internet was built in 1999, and the internet did happen. But the companies that built the bridge in 1999, the fiber optics, they went bankrupt. They were right, but they were too early. Ulanqab is a 1999 fiber optic company. They are building the future, but they will not see the profit. The profit will go to the next generation of builders who are not stuck with the sunk cost. Keep watching the operational MW. That's the only honest metric. The rest is noise.