From the ashes of 2022, we planted seeds for 2030. But the seeds we’re seeing in Ulanqab, Inner Mongolia, are not the kind that grow forests—they’re the kind that might choke the soil with promises before a single root takes hold.
A city now claims it will host 12.5 gigawatts of data center capacity. That’s more than OpenAI’s entire Stargate ambition. It’s the stuff of legend—a digital fortress built to power the next generation of AI. But when I dig into the numbers, the legend cracks. The actual operational capacity today? 1.2 GW. That’s a gap of over 10x. And 70% of those commitments were made in the last twelve months—a frantic rush that smells less like organic demand and more like a land grab orchestrated by policy and hype.
Context: The Geography of Hype
Ulanqab sits in a cold, wind-swept corner of China. Its climate is a gift for data centers—low PUE, cheap electricity, ample land. A fiber optic cable runs to Beijing with under 5ms latency, making it an ideal candidate for latency-sensitive AI workloads. The government’s “East Data West Computing” strategy has blessed it as a hub. Companies like DeepSeek (1GW commitment), Xiaohongshu (600MW), ByteDance, and Alibaba have all signed on. It reads like a dream: a centralized AI compute paradise.
But I’ve seen this movie before. In 2017, during the ICO craze, we saw whitepapers promise decentralized compute networks that would revolutionize the world. Golem, Bitconnect—they had grand visions, but the actual utilization was near zero. The gap between promise and reality was the same: a chasm that only the most disciplined of builders could cross. Ulanqab is no different. It’s a whitepaper in concrete form.
Core: The On-Chain Audit of a Physical Project
Let’s break down the 12.5GW figure. To put it in perspective, 1GW of data center capacity can support roughly 100,000 high-end GPUs (like H100s) running at full tilt. 12.5GW would mean over a million GPUs. That’s a hardware supply chain that doesn’t exist yet—especially with US chip export restrictions. Each GPU needs cooling, networking, and power infrastructure. The engineering challenge is immense, but the financial one is even greater.
From my experience auditing DeFi protocols, I’ve learned to spot the difference between “TVL promised” and “TVL audited.” Here, the TVL is 12.5GW committed, but the “audited” operational capacity is 1.2GW. The gap is a red flag. It suggests that most of these commitments are options—not binding contracts. Companies lock in land and electricity rights to secure future optionality, but they aren’t paying for the compute yet. If the AI boom slows, or if capital tightens, these options expire worthless.
Moreover, the concentration of demand in a single location creates a single point of failure. If Ulanqab’s grid fails, or if a policy shift occurs, the entire AI ecosystem of these companies could stall. Compare this to the decentralized Web3 ethos: we build networks that are resilient because they are distributed. DePIN projects like Render or Akash aim to spread compute across many nodes, not consolidate it into one mega-fortress. Ulanqab is the antithesis of that—it’s a centralized monolith that could become a regulatory choke point.
Contrarian: The Pragmatism Test
But let’s not dismiss the entire plan. The core advantage—low latency to Beijing—is real. For AI inference, every millisecond matters. Distributed compute networks often struggle with latency because data has to travel long distances. Ulanqab, by being physically close to the demand center, offers a performance edge that no decentralized network can match today. That’s a legitimate utility.
However, the contrarian angle is that this very advantage creates a trap. Once the infrastructure is built, the operators will have enormous power. They can raise prices, dictate terms, or even be forced to comply with government surveillance. The clients—the DeepSeek, ByteDance, and Alibaba of the world—will be locked in. Their data, their models, their entire business will be in one physical location. That’s a national security risk, not just for the companies, but for the users who rely on them.
The CBDS and cryptocurrencies are fundamentally opposed: one seeks total surveillance, the other seeks privacy and freedom. Ulanqab’s centralized model, while efficient, is a step toward the surveillance state. It’s the infrastructure for a “panopticon AI.” We must ask: do we want our AI compute to be as opaque as the government’s firewall?
Takeaway: The Seeds of Tomorrow
From the ashes of 2022, we planted seeds for 2030. But the seeds we plant today must be resilient, not monolithic. Ulanqab’s 12.5GW promise is a bet on centralization. It might yield returns for a few years, but the harvest will come with a cost: dependency, fragility, and loss of autonomy. The decentralized alternative may be slower and more expensive today, but it builds a forest that can survive any storm. The question is not whether Ulanqab can build its castle. It’s whether we will have the wisdom to build a network of villages instead.