The Mirage of Gigawatts: What Ulanqab's 12.5GW Promise Really Tells Us About the AI Arms Race

There is a number that has been haunting my sleep since the Goldman Sachs report crossed my desk: 12.5 gigawatts. That is the amount of data center capacity that Ulanqab, a city in Inner Mongolia that most of the world has never heard of, has promised to build. To put that in perspective, it surpasses the entire Stargate project that OpenAI has been hyping for the last year. But here is the thing that nobody wants to talk about. The actual operating capacity today is just 1.2 gigawatts. A tenfold gap. And 70% of those commitments were made in the last twelve months. This is not infrastructure building. This is a land grab disguised as a technological revolution, and it tells us more about the psychology of the AI bubble than any model benchmark ever could. I have seen this movie before, back in 2017 when I watched fifteen friends lose their savings to ICO promises that were equally bold and equally empty. The names change, but the pattern remains disturbingly familiar. For those unfamiliar with the geography of Chinese computing, Ulanqab is not a random choice. It sits in a cold, windswept plateau roughly 300 kilometers from Beijing, connected by fiber optic links that deliver sub-5 millisecond latency to the capital. That is the killer feature. Most western data centers in remote locations are built for latency-insensitive workloads, backups, archival storage, the digital equivalent of a cold storage warehouse. But 5 milliseconds is fast enough to run real-time AI inference, search queries, and recommendation algorithms. It means Ulanqab is not just a backup site. It is designed to be a compute suburb of Beijing, a place where the heaviest AI workloads can live while still feeling virtually adjacent to the city's tech ecosystem. Add in the cold climate that naturally lowers PUE ratios, cheap land, and abundant wind and solar power, and you have what looks like a perfect storm of favorable conditions. DeepSeek has committed to a gigawatt. Xiaohongshu, the Instagram-like platform that is taking China by storm, has signed up for 600 megawatts. ByteDance and Alibaba are also in the mix. These are not small players. These are the heavyweights of Chinese AI, and they are all betting that Ulanqab is the future. But here is where my auditor's lens starts to itch. Let me walk you through the unit economics, because that is where the fantasy starts to crack. The business model for these massive data centers is essentially real estate plus electricity. You buy land cheap, build massive sheds, fill them with servers, and rent them out at wholesale rates to the big tech firms. The margin comes from your operational efficiency, primarily your electricity cost and your PUE. Ulanqab has excellent fundamentals on both counts. But the capital expenditure required to go from 1.2 gigawatts to 12.5 gigawatts is staggering. We are talking about billions of dollars in construction, cooling systems, power distribution, and networking infrastructure. The depreciation alone will eat profits for a decade. The investment payback period is 10 to 15 years, assuming everything goes perfectly. And here is the dirty secret that the Goldman report glosses over: most of those 12.5 gigawatts of commitments are not binding contracts. They are letters of intent, strategic reservations, and political signaling. They lock in land and power allocations, and they secure government subsidies. But they do not represent funded, shovel-ready projects with committed capital. In my experience auditing failed projects back in 2017, I learned that there is a massive difference between a promise and a purchase order. This is the same dynamic playing out at a much larger scale. The commitments are a hedge against future scarcity, a way for these tech giants to ensure they have options if the AI boom continues. But if the boom stalls, these commitments evaporate faster than a GPU allocation on a Monday morning. Let me now give you the contrarian angle, because I think most coverage of this story has the framing backwards. The conventional narrative is that Ulanqab is a sign of Chinese AI strength, a direct challenge to American dominance in compute infrastructure. But I would argue that this is actually a sign of fragility, not strength. Think about what it means when a city has to promise ten times its current capacity to attract customers. It means the demand is not there yet. It means these companies are being courted with promises of future abundance, not responding to current need. This is a buyers' market, and the data center operators know it. They are offering massive discounts and preferential terms to lock in anchor tenants. The power dynamics are all wrong. In a healthy market, the infrastructure providers have some leverage. Here, they are begging for business. And the customers know it. ByteDance and Alibaba are not just tenants; they are potential competitors. They have the engineering talent and the capital to build their own data centers if the economics shift. They are using Ulanqab's desperation to negotiate favorable terms, and they will walk away the moment the price is no longer right. This is not a partnership. This is a hostage situation where the hostage-takers are the ones with the checkbooks. There is also a deeper issue that nobody in the mainstream coverage is addressing, and it is the one that scares me the most. We are building a massive amount of compute infrastructure based on the assumption that AI demand will continue to grow exponentially. But what if it does not? What if the current wave of AI applications, chatbots, image generators, coding assistants, hits a plateau? What if the next generation of models is significantly more efficient, requiring less compute for the same output? The semiconductor industry has a long history of architectural improvements that reduce compute requirements. We are already seeing quantization techniques and model distillation that deliver comparable performance with a fraction of the compute. If that trend accelerates, we could be left with a massive oversupply of data centers, all burning electricity and all losing money. The 2022 crypto winter showed us what happens when speculative infrastructure meets reality. Billions of dollars in mining farms became worthless overnight when the price of Bitcoin crashed. The same thing can happen to AI data centers, except the scale is orders of magnitude larger. And unlike crypto mining, which can be turned off and on, data centers have long-term power purchase agreements and debt obligations that cannot be easily escaped. The sunk cost fallacy will keep these facilities running long after they become economically unviable. I want to share a personal story that I think illustrates the human dimension of this risk. In 2020, I co-founded a community called Ethos Circle, a Discord group dedicated to demystifying DeFi for non-technical professionals. We had 2,500 members, and when the October attacks hit, panic spread through the server like wildfire. I spent 72 hours straight moderating chats, translating exploit reports into simple safety checklists. We retained 85% of our members, not because we had superior technical analysis, but because we had built trust. That experience taught me something that applies directly to the Ulanqab situation. Trust is the only protocol that matters. And right now, the trust in these AI infrastructure commitments is based on nothing more than hype and hope. The people making these decisions are not bad actors. They are rational actors responding to insane incentives. The city government gets political credit for attracting big tech investment. The tech companies get optionality and negotiating leverage. The investment banks get underwriting fees. Everyone benefits from the fantasy, except the end user, who will ultimately pay for all this overbuilding through higher prices and stranded assets. Code is law, but people are the context. And the context here is that we are repeating the exact same mistakes we made in 2017, just with bigger numbers and more sophisticated marketing. The blockchain space spent years learning that community over coin is the only sustainable approach, that building for real users matters more than chasing speculative valuations. The AI infrastructure space is about to learn that same lesson, but the tuition bill will be measured in gigawatts. I have seen what happens when a community loses faith in the underlying technology. It takes years to rebuild. And when you have billions of dollars in physical infrastructure that relies on that faith, the collapse is not just digital. It is physical. It is concrete and steel and fiber optic cable that has to be written off. So what should we actually watch for in the coming months? I would ignore the press releases and focus on the operational data. If Ulanqab's actual operating capacity doubles from 1.2 to 2.5 gigawatts in the next year, that is real demand. That means the commitments are converting to construction. If it stays flat, if the announcements keep coming but the cranes do not appear, you know the whole thing is a paper castle. Also watch the capital expenditure reports from DeepSeek, ByteDance, and Alibaba. If they are actually spending money on Ulanqab infrastructure, that will show up in their earnings calls. And perhaps most importantly, watch the regulatory environment. If China starts tightening energy consumption standards for data centers, that will slow the buildout dramatically. The green energy advantage that Ulanqab claims is real, but it depends on the construction of massive new wind and solar farms, which face their own supply chain and grid integration challenges. The infrastructure buildout is not just about the data centers themselves. It is about the entire energy ecosystem that supports them, and that is a much harder problem than simply pouring concrete. Anonymity is a shield, not a lifestyle, and similarly, gigawatts are a promise, not a proof. The question we need to ask ourselves is not whether China can build 12.5 gigawatts of data center capacity. They probably can, given enough time and money. The question is whether they should, and whether the demand will be there to justify it. I have spent the last seven years watching communities form and dissolve, watching speculative bubbles inflate and pop, watching smart people make terrible decisions because they were all making the same terrible decision at the same time. This is that moment again. The AI arms race is real, but it is also a psychological phenomenon, a collective belief that compute is the ultimate moat. It might be. But it might also be that the moat turns out to be a swamp. I do not have the answer, but I know that when I look at the gap between 1.2 and 12.5, I do not see ambition. I see anxiety. And anxiety, when it is amplified across an entire industry, rarely ends well. The only way to navigate this is to stay grounded in the fundamentals, to keep asking whether the demand is real, whether the economics work, and whether we are building for the long term or just for the next quarterly report. Community over coin, always. And in this case, community over compute. The people who will survive this cycle are the ones who remember that.