Hello! I'm your financial analysis assistant. This article from the IT Times is packed with valuable information. It's not just about the activities of an AI company called DeepSeek; it also reveals a core principle of the second half of China's AI industry: the competition is shifting from a technical race to a race of energy and cost.
To help you understand this easily, I'll first summarize the main points in one sentence and then break it down in five simple sections.
📝 Core Summary
DeepSeek is turning the production of tokens (AI-generated text/data) into a battle of extreme cost efficiency. They plan to build a massive data center in Ulanqab, Inner Mongolia, equipped with a large number of Huawei chips, and are preparing to go public to raise funds. Ulanqab has been chosen as the battleground for this competition because it boasts the lowest latency network closest to Beijing, extremely low electricity prices (especially with direct supply of green energy), and abundant renewable resources. The future of AI competition will no longer depend on who has the smartest models, but on who can produce the cheapest tokens using the lowest electricity costs and the highest utilization rates.
---
🔍 In-Depth Analysis: Five Key Points to Understand the “Token Cost War”
1. Why Ulanqab? More Than Just Cold Weather
Many people think data centers are located in the north because it's cold, which saves on air conditioning costs. That's partly true, but there's another hidden advantage: its proximity to Beijing.
- Distance equals money (latency): Beijing is the hub of AI applications in China, with many tech companies and users. It only takes 1.5 hours by high-speed train to travel from Beijing to Ulanqab, and the fiber-optic transmission latency is only 2.1 milliseconds. This means that using AI in Beijing and processing data in Ulanqab with almost no delay. If the center were built in Guizhou or Gansu, although the electricity would be cheaper, the increased network latency would reduce user experience.
- The unique Mengxi Power Grid: Inner Mongolia's western region has its own independent power grid system, which is more flexible in power trading compared to the national grid. Data centers can directly sign long-term power purchase agreements with wind and solar power plants, locking in electricity costs for several years in advance.
- Direct green energy connection: Instead of the traditional power transmission process, Ulanqab connects wind farms directly to the data centers, eliminating intermediate steps. This not only makes the energy usage more sustainable but also cheaper and more stable.
2. DeepSeek's Strategy: Spending Big to Drive Down Prices
The article mentions two seemingly contradictory actions:
1. Massive investment: They plan to build a 1GW data center, purchasing 160,000 Huawei Ascend chips, which could cost over 40 billion yuan just in chips.
2. Drastic price cuts: After the release of the new V4.1 Flash model, the price of tokens dropped significantly, with idle periods costing as little as 0.02 yuan per million tokens.
What's behind this strategy?
- Economies of scale: The AI industry has high fixed costs (building data centers, purchasing chips) but low marginal costs (electricity per additional token generated). Once the scale reaches 1GW, the cost per unit of electricity can be significantly reduced.
- Attracting users and generating more data: By offering low-token prices, DeepSeek attracts developers and businesses to use their models. The more users, the more data generated, and the faster model updates. The large amount of computing power also helps spread the high hardware costs.
- Raising funds for expansion: They are working with CITIC Securities to list on the STAR Market, indicating a need for capital to sustain this high-investment, low-profit expansion model. Their message to investors is: “We have the lowest electricity costs and the largest market potential; give us money, and we can outcompete our rivals.”
3. Ulanqab's Billion-Dollar Bet: From Selling Land to Selling Electricity
In August 2026, the city signed contracts worth over 100 billion yuan, with annual production capacity soaring from 3.3GW to 12.5GW. This indicates a major industrial shift.
- Industry chain expansion: Data center construction used to focus on building buildings and purchasing servers, but now, with high power consumption, electricity has become a key factor. Companies in wind power, energy storage, and cloud services are all moving in.
- From IDC to Energy Hub: Ulanqab is transforming from a place for storing servers to a green energy processing center. It offers cheap and stable green energy and a network close to Beijing, enabling the conversion of electricity into computing power and services for Beijing.
- Goldman Sachs' Data: Ulanqab has become the fastest-growing AI computing hub in the Asia-Pacific region, indicating a global or national shift in AI computing demand.
4. The Biggest Challenge: Low Electricity Prices, but Low Utilization
A sobering point in the article is that the average utilization rate of GPUs in domestic AI centers is less than 30%.
- What is utilization? If you buy 100 supercomputers and only 30 are in use, the other 70 are idle, resulting in high costs.
- Ulanqab's dilemma: Although the electricity is cheap (around 0.358 yuan per kilowatt-hour), if there are no enough AI applications or users to utilize the computing power, the chips will remain idle, wasting money.
- Realistic Constraints: Land availability is limited, and some projects are still in the planning stage. Local officials admit that the region is still in a loss-making phase, with untapped economic value and tax revenue.
5. The Ultimate Question: Who Will Win the “Token Factory” Race?
The article raises a crucial question: How much does it cost to produce one million tokens?
- The shift in the AI industry: In the first half, the focus was on who had the smarter models; in the second half, it will be on who can produce tokens at the lowest cost.
- Cost formula: Token cost = (electricity cost + depreciation + maintenance) / number of tokens produced. Ulanqab has an advantage in electricity prices. Although Huawei chips are expensive, the large scale can reduce costs. Effective utilization is the key variable. If DeepSeek's models are effective and attract many users, the cost can be significantly lowered.
- Conclusion: By choosing Ulanqab, DeepSeek is betting on solving the utilization issue and converting cheap electricity into the cheapest tokens. If successful, they will create a significant competitive advantage.
💡 Insights for Everyone
1. AI services will become cheaper: As giants like DeepSeek establish large, low-cost computing centers in places like Ulanqab, the cost of using AI services (writing, programming, customer service, etc.) will decrease, potentially becoming free or very low.
2. Energy is the new commodity: Regions with stable, cheap, and green energy resources (like Inner Mongolia and the northwest) are gaining strategic importance.
3. Focus on utilization: To evaluate the reliability of AI companies or data centers, look at their computing power utilization and actual token revenue, not just the size of their facilities. Only by effectively using computing power can they make a profit.
In one sentence: DeepSeek's layout in Ulanqab represents a comprehensive cost challenge involving electricity, network, chips, and models. Whoever can convert the most electricity into the cheapest tokens will win the second half of the AI industry.