Summary of Key Points
Amid the wave of AI, technology companies that traditionally relied on "light assets" (such as software and algorithms) are shifting towards "heavy assets"—building their own data centers, securing land, and acquiring electricity. The real estate industry has also identified data centers as a new opportunity, but the capital market is concerned that some projects may merely represent the traditional business of renting server cabinets, which could lead to bubbles. What truly holds value are intelligent computing centers capable of providing AI processing power.
1. Why Are Technology Companies Suddenly Becoming More "Heavy"? — From Renting Processing Power to Building Their Own Infrastructure?
In the past, technology companies that worked on large AI models mostly rented processing power from other data centers. Things have changed:
- Surging Demand: AI training requires massive amounts of computing power; for example, training a model with hundreds of billions of parameters might take thousands of GPUs running for several months. External processing power is either unavailable or too costly.
- Controlling Core Resources: By building their own data centers, companies can directly manage critical aspects such as electricity supply, cooling (e.g., using liquid cooling technology), and networking, without relying on others.
For instance, DeepSeek is hiring civil and electrical engineers to build a data center from planning to operation. ByteDance established a subsidiary in Zhongwei, Ningxia, due to the low land costs and abundant electricity supply there, which are ideal for large-scale computing facilities.
In short, it’s like going from "borrowing a kitchen to cook" to "building your own kitchen and using it as you wish."
2. Data Centers Become a New Track in Real Estate — Real Estate Companies Shift from Building Houses to Building "Computing Facilities"
The profits from traditional residential development are declining, so real estate companies see data centers as a profitable opportunity:
- Government Support: Regions like Ulanqab and Yancheng have favorable conditions, such as cold climates (natural cooling that saves on electricity) or access to wind power (cheap electricity), and the government offers incentives to attract companies to build data centers.
- Business Transformation for Real Estate Companies: Instead of building residential homes, they are now building data centers and operating industrial parks for technology companies. A well-known real estate firm is actively pursuing data center construction projects.
For example, Ulanqab has signed 84 data center agreements with a total investment of over 500 billion yuan, aiming to reach 200,000 P of computing power by 2026 (enough to train dozens of large models). Yancheng, Jiangsu, is attracting companies like SenseTime to build a 3,000-P intelligent computing center thanks to its wind power resources.
This represents a shift for real estate companies from selling houses to selling space that can host computing services.
3. The Strategic Layout of Data Centers — Where Is Training Suitable? Where Is Inference Suitable?
AI computing power requirements differ significantly, leading to different locations for data centers:
- Training Power: For training large models like ChatGPT, which don’t require real-time responses but need massive amounts of processing power, remote areas with low land costs and electricity prices (e.g., Ulanqab, Zhongwei, Ningxia) are ideal.
- Inference Power: Applications that require immediate responses (e.g., AI-powered chat or image generation) should be located near users (e.g., Beijing, Shanghai) to minimize network latency.
Therefore, technology companies will set up training bases in remote areas and inference nodes in major cities, achieving cost-effectiveness.
4. Capital Market Concerns: Is the Boom Real or a Bubble?
Although data centers seem popular, not all projects are profitable:
- Traditional IDCs Focus on Renting Cabinets: Many data centers simply rent out server cabinets, similar to renting out housing, relying on long-term contracts with major clients who often have significant bargaining power.
- Intelligent Computing Centers Are the Future: Unlike traditional IDCs, intelligent computing centers provide comprehensive services like GPUs, liquid cooling, and high-speed networking, focusing on selling processing power rather than just space. These projects are more competitive but require higher technical expertise.
- Bubble Risk: The data center industry experienced a boom from 2021 to 2023, but some projects failed due to a lack of stable clients and low utilization rates, resulting in long investment recovery periods.
The capital market is wary that following the trend without proper infrastructure may lead to empty data centers. Only projects that can genuinely meet AI computing needs will be successful.
5. The True Competition: Forcing Control over Core Resources in the AI Ecosystem
Technology companies’ pursuit of land, electricity, and data centers is not about entering the real estate market; it’s about gaining control over the critical resources needed for AI:
- Electricity: AI consumes a lot of power, and access to cheap and stable electricity is crucial.
- GPUs: While GPUs are essential, they are ineffective without proper support systems.
- Land: Land suitable for data centers (e.g., with low temperatures and proximity to major cities) is becoming increasingly scarce.
In essence, those who control these resources will have a competitive advantage in the AI ecosystem.
In Conclusion: AI has forced technology companies to shift from a light-to-heavy approach, while real estate has found a new growth area. However, we must be cautious of potential bubbles. Only projects that can provide genuine AI computing power will be successful in the long run. This is similar to the internet era’s competition for servers; in the AI era, the focus is on acquiring "super server centers" along with the underlying land and electricity resources.