The "Hidden Barriers" of the AI Computing Power Race: When Chips Are No Longer the Only Factor
Hello everyone, I'm your financial journalist. The news we're going to discuss today might be different from what many people think of as "AI = buying graphics cards."
In the past, we always thought that competing in AI meant seeing who had the most GPUs (graphics processing units) or the fastest chips. However, Chen Zhenkuan, an executive at Lenovo Group, recently revealed a very practical issue: AI servers are now too heavy to be placed on ordinary factory floors; they also consume so much electricity that regular factories can't supply it.
This is not just an engineering problem; it reveals a fundamental shift in the AI industry: AI infrastructure is evolving from a "light-asset technology race" into a "heavy-asset infrastructure race." It's like how we used to compare whose phone screens were brighter; now it's about whose power plants are larger and whose foundations are more stable.
Below, I'll break down this news into five aspects to explain the logic behind it in simple terms.
---
1. **Physical Limits: Why Must AI Servers Be Located on the Ground Floor?**
In the past, server rooms could be set up on any floor, as a single server didn't weigh much and the floor could support up to a ton. But with the emergence of "super nodes" (high-performance computing units consisting of dozens or even hundreds of GPUs), things have changed:
- Significant Weight Increase: Super nodes are extremely heavy. Chen Zhenkuan mentioned that new factory floors need to support at least 3 tons per square meter, sometimes even 5 tons. This means the equipment must be placed on the ground floor, and underground garages cannot be dug (the foundation must be extremely deep and stable).
- Increased Power Demand: A super node production facility requires at least 30 megawatts of electricity, or even more. For comparison, ordinary factories might only need a few megawatts, but these facilities need specialized 110-kilovolt power transformation equipment.
In simple terms: It's like running a small noodle shop from your home kitchen; now you need a central kitchen with industrial-grade stoves, dedicated high-voltage power lines, and a reinforced foundation to prevent the stoves from collapsing the floor. Physical space, load-bearing capacity, and electricity—these previously overlooked infrastructure details—have become the first barriers to entering the AI computing power market. Many old factories are now unusable and must be rebuilt or renovated, which is a huge additional cost.
---
2. **The Blending of Technical Boundaries: It's No Longer Just About Buying Chips**
In the past, servers, storage, networks, and operating systems were sold separately with clear boundaries. You bought a CPU, memory, and a network card, and then assembled them together. But super nodes have broken these boundaries. They are no longer individual servers; they are a single, integrated unit that includes:
- Computing Cabinets: For housing the GPUs.
- Power Cabinets: For supplying electricity.
- Storage Cabinets: For storing data.
- Connection Cabinets: For high-speed data transfer.
This leads to a complex issue: Lack of Standards: Chen Zhenkuan mentioned that there are 12 manufacturers and 13 different protocols just for high-speed interconnection within a single node.
In simple terms: Imagine assembling a computer; in the past, any USB port would have worked. Now, each manufacturer's connectors are different in shape and speed. Manufacturers like Lenovo and Huawei not only need to consider which chips to use but also reserve space for potential upgrades or changes in chip manufacturers, which means they have to redesign the power supply, cooling systems, and connection cables.
The result: The impact extends beyond GPUs to switches, PCB boards, high-speed cables, power modules, and liquid cooling equipment. The technical complexity has increased exponentially, along with research and development costs.
---
3. **The New Game of Finance: From "Making as Much as You Spend" to "Borrowing for Infrastructure**
This is the most shocking part for most people. AI infrastructure is so expensive that companies can't afford it with their own cash flow.
- Exponential Scale: Data from CITIC Construction Investment shows that during the past two decades of the mobile internet era, the annual capital expenditure in China and the US was about $100 billion each. This year, North American cloud companies spent about $800 billion, and China spent about 1 trillion yuan. The investment has increased by 10 times, possibly 20 times when considering the entire supply chain.
- Changes in Financing: Companies used to invest after making a profit, but now they have to borrow money. Financing sources have expanded from simple profit retention to bonds, financial leasing, private equity, and credit arrangements.
- Interest Rates as a Lifeline: Wang Qing from Chongyang Investment pointed out that projects backed by large cloud companies like Alibaba, Tencent, and Microsoft have interest rates of around 5.3%-6.6%, while those without such support have rates as high as 8.5%-9%.
In simple terms: It's like two real estate developers building a building. Developer A (like Vanke) gets a 5% interest rate from the bank because they're considered reliable, while Developer B (a smaller developer) gets a 9% interest rate due to higher risk. Even though they're building the same building, Developer B might end up paying more in interest each year, potentially eating into all their profits or even going bankrupt.
In the AI infrastructure game, financing costs (interest rates) directly determine whether a project is profitable. A few percentage points in interest can make a huge difference over billions in investments. Therefore, credit and financing capability have become as important as chip technology.
---
4. Is It a Bubble or the Future? The Time Difference Between "Money" and "Return?"
The biggest debate in the market is whether AI is a bubble. Wu Chaoze from CITIC Construction Investment pointed out a harsh reality: The computing power industry chain is taking more than 90% of the revenue and profits, while the model and application layers are still losing money, and commercialization is lagging behind.
- Upstream Profits, Downstream Challenges: Companies selling the necessary components (chips, servers, electricity) are making huge profits, but those developing AI products and services for users are still struggling to find ways to make money.
- Externalities in Technology: Peng Wensheng from Gaojin Think Tank believes that technological innovation has externalities. This means that investing in AI may not immediately yield profits for the investor, but it can boost overall societal productivity, benefiting everyone. Focusing only on a company's short-term financial reports can underestimate the value of AI.
In simple terms: It's like building a highway. The construction company might lose money or only earn a small fee, but once the road is completed, truck drivers, bus operators, and businesses along the route all benefit.**
Current AI infrastructure is like building a highway to the future. The question is: When will the traffic (commercial returns) arrive? If no traffic comes for 10 years, the construction company will fail. But if the traffic is heavy, the decision to build the road was wise. We're currently in an awkward period where the road is built, but the traffic is yet to arrive.
---
5. **The Ultimate Conclusion: A "Full-Element" Heavy-Asset Endurance Race**
Let's take a broader perspective. The so-called "computing power race" is no longer just about chips. It's a heavy-asset race that involves:
1. Obtaining GPUs (chips).
2. Ensuring high-bandwidth storage, interconnection, and cooling systems.
3. Building factories, providing power, and constructing stable foundations.
4. Managing financing costs (interest rates).
5. Generating sufficient commercial returns.
Any failure in any of these steps can paralyze the entire system:
- Even the fastest chips are useless without power.
- Even with sufficient power, a weak foundation can cause equipment to collapse.
- Even the most stable equipment can lead to losses if financing costs are too high.
- Even if the finances are manageable, if there's no demand from applications, it's a bubble.
What does this mean for ordinary people?
1. High Barriers to Entering the AI Industry: Small players can hardly enter the core computing power market by just buying a few cards. This is a domain dominated by giants and national teams.
2. Focus on Infrastructure, Not Just Models: In the coming years, companies in the fields of electricity, data centers, liquid cooling, and high-speed interconnection may be more stable and profitable than pure software companies.
3. Patience and Risk: The return on investment in AI has slowed down. Don't expect explosive profits next year, but don't dismiss the value of infrastructure just because there's no immediate application growth. This is a marathon that tests endurance, financial strength, and system engineering capabilities.
In summary: The second half of the AI race is about who is more "heavy" in terms of infrastructure, financing, and supply chain integration.