Summary of Key Points
The global data center industry is set to experience explosive growth over the next five years, with capacity increasing from 82 gigawatts to 219 gigawatts, corresponding to capital expenditures in the trillions of dollars. However, the market has begun to question the speed at which cloud providers are spending money. Although the physical hardware of data centers (such as power supply and cooling systems) will not be immediately discontinued due to their long construction cycles, the expenditure on computing power for training AI models has entered a more cautious phase. Future demand will shift from "AI training" to "AI inference" (using learned knowledge to perform tasks). Nevertheless, this period of prosperity could face risks such as oversupply, declining demand, or a shortage of financing. The key lies in whether cloud providers' revenue growth can keep up with their capital expenditures and whether the pressure on financing will increase.
1. How much money is being spent on data centers?
Over the next five years, the global data center capacity will increase by approximately 137 gigawatts (from 219 to 82 gigawatts). According to NVIDIA's Jensen Huang, the cost of each gigawatt of computing power ranges from $80 billion to $100 billion, resulting in total investments of $10 trillion to $13 trillion—this is roughly equivalent to 10% of the world's annual GDP. Such substantial investment has caused concern in the market: the Philadelphia Semiconductor Index (a benchmark for the chip industry) has dropped by 13% in recent months, raising concerns that cloud providers may be spending too much without generating sufficient returns.
2. The physical layer cannot be stopped, but the training layer needs to slow down
Data center construction consists of two parts:
- Physical layer: Hardware components such as power equipment, transformers, and cooling systems take 2 to 5 years to build, and projects that have already started cannot be halted easily, so these expenditures will continue.
- Training layer: The computing power used for training AI models (e.g., high-end GPUs) is now being approached with caution. Josik suggests that the "training layer has reached its later stages," meaning demand for AI training is nearing its peak, and further investment could be wasteful, as not every company needs to train large models like ChatGPT.
Even if we account for equipment depreciation (e.g., GPUs becoming less valuable over time), the conclusion remains the same: the key is whether cloud providers' revenue growth can keep up with their spending. If spending grows much faster than revenue growth, "capital expenditure watchdogs" will emerge—investors may sell stocks to force companies to reduce their spending.
3. Shifting from "AI training" to "AI inference": A change in demand
Currently, data centers are primarily used for training AI models, but after 2027, more computing power will be allocated to "inference" tasks (e.g., using AI to answer questions, recommend products, and make autonomous driving decisions). The reason is simple: training is a one-time process, while inference is an ongoing activity that requires continuous use of AI.
However, there is a challenge: the benefits of the current investment in training hardware will only be visible in five years (e.g., improved efficiency in businesses). If cloud providers' revenue slows down before these benefits materialize, the market may adjust before the returns are realized. Josik predicts that there could be a period of reduced spending in the next two years until demand for inference truly takes off.
4. How to identify the end of this boom?
Josik believes this data center boom could end similarly to three previous crises, as indicated by three observable indicators:
1. Oversupply (similar to the 1999 internet bubble): Check the vacancy rate (currently only 1% in North America, which is still considered safe) and GPU rental prices (if they haven't dropped, demand is not excessive).
2. Declining demand (similar to 2022): The market will react negatively if cloud providers' capital expenditures exceed revenue growth by more than 1.5 times (for example, Amazon's and NVIDIA's stock prices fell by over 50% in 2022).
3. Lack of financing (similar to the 2008 financial crisis): Monitor the proportion of companies' capital expenditures as a percentage of their cash flow—Amazon is approaching 100%, and Google has already started issuing bonds. If this ratio exceeds 100%, investors will be reluctant to lend to them.
The two most critical indicators are the growth rate of cloud services and the spending plans of cloud providers. If companies begin to cut back on their spending plans, it indicates that demand is indeed weak.
5. The debate over depreciation: Will GPUs become obsolete quickly?
Some worry that the rapid depreciation of data center equipment (especially GPUs) will erode profits. Josik argues that there's no need to panic:
- High-end GPUs used for training can be repurposed for inference tasks, which require less computing power.
- A secondary GPU market already exists, allowing for the reuse of equipment and extending its actual lifespan beyond what accounting estimates suggest.
He is more concerned about whether companies will be forced to borrow money due to insufficient cash flow. If borrowing becomes excessive, the market will exert pressure on them to adopt more "capital discipline" (i.e., more cautious spending).
In conclusion: Data centers are still growing, but the phase of reckless spending may be coming to an end. There could be adjustments in the next two years, but the demand for inference tasks will become a new driver of growth in the long term. Ordinary investors should pay close attention to cloud providers' revenue growth and spending plans as key indicators of market trends.