Popular Summary of Key Points
Several seemingly unrelated financial and economic news stories recently highlight a significant shift in the AI industry: DeepSeek has reduced the pricing of its large-scale models by 40%, Elon Musk is investing in building his own natural gas power plants for AI use, there are rumors that DeepSeek is preparing for an IPO on the Science and Technology Innovation Board, and Apple has quietly increased the price of its new iPhones. These developments reveal that the focus of the competition in the AI industry has shifted from who has the smartest models or the most advanced parameters to who can minimize the cost of delivering AI services. While AI technology may seem to be becoming more affordable, the truly valuable assets are electricity, data centers, and the ability to efficiently manage computing resources.
---
Detailed Analysis in Four Dimensions
1. The Reduction in Large-Model Pricing is Not a Loss-Making Strategy, but a Way to Gain Access to the Capital Market
Let’s do some math: DeepSeek has lowered the cost of using its large-scale models to 0.02 yuan per million tokens during off-peak times. One token is roughly equivalent to 0.7 Chinese characters, meaning you can get 700,000 words of AI-generated content for just 2 cents—cheaper than half a tissue paper. This represents a 60% reduction in price. Some people call this a vicious price war, but it’s actually similar to a奶茶 shop reducing the price of a 20-yuan drink to 10 yuan: if sales increase by more than three times, total revenue will be higher, and inefficient players will be eliminated. DeepSeek is already in talks with securities firms for an IPO. When it submits its prospectus, the capital market won’t be interested in how impressive its model parameters are or how world-leading its capabilities are. Instead, the market will focus on three key figures: first, how much actual money users are paying each month (not just fake usage counts); second, how much profit is left after deducting all costs for every million tokens sold; and third, how much money is needed to expand computing capacity each month and whether there is enough cash on hand. If the increase in usage outpaces the price reduction, it’s a profitable strategy that will boost the company’s valuation. If no one uses the services despite the lower price, then it would be a losing proposition that would severely damage the company’s valuation.
2. Although AI Services Are Cheaper per Use, the Costs of GPUs and Data Centers Have Not Decreased
Many people think the price drop is due to cheaper GPUs, but that’s not the case. To illustrate, training large-scale models is like building a superhighway with hundreds of lanes, while our daily use of AI, such as writing copy or processing data, is like driving on that highway. Previously, companies competed to build the widest highways; now, the goal is to generate revenue from traffic. However, traffic patterns are highly uneven—for example, there’s a peak in usage during working hours for tasks like report writing and coding, but almost no usage at night. Maintaining a fully loaded GPU cluster all year round just to handle the peak traffic would be inefficient and wasteful. Solutions like cache reuse and flexible scheduling aim to make better use of GPUs, allowing one GPU to perform the work of several, thereby reducing the cost per use. For AI to become widely adopted, the cost per use must be significantly lowered.
3. Elon Musk’s Investment in Gas Power Plants Indicates That the Scarce Resource in the AI Industry is No Longer GPUs
Previously, the entire industry was competing for NVIDIA GPUs, and whoever could obtain them was considered a leader. Now, companies realize that having thousands of GPUs without a place to install them or a stable power supply makes them almost useless. Musk’s investment in a 1.2-gigawatt natural gas power plant for xAI, and even his efforts to have SpaceX manufacture key components for gas turbines, show that the industry can no longer wait for the traditional power grid expansion process, which can take three to five years. Building a power plant on its own allows for immediate power supply. The pace of change in the AI industry is rapid: if you launch computing resources three months late, your customers will be taken by competitors; if your power plant is not operational for a year, the GPUs you’ve purchased will sit idle in storage. This year, GE’s AI-related power generation orders have already exceeded the annual forecast for 2025. What’s in demand is not just any power supply but a stable one that can be delivered on time, 24/7, and without sudden price spikes—this is even more scarce than NVIDIA GPUs.
4. Computing Resources Will Be Hierarchized, with GPUs Becoming the Least Valuable
NVIDIA’s stock price has been declining recently, not because the market is pessimistic about AI, but because people are realizing the long-term costs of investing in GPUs and data centers. In the future, computing resources will be categorized into three tiers:
- The lowest-tier resources, which are just piles of GPUs without stable long-term customers or power supply, will depreciate rapidly and become worthless.
- The mid-tier resources, which can efficiently manage computing resources and reduce costs, will generate steady profits.
- The highest-tier resources, which include access to power supplies, necessary permits, and long-term customers, will be in high demand and considered the most valuable assets in the AI industry.
Once DeepSeek goes public, it will set a clear benchmark for the entire large-model industry. Future success will no longer rely on claims about having the “world’s most advanced parameters” or the “smartest models.” Instead, the key indicators will be usage volume, profit margins, and cash flow—these are what the capital market will evaluate. Stories about “smart models” won’t be enough to attract investment anymore.