Core Summary
DeepSeek has adjusted the prices of its AI models twice within a week (first increasing the peak prices and then applying the lower prices throughout the weekend). This apparent contradiction is actually a strategy to distribute the high demand for computing power more evenly. This marks the end of the “low-price competition” in the AI model industry, as companies are now focusing on competing on the value of their technology. The market has reacted differently: stocks in the computing power hardware sector have declined in the short term, but Alibaba’s fundraising of 80 billion yuan for AI infrastructure has been eagerly sought after by long-term investors. The increased prices have benefited the upstream components of the computing power chain (such as optical modules and servers), and the leasing model has shifted from fixed rents to a revenue-sharing system based on Token usage. However, the sustainability of this trend depends on whether users accept the price increase and whether the technological path and trade policies remain stable.
1. Price Adjustments Are Not Arbitrary; They Signal a Shortage of Computing Power
DeepSeek’s first price increase raised the output prices by 350% during peak hours (9-12 AM and 2-6 PM) and implemented tiered pricing. The second adjustment applied the lower prices during the weekend. Why? Because the demand for computing power during peak times is genuinely insufficient—for example, its weekly usage increased by 570%, often resulting in “capacity shortages.” Nationally, the usage of Tokens has increased 1000 times in the past two years, but there is a “structural gap” in supply: 80% of the real-time demand is concentrated in the eastern regions, while 80% of the training tasks are in the western regions. Data centers can be built in 100 days, but the necessary infrastructure (power supply) takes two to three years to develop.
Price adjustments essentially use price mechanisms to direct demand: less urgent tasks (such as batch data processing) are shifted to nighttime or the weekend, while the computing power during the day is reserved for real-time needs that are more costly (such as user interactions and real-time translations). The market quickly understood the implication of the first price increase, with computing power stocks hitting their daily limits on the same day, as people realized that it was a shortage of computing power, not an increase in the cost of Tokens.
2. Price Increases Indicate a Shift from Low-Price Competition to Competitiveness Based on Technology
For the past two years, AI companies have engaged in price wars, selling Tokens at very low prices, implying that the focus was on acquiring users as quickly as possible, assuming the technology was not valuable. However, low prices are not sustainable for research and development; the smarter the models, the more expensive their training and maintenance become. Now that DeepSeek has led the price increase by 350%, followed by companies like Zhipu and Tencent Cloud, it indicates a shift in the industry from competing on price to competing on the intelligence of their models.
A Morgan Stanley report emphasizes that the intelligence of models is the key to long-term competitiveness, as low margins cannot support the development of next-generation models. Analysts suggest that the low-price competition among domestic AI companies has officially ended, and prices are beginning to return to a more reasonable level. The end of the price war sends a signal to the entire industry that investments in computing power need to be profitable, which is reflected in the stock market.
3. Where Does the Extra Money From Price Increases Go?
The extra revenue generated by the price increases is flowing back into the upstream components of the computing power chain: AI companies use it to purchase more computing power, through the following channels: AI companies → cloud service providers → servers → optical modules → chip manufacturers → equipment manufacturers. Recent financial reports from upstream companies show significant growth:
- Zhongji Xuchuang’s revenue increased by 182% in the first half of the year, with a 242% increase in net profit, and orders extended into 2027.
- Huahong Hongli, a chip manufacturer, saw its revenue reach a record high in the second quarter, with 60% of the growth coming from price increases.
- Zhongwei, an equipment manufacturer, reported a 300% increase in net profit.
More importantly, the leasing model for computing power has changed: previously, rents were fixed based on the number of GPU cards, but now they are split from the revenue generated by Token usage (as seen in contracts between Xingyun Technology and leading AI companies). This means that computing power has shifted from being a cost to becoming an asset—the more expensive the Tokens, the more revenue computing power companies can earn. This model is sustainable because the demand for daily tasks (such as chat and office AI) is growing and is a stable source of revenue, unlike the more intermittent needs for model training.
4. How Long Will This Trend Last?
The sustainability of this trend depends on two key factors:
- Will users switch to other options? Alibaba has made its flagship models open-source, allowing companies to deploy them on their own, which could potentially offset the impact of DeepSeek’s price increases. We need to monitor three indicators: whether usage decreases, whether the utilization of computing power during off-peak times increases, and whether the price difference between peak and off-peak usage narrows (indicating user willingness to use cheaper alternatives).
- Technical and Policy Risks: Even with positive financial reports, stock prices may still decline if there are concerns about potential changes in technology (such as the emergence of cheaper alternatives) or trade policies (such as restrictions on the export of optical modules). Zhongji Xuchuang, for example, receives 94.8% of its revenue from overseas, so a ban could severely impact its business.
In short, if Tokens become more valuable and computing power becomes more expensive, upstream companies will have more orders. DeepSeek’s price increases serve as a barometer, reflecting the intensity of demand as well as the stability of technological and policy factors.
5. Market Divergence: Short-Term Focus on Performance, Long-Term Investment in AI Infrastructure
On August 24th, stocks in the computing power sector declined, but Alibaba’s fundraising of 80 billion yuan was eagerly sought after by long-term investors, with demand exceeding the initial target by three times. The reasons for this contrast are:
- Short-Term Investors: They were previously focused on the perceived shortage of computing power but are now evaluating actual performance. Although Zhongji Xuchuang’s financial reports are positive, they are concerned about future risks, leading to sales.
- Long-Term Investors: Alibaba’s investment in AI infrastructure indicates their belief in the long-term need for substantial computing power, and they are willing to wait for returns in the future.
This divergence reflects different perspectives on short-term fluctuations and long-term trends. The core message of this news is that the supply-demand imbalance in the AI computing power market has reached a point where price adjustments are necessary. The industry is shifting from a focus on acquiring users at all costs to investing in research and development. The sustainability of this trend depends on user acceptance and the stability of technological and policy factors.
For the general public, understanding this logic helps explain why some AI stocks are declining while other investors continue to invest in the AI sector.