Summary of Key Points
Recently, there has been a “reverse operation” in the prices of large AI models between China and the United States: giants such as OpenAI and Google in the U.S. have successively reduced their prices (OpenAI by 20%-80%, Google by 50%), while Anthropic has even frozen its price increase plans; in contrast, China’s Kimi K3 has seen a price increase of 3.5 times, and DeepSeek has announced a significant price hike. Behind this contrast lie two different approaches to AI development—the U.S. follows a “high-cost, high-pricing” “oil model” (treating AI as a scarce strategic resource), while China adopts a “low-cost, high-value” “tap water model” (making AI accessible to various industries). China’s cost-effective approach has challenged the U.S.’s dominance in pricing, and the price increase in China also reflects the shortage of computing power and the need to move towards sustainable development.
Detailed Analysis
1. Price Reversal: Why Are U.S. AI Models Suddenly on Sale, While Chinese AI Models Are Getting More Expensive?
- Why Are U.S. Prices Falling?
The surface reason is “model optimization to reduce costs,” but in reality, it’s due to competition from affordable AI models in China. For example, OpenAI’s Luna model was reduced by 80%, Google’s Gemini is available at a 50% discount until 2027, and Anthropic canceled the price increase for Sonnet 5—all in response to competition from Chinese models like DeepSeek and Kimi (Wall Street refers to this as the “Kimi Moment”).
- Why Are Chinese Prices Rising?
The increase is not arbitrary: Kimi K3 uses a MoE architecture with 2.8 trillion parameters, which improves performance but also increases costs; DeepSeek’s price hike is due to a shortage of computing power. However, even with the price increase, Chinese models still have advantages: Kimi’s price is on par with mid-range models from the U.S., but 70% lower than the flagship models; moreover, Kimi’s caching technology reduces the actual cost of programming tasks significantly (with a caching hit rate of over 90%).
2. Two Different Approaches to AI: The U.S. and China
- The U.S. Model: AI as Expensive “Oil”
In the U.S., AI is treated like a costly, high-profit strategic resource. Tech giants (OpenAI, Google), top universities, and venture capital form alliances to invest billions of dollars in pursuit of the “strongest models and highest computing power,” earning high profits through technology monopolies. For instance, Uber spent its entire annual AI budget in the first four months of the year, and Salesforce pays $300 million to Anthropic annually.
- The Chinese Model: AI as Accessible “Tap Water”
In China, AI is seen as an affordable resource, focusing on open-source development, widespread adoption, and integration with real industries. Limited by computing power (lack of H100 chips), Chinese companies use innovative methods (MoE architecture, caching optimization, and engineering improvements) to reduce costs: training costs are one-tenth of those in the U.S., and the computing power used for inference is only one-third to one-half of the U.S. level, with GPU utilization exceeding the industry average by more than 20%. Additionally, Chinese AI API services generate a 20%-40% gross margin, making low prices sustainable rather than a loss-making strategy.
3. The Real Reason for China’s Price Increase: Not a Profit Driven Move, but a Step Towards Sustainable Development
- Computing Power Shortage Driving Prices Up
When Kimi K3 was launched, user requests pushed the system to its limits, forcing the suspension of new subscriptions. Chinese AI companies need more computing power, but devices like GPUs are expensive and hard to obtain, so price increases are necessary to cover real costs.
- From a “Land Grab” to a Focus on Quality
Previously, Chinese AI relied on low prices to capture the market; now, it’s shifting to a focus on quality. Kimi K3’s improved performance enables it to handle more complex tasks, so the price reflects its value. Even with the increase, Chinese models are still 70% cheaper than U.S. flagship models, maintaining a high cost-effectiveness.
4. The Crisis in the U.S. AI Model: High Costs Not Leading to High Returns?
- Low ROI on AI Investments
Companies spend $1 on AI tokens, only 18 cents of which generate actual value; 44 cents are used to fix bugs, and 27 cents go toward rework—money is spent with poor results. For example, Uber’s employees perform meaningless tasks to increase token usage, leading to uncontrollable costs.
- Narrowing Performance Gap
A Stanford report shows that the performance gap between top AI models in China and the U.S. has narrowed from 17.5% to 0.3% (2024 data; the gap is even smaller now, 2026). As performance levels converge, China’s cost advantage becomes a competitive advantage, forcing the U.S. to lower prices.
5. Behind the Computing Power Race: Big Companies Investing in AI
- Alibaba: Selling Games, Issuing New Shares
Alibaba’s capital expenditure increased by 75% in the second quarter, leading to the sale of its gaming business, Lingxi Huyu, and the issuance of HK$800 million in new shares to fund AI infrastructure.
- Tencent: Reducing Holdings in Kuaishou, Investing in Computing Power
Tencent’s capital expenditure in the second quarter was three times that of the previous year, mainly on purchasing computing equipment. To raise funds, it reduced its holdings in Kuaishou by 273 million shares (realizing HK$12 billion) and invested in Kuaishou’s AI video model, Keling AI.
- ByteDance: Planning $70 Billion in Capital Expenditure
ByteDance is considering increasing its capital expenditure to $70 billion in 2026, all for AI computing power.
In Conclusion
The price reversal in AI models between China and the U.S. reflects a clash between two models: one based on high costs and monopolies, and the other on low costs and widespread accessibility. Chinese AI is no longer just a follower in price competition; it’s changing the global AI landscape, making AI more affordable for everyone. This is good news for developing countries but poses a challenge to the U.S.’s high-pricing model.