Summary of Key Points
The K3 model released by Kimi has attracted attention for its 2.8 trillion parameters (the largest among domestic models) and for surpassing Claude in the Code Arena programming competition, becoming the first open-source model to do so. However, its high price of $2.3 per million tokens has sparked controversy, as it represents a new record for domestic models. The article compares Kimi with companies like Moutai and Apple to highlight that large models currently only have a “short-term pricing window” rather than long-term pricing power. Pricing power relies on barriers that users cannot easily overcome, such as ecological lock-in or access to scarce resources. Additionally, the low cost of switching between large models and the short lifespan of their “SOTA” (State-of-the-Art) status make it difficult for them to maintain high prices in the long term. The article suggests that the true sources of pricing power for large models may lie in “upper-layer application entry points” (where users form habits) and “lower-layer scarce resources” (such as computing power or chips). During this window period, companies should focus on building user loyalty rather than simply raising prices.
I. Kimi K3’s Explosive Success and Pricing Controversy
The K3 model has gained prominence for two reasons: its large number of parameters (2.8 trillion) and its victory in the Code Arena programming competition, where it surpassed Claude fable5 to become the first open-source model to top the rankings. This success caused a significant drop in the stock price of Zhispu (a Chinese AI company), with its value evaporating by HK$300 billion in just two days, and it also impacted the valuation expectations of Anthropic.
However, controversy arose quickly over the high pricing of K3’s API at $2.3 per million tokens, which is a new record for domestic models. Some argued that being the “world’s best” should justify the higher price, while others questioned the fairness of such prices for open-source models. Although Kimi’s representatives claimed that open-source models shouldn’t be labeled as cheap, the author questioned whether SOTA models truly possess pricing power.
II. What Really Constitutes Pricing Power? Three Examples Will Make It Clear
Pricing power does not mean being able to raise prices just because you are temporarily ahead; it means that customers remain loyal even after price increases. Warren Buffett once said, “A good business is one where you can raise prices without losing customers; a bad business is one where you have to pray for a 10% increase to still retain them.” Here are three examples of companies with true pricing power:
- Moutai: It’s not just delicious; its value also lies in the social respect it commands. Moutai is often used in business settings as a symbol of importance, and those who need it have the financial means to afford it (you can’t replace it with another brand like Erguotou).
- Apple: Users’ apps, photos, and workflows are all integrated into the iOS ecosystem. Switching to Android requires re-adaptation and data transfer, which is costly (the author experienced this personally).
- NVIDIA: Approximately 80-90% of large model computing power uses NVIDIA chips. Other chip manufacturers either have inferior performance or lack compatible software ecosystems, leaving customers with no choice.
The common factor in these companies is that their products are irreplaceable for their customers, which is the essence of pricing power.
III. Why Do Large Models Lack Pricing Power?
Large models face two major obstacles to establishing pricing power:
1. Low switching cost: Switching to a different model is like changing a website address; users can easily switch from K3 to Claude-fable5 in just ten minutes, with almost no barrier to entry.
2. Short lifespan of SOTA status: The technology landscape changes rapidly, and the “SOTA” status only lasts for about half a year. Models like Gemini from last year are now considered outdated, and others like DeepSeek, Zhispu, and Kimi also only remain popular for short periods. A model that frequently changes leaders cannot claim true pricing power.
Therefore, the high prices of large models are more of a “pricing window” – an opportunity to charge a premium while there are no competitors. Once competitors catch up or open-source models lower prices, those high prices will likely fall. This is not a sign of long-term pricing power but rather a temporary advantage.
IV. Where Pricing Power for Large Models Might Lie
It’s difficult for models to establish their own pricing power, but two areas in the industry could potentially provide it:
1. Upper-layer application entry points: By integrating users’ workflows, data, and habits into the model platform, companies can create a loyal user base. For example, if users develop custom commands and project contexts using an intelligent office tool, they may be reluctant to switch to another platform.
2. Lower-layer scarce resources: Access to valuable resources such as computing power or advanced chips can give a company pricing power. NVIDIA’s chips are essential for large model training; without them, other models struggle to perform effectively.
Models themselves are not highly valuable on their own, but when combined with the necessary underlying technologies, they can command higher prices from users.
V. What to Do During the Pricing Window Period
Even without direct pricing power, the temporary lead provides an opportunity to build user loyalty:
- Smart strategy: Use this time to establish a strong user base by encouraging them to use your model and accumulate valuable data and habits. For example, let users create projects or save custom commands on your platform. When competitors emerge, users may find it too inconvenient to switch, thus maintaining your position.
- Wrong strategy: Relying solely on price increases during the window period can lead to customer loss. Once the window closes, you’ll have to start from scratch again, similar to running on a treadmill without ever truly catching up with competitors.
In summary, while the performance of large models is impressive, their true value lies in their ability to retain users or control scarce resources. Relying solely on temporary SOTA advantages for price increases is not a sustainable strategy.