Summary of Key Points
Recently, the AI industry has been in the midst of a heated debate about whether open-source models should be restricted. Sam Altman led a public letter advocating for the openness of weights, with 25 organizations including Microsoft and Meta signing on in support, while Elon Musk also voiced his approval. OpenAI and Anthropic, which had initially not signed the letter (and had previously sought restrictions on Chinese open-source models in Washington), later joined as well, raising the total number of signatories to 77 within three days. Behind this controversy lies a battle for interests among different companies: closed-source model providers (such as OpenAI) fear that open-source models will steal market share, while those selling computing power and cloud services (like NVIDIA) prefer greater model popularity.
At the same time, the AI model industry is exhibiting three key trends: the lead-time between innovations is getting shorter; price wars initially involve free offers followed by price increases; and corporate customers are the real source of profit. Ultimately, what determines a company's success or failure is not the model itself, but the advantages it has beyond the model—such as access points, customer retention, products, and industry rules.
1. The More Money Burned, the Shorter Lead Time? Because There Are No Secrets in Technology
You might think that AI companies spending so much on training models should gain a monopoly advantage, but the reality is that the lead-time between innovations has shortened to just months.
- The gap is minimal: A Stanford report shows that the difference in capability among the top four models is less than 25 percentage points, and in professional tests, the difference among the top 15 models is only 3 percentage points. Just as oil companies don't charge ten times more for crude with a 3% higher purity, model differences of this magnitude don't justify higher prices.
- Rapid turnover: In the past 39 months, the leader in model performance has changed 21 times, with an average tenure of less than two months. For example, Gemini 2.5 Pro caught up to OpenAI's o3 at a quarter of the price, and OpenAI immediately reduced its pricing by 80%—once the lead was challenged, the power to set prices disappeared.
- Lowering the cost of catching up: Training OpenAI's GPT-4 is estimated to cost hundreds of millions, while DeepSeek R1 managed to match o1’s performance for just $5.57 million; Kimi K3, being open-source from the start, quickly rose to the top in code-based tests. Why? Because there are no secrets in technology: architectures are published in papers, training methods are documented, computing power is clearly priced, talent moves freely, and models can be rapidly improved by using others' outputs as a basis for further development. This industry is like commodities without patent barriers; anyone can enter.
2. The Final Outcome of Price Wars: Free Offers to Capture Market Share, followed by Elimination and Price Increases
The Chinese market has already played out this price war scenario globally:
- Step one: Free offers to attract users: In May 2024, DeepSeek reduced its API prices to 1% of GPT-4’s rate, while ByteDance’s model was offered for free. However, using models consumes energy and chips, so free offers mean more losses for the companies—large firms can afford these losses (by using cloud services to recover costs), but startups cannot.
- Step two: Eliminating weaker competitors: In January 2025, Li Kaifu’s ZeroOneWanShi company abandoned its large-scale models and integrated them into Alibaba Cloud, marking this year as the “year of commercial elimination.”
- Step three: Surviving companies raise prices for modest profits: The survivors start charging higher fees, with over 70% of manufacturers increasing their prices. For example, DeepSeek first reduced prices before doubling them for its flagship models, with prices rising from low levels to match those of Claude. In the end, it’s like in the oil industry: those unable to compete on cost will exit, while the remaining companies make small profits.
3. Who Really Makes Money Behind the Heavy Spending? Corporate Customers Are the Real Cash Cows
Model companies burn a lot of money, but not everyone benefits:
- OpenAI: High revenue but “leaking cash”: In 2026, OpenAI’s revenue exceeded $25 billion, but only 59 million out of its 900 million users paid. Free users drag down the gross margin, and it will need to continue spending $218 billion by 2030 to become profitable.
- Anthropic: Profitable from corporate customers: In the third quarter of 2026, Anthropic reported a profit of $1 billion, with 60% of its revenue coming from corporations. The key is the “switching cost”—once companies integrate the models into their workflows, employees become accustomed to them, and data accumulates, making it more costly to switch suppliers. For example, customers who paid $100 last year may pay $500 this year, resulting in a customer retention rate of over 500%.
- Domestic companies: Many users but lower revenue: Chinese companies like Zhipu generate annual revenue of 700 million RMB, compared to the American leaders’ 300 million USD. However, Chinese models are being used more frequently in China than American ones on OpenRouter, thanks to their open-source approach and lower prices.
4. Advantages Beyond Models: The True Moats
Models alone cannot sustain a business model; companies need other factors to survive:
- Access points: The scale of user base creates a distribution network—OpenAI’s 900 million active users are a natural traffic pool that others cannot steal.
- Customer retention: High customer retention rates mean customers are reluctant to switch to different models.
- Products: Specialization in specific fields: For example, Anthropic’s Claude Code generated $1 billion within six months of being released, thanks to its specialized capabilities in coding.
- Industry rules: Influencing government regulations is the fastest way to maintain a competitive edge. If a model can’t maintain its lead, companies try to persuade governments to establish rules in their favor. For instance, OpenAI wanted to define “model distillation” as theft, but Sam Altman defended it (because more open-source models mean better sales for NVIDIA’s chips).
In Conclusion
AI models are becoming like commodities, and relying solely on model superiority is no longer enough. Future survivors will either have a massive user base, stable corporate customers, or the ability to influence industry rules. In other words, companies need to find their own “moats” beyond the models themselves. Those that sought restrictions in Washington realize that the advantages of their models are no longer sustainable.