Summary of Key Points
This article challenges the conventional perception of the risks associated with AI investment: People debate whether AI represents a “real revolution” or a “bubble,” assuming that the main risk lies in the technology’s shortcomings. However, the article argues that the real threat to AI is its tremendous success and consequent low cost. Similar to the photovoltaic industry in the past, Chinese companies have used their aggressive pricing strategies to turn AI models into commodities, squeezing profit margins. As a result, companies that solely sell AI models or computing power are unable to make a profit, with value shifting towards “physically scarce resources” (such as electricity and advanced hardware) and “user-centric applications” (such as workflows and private data).
1. The Risk of AI Lies Not in Overstatement, but in Its Excessive Success and Low Cost
The market is divided between proponents (optimists) and skeptics (pessimists) of AI, both based on the assumption that the danger lies in the technology not being as powerful as expected. But the article suggests this might be completely wrong. The real threat is when AI becomes too advanced:
- Technological maturity lowers the barriers to entry, allowing anyone to enter the market;
- Chinese companies, with their low-cost strategies (for example, DeepSeek reducing model prices to one-twentieth of Western levels), have made AI models extremely affordable;
- As a consequence, corporate customers are unwilling to pay a 25-fold premium for a 0.2% improvement in performance, forcing Western giants like OpenAI and Google to lower their prices as well, thereby eroding profits.
2. The Hard Lessons of the Photovoltaic Industry: Technology Wins, but Companies Lose
The article uses the photovoltaic industry as a metaphor to illustrate a harsh reality: successful technology does not necessarily lead to profitable businesses. Around 2010, everyone saw photovoltaics as the future, and Chinese companies and governments invested heavily (in land, loans, tax incentives). This quickly eliminated technical barriers, allowing anyone to produce photovoltaic components, with China accounting for a large share of global production capacity.
- The outcome was disastrous: prices plummeted by 40% in 2011, making solar energy the cheapest form of power;
- Many companies went bankrupt or restructured, with some experiencing gross margins as low as -60% (losing 60% on every dollar sold);
- Even if you accurately predicted the success of photovoltaics, investing in related stocks could still result in losses due to overcapacity and resulting inered profits.
3. AI is Becoming an Inexpensive Commodity: Chinese Low-Cost Models are Pushing Western Giants
Why is AI more likely to become a commodity than photovoltaics?
- Zero switching cost for models: Large AI models are available through APIs, allowing developers to easily switch between them without significant barriers (unlike the Office software ecosystem);
- Chinese low-cost strategies: Companies like DeepSeek have made models extremely affordable (1 dollar per million tokens, compared to Western prices of 25–30 dollars), with only a slight difference in performance;
- Low cost for widespread use: Future intelligent agents (e.g., booking flights or creating travel itineraries) will frequently use these models, so they need to be inexpensive;
- Zero replication cost: While photovoltaic production requires mining silicon and building factories, AI model training has virtually zero replication costs, making price competition even more intense.
Western giants’ current price cuts (e.g., lightweight and fast versions of models) are a direct replica of the strategies used by photovoltaic companies to survive in a competitive market.
4. Where Does the Value Go? Two Directions: Physical Scarcity and User-Centricity
If AI models become as ubiquitous as free water, where will the value flow? Drawing from the photovoltaic example:
- Physically scarce resources: While AI models can be optimized, resources like electricity and advanced hardware (e.g., TSMC’s manufacturing services, NVIDIA’s GPU packaging) are difficult to replicate; these demands remain rigid.
- User-centric applications: Although large AI models themselves have no barriers, companies that integrate them into daily workflows (e.g., CRM systems) or possess unique private data can profit. Just as bottled water is more expensive than tap water due to processing and distribution, those who can connect free models with user habits gain control over pricing.
5. Investment Insights: Avoid Companies That Sell Only Models, Focus on Two Areas
The article provides clear investment advice:
- Avoid companies that solely sell AI models or computing power: Their profits will be eroded by price wars, similar to what happened to photovoltaic manufacturers;
- Invest in physically scarce resources: Such as electricity (required for running AI models) and advanced hardware (e.g., TSMC’s high-end packaging);
- Focus on user-centric companies: Those with proprietary workflows (e.g., enterprise software) or private data (e.g., healthcare, finance): They can use free models to build strong ecosystems and earn money from providing related services.
The final thought is poignant: “The people who change the world are often not the same ones who make money.” AI will transform the world, but companies that only sell models may become victims of its own technological success, similar to the photovoltaic manufacturers of the past.
The core logic of this article is that the ultimate risk with AI is not technical failure, but rather the resulting price competition caused by its excessive success. Chinese companies’ low-cost strategies have accelerated this process, turning AI from a high-profit technology into a commodity, with value shifting to two key areas: physically scarce resources and user-centric applications. Investors should avoid the intermediate “model layer” and focus on these more valuable aspects.