Summary of the Key Points
This article is a sobering critique of the current AI hype by technology thinker Benedict Evans, who highlights eight “truths” about the commercial reality of the AI industry: Basic model companies may become mere “pipelines” that earn meager profits; the current low subscription fees are merely a illusion fueled by subsidies; only the programming field has truly embraced AI; there is a contradiction between the perceived “definite demand” for AI and enterprise software; the concept of a “universal model” is an illusion; integrating AI into specific workflows is the key to success; the future of AI cannot be accurately predicted by historical analogies; internal talent and process issues within the industry are more critical; and when AI truly changes the world, it will likely become ubiquitous yet somewhat mundane. Essentially, Evans wants to convey that AI is not omnipotent, and behind the hype lies a complex web of commercial logic. Only by finding solid applications and profit-making models can companies survive in this landscape.
Detailed Analysis
1. Basic Model Companies May Become “Pipe Workers”
Imagine basic model companies (such as OpenAI or Anthropic) as those who build highways: they invest billions to create infrastructure that could transform the world, but ultimately they only earn a small fee for usage (e.g., through token charges). The real profit comes from businesses that operate on top of this infrastructure, such as those that develop applications using these models. This is a common pattern in technology history—companies that laid the groundwork for internet connectivity, like those who provided bandwidth, later reaped the benefits while companies like Google and Facebook made substantial profits. AI model companies face the same risk: no matter how advanced their models are, others can build applications on them, leaving them with minimal profits.
2. $20 Subscriptions Are a Loss-Making Strategy
Currently, paying $20 per month to use services like ChatGPT or Claude seems like a good deal, but in reality, the model companies are losing money. The cost of each interaction and each generated word (represented by tokens) is high, and $20 does not cover these expenses. This is similar to how ride-hailing companies subsidized their services to attract users. In the future, AI fees will likely be based on usage, just like mobile data plans—only what you use will be charged. For example, writing a long report with AI might cost several dollars rather than a fixed fee.
3. Only Programming Has Truly Embraced AI; Other Applications Are Just for Show
To determine if AI has been truly integrated, look at whether users use the services daily and continue to do so (high retention rates). So far, “agent programming” (where programmers use tools like Copilot to write code) is the exception. In other areas, such as chatbots or AI-generated art, users often give up after a short trial period. These are more superficial uses and not represent genuine demand. Don’t be misled by headlines claiming that AI will revolutionize industries; most applications are still in the “trial phase.”
4. Integrating AI into Enterprise Software: High Uncertainty Leads to Low Profit Margins
Traditional enterprise software (e.g., financial or customer management systems) requires absolute accuracy—clicking a button should always produce the expected result. However, large language models (LLMs) are probabilistic and may generate incorrect outputs (e.g., miscalculations or missing contract terms). Introducing such uncertain technology into critical applications leads to increased costs for verification and reduced user satisfaction, resulting in lower profit margins. Companies will be less willing to pay for tools that frequently fail.
5. Don’t Believe in “Universal AI Models”
Many people dream of a model that can solve everything, but Evans argues this is unrealistic. The real competitive advantage lies in integrating AI into specific industry workflows. For example, a law firm doesn’t need an AI that can write all contracts; it needs an AI that can automatically generate draft contracts and match them with relevant cases, seamlessly integrating into existing processes.
In Conclusion
The AI boom is not about easy profits but about understanding commercial logic: avoid becoming a “pipeline” that merely collects fees. Don’t be fooled by subsidized low prices; focus on practical applications that address real work challenges. Only by finding meaningful use cases can companies thrive in the AI landscape. When AI becomes as ubiquitous and reliable as water and electricity, it will truly transform the world.