The competition in AI is fierce. "Expanding from models down to scenarios" and "Integrating models upward from scenarios and mastering data downward" are two different paths. Everyone expresses their views based on their own positions. Here, I'll restate the viewpoint expressed in Karp's interview, putting it more bluntly.

"If AI is really worth that much money, then let the results speak, don't use those subscription-based gimmicks. The current token-based charging model has actually exposed a weakness of frontier labs: they only understand models, not business delivery.

What enterprise clients truly want is to solve specific business node problems, not to buy a chatbot that can chat and write code.

Karp directly pointed out that these model vendors are exploiting information asymmetry to sell over-packaged computing power to large companies.

The current Fortune 500 companies are like free coaches for AI companies.

Paying high tuition fees and also providing the highest quality training materials for free. The chemical reaction generated by the combination of computing power, models, and business logic, which should have been a corporate trade secret, is now all exposed on the servers of AI labs.

This business model works purely because everyone is still in a state of technological panic. Afraid of being left behind by the times if they are a step slow. But Karp said: this kind of robbery-style growth is unsustainable.

The AI that enterprises truly need shouldn't be this kind of outsourced brain that could betray them at any time. It should be a system that grows on the enterprise's own architecture, capable of closed-loop control over data weighting and computing power allocation.

The next revolution in the AI industry probably won't be the model parameters multiplying several times over, but how enterprises take back data sovereignty from these labs."