AI Cost Reflection: Calculate ROI Before Entering the Productivity Market

Recently saw Wang Puzhong reflecting on the waste caused by universal AI adoption, with daily compute costs peaking at 10 million yuan. This detail is more worth watching for valuation circles than the productivity market itself. Big tech pulling AI from slogans back to cost sheets shows that the "deploy models first, find scenarios later" approach has reached its homework submission time.

The ceiling of this track depends on whether customers are willing to keep paying for results. Model parameters and free quotas don't determine it. Liu Chiping calls the new beachhead for AI commercialization the "productivity market," which is essentially about grabbing workflow entry points. The problem is pricing is still rough—subscriptions, pay-as-you-go, and free subsidies are mixed together. Generating a weekly report vs. an investment research draft might use the same number of Tokens, but their commercial value differs vastly.

I've been using GPT-4 and LiST for data sorting recently, used DeepSeek Harness for two weeks, and also built workflows with Trae and QoderWork. The feeling is quite clear. What really saves money is whether the model can turn dirty, tedious work into verifiable deliverables. Recently helped a friend look at a SaaS company purchasing AI office tools. The buyer asked if it could replace positions, how many months to recover costs; whether it could write code was actually secondary. The seller was still telling stories based on call volumes.

Previously I worried about securitizing AI infrastructure assets because chips become obsolete and software iterates too fast. Now my thinking has shifted slightly. Compute can be priced, but shouldn't be sold as a resource—it must be sold by result. Whoever can get enterprises to move AI budgets from innovation expenses to operating costs has pricing power. How's the competitive landscape? Short term relies on subsidies, mid-term looks at scenario understanding, long term competes on ecosystem. Big tech has traffic and cloud, SMEs understand processes better. In the end, it likely won't be winner-takes-all, but it will wash out those who only know how to sell Tokens.