Large model prices have fallen into the li era, and the people who should catch this wave of gains are not coders.
Last month a friend of mine who runs a tea business threw his own gift-box price list at a large model and asked it what Jin Jun Mei gift boxes generally sell for on the market, whether his custom price was expensive, and where the goods should be sourced from. The answer was beautiful, fluent prose. Unfortunately, all of it was invented.
That is not the model's fault. It is that what we fed it was wrong. And this is exactly the signal most worth understanding in the industrial shift now under way.
1. A price collapse: compute, for the first time, absurdly cheap
Start with the facts that have just landed.
Alibaba Cloud pushed the input price of its main-tier large model β the one benchmarked against GPT-4 β from two fen per thousand tokens all the way down to about half a li, a drop of 97 percent. The flagship Qwen-Max also came down to 0.04 yuan, a fall of 67 percent. ByteDance, Zhipu and Baidu all cut prices this month too.
The industry calls it the li era.
What does that mean? In the past, if you wanted AI to do work that needed a bit of brains, the calling cost had to be counted on your fingers. Today that money has become cheap enough to be almost negligible. Guangming Daily put it precisely in a report on 3 September: China's large model industry is moving from capability premium towards economic rationality, entering a dual competition paradigm where capability is king and cost decides the winner.
Put that in plain language: capable large models will become cheaper and more widespread. The water, electricity and coal era of AI has genuinely arrived.
2. But having a model does not mean having anything useful
With compute this cheap, AI should be blooming everywhere. Reality disagrees.
A 2026 report from Deloitte and Google Cloud gives a set of numbers that sting: companies are piling into AI agents, but only just over 30 percent actually reach production and genuinely do work.
The bottleneck is not the model β the model is already smart enough and cheap enough.
Where is the bottleneck? In the data, and in deployment.
Think about that tea shop. What is its most valuable asset? Not a chatbot, but the things it holds that nobody else does: real Jin Jun Mei market prices, the floor price on its own supply, procurement channels upstream and down that can be trusted. Those things are scattered across the owner's head, Excel sheets and chat logs with suppliers, and have never been properly gathered and organised into knowledge an AI can understand and responsibly cite.
A bare large model, facing the question of what Jin Jun Mei gift boxes cost on the market, has exactly one option: to lie with a straight face. It has never seen the market; it has no source anchor.
So the conclusion is clear: having a model is not the same as having something useful. The model is the engine, data is the fuel, and the people who refine the fuel and feed it into the engine are the ones who are genuinely valuable.
3. Industry know-how has become the new moat
The 2026 enterprise AI trend confirms this: AI is moving from being able to chat and put on a demo to cross-system production execution β it is starting to answer customer service, manage inventory, reconcile accounts and write reports for companies.
Once AI has to do work, it can no longer rely on inventing things. What it needs is your price list, your inventory numbers, your customer ledgers, your compliance boundaries. And those are exactly the industry know-how a company has spent twenty years accumulating and outsiders cannot copy.
Local private deployment and tiered governance have become hard requirements in high-compliance settings such as finance and healthcare for the same underlying reason: turning a company's data β safely and within boundaries β into real material an AI can use.
Note that the precondition has changed. In the past, feeding industry knowledge to AI was hard because it was expensive: expensive calls, expensive trial and error. Today compute is nearly free, and the real cost has become the work of organising knowledge itself β how to turn the pricing logic in the owner's head into an industry handbook that an AI can read, answer from accurately, and never step outside of.
4. Whose hands do the gains fall into? Not coders
This is the core judgement I want to make.
The people who should catch this li-era dividend are not the ones who can write code. There are plenty of capable coders, and the models themselves are doing their work for them.
What is genuinely scarce is people who understand an industry and are willing to bury AI in the ground, to make it land in concrete places.
In translation, that means businesses like these:
First, brokers and organisers of industry data. Help a tea shop turn scattered market prices, supply and quotes into a knowledge base an AI can call directly. The work is not hard, but every industry is a gold mine, and nobody can take your understanding of that industry away from you.
Second, industry-specific AI deployment services. Not selling a general-purpose chatbot, but customer service that knows this trade, an assistant that can check inventory, a clerk that can verify prices. Wire AI into a company's real business flow so that it genuinely works on people's behalf.
Third, gatekeepers of an industry's benchmark prices. Whoever holds the most accurate market anchor in an industry is the one who makes AI answer reliably in that industry β and that in itself can grow into a business.
See what these have in common? None of them is a model problem; all of them are industry problems. All of them rest on how deeply you know a trade, not on whether you can tune parameters.
5. Three straight words for anyone thinking of getting in
I am not selling anxiety and not hyping a trend. Just a few things learned from stumbling.
First, do not chase new models; take stock of the data in your hands first. What you can really protect has never been an API; it is the industry you understand better than others.
Second, go small, go deep, go concrete. Do not try to empower a whole industry from day one. Start with one tea shop, or one small factory, and do one thing solidly: answering the dozens of questions customers ask most, quickly and accurately. One small wedge that works is worth more than ten grand demos.
Third, sell reliability as a product. A bare large model makes things up; feed it until it does not invent, answers accurately and cites its sources, and that in itself is a moat. Whoever first turns AI in their own industry from able to chat into able to use holds the next ticket.
One more thing that is easily overlooked: the cost collapse is quietly lowering the bar for trial and error. In the past, when an SME wanted to try AI, the calling cost alone could eat a budget line; they could not afford to try, so they simply did not. Today the same call costs a few li, and a company can spend a little, run it and see whether it is worth investing in. What does that mean? It means that using AI to do work β something only big companies could afford in the past β is for the first time genuinely on the table for small shops, small factories and sole traders. Opportunities moving downstream are often more worth noticing than the opportunity itself: you do not need to compete with the giants on technology, you only need to get into the corner of one specific industry one step before they do.
After the compute collapse, only two scarce things remain: real data, and industry people who can turn that data into value.
Coders do not have those two things, and models cannot grow them on their own. They are, precisely, in your hands.