It Has Been Three Years — Why Haven't AI Applications Exploded Yet?
Recently Meta circulated a scary story about compute overcapacity and the AI stock market dropped in response. Everyone is anxious.
People keep asking: can this AI game keep going? When will AI applications finally explode? When will the "AI version of TikTok, WeChat or Xiaohongshu" show up? Is AI about to repeat the dot-com bubble of 2000?
Here is my prediction: within half a year, the AI infrastructure bubble will likely burst; within a year, enterprise AI applications will start to take off; only two years out will consumer AI applications truly arrive.
Here is why.
Today's AI looks a lot like the personal computer of forty years ago
Current AI is genuinely strong at chatting and assisting with work, but it is still far from the "open it and use it, get hooked instantly" experience ordinary people imagine. Ask it to write your weekly report or plan a trip, and your first reaction is usually not "how convenient" but "how am I supposed to phrase this?"
Early computers were exactly the same. They were clumsy bricks — no games, no short video, no social networks, and in the earliest days not even a graphical interface. Users stared at a black box with a blinking cursor, typed commands themselves, learned the operations themselves, and swallowed bizarre error messages themselves.
In 1981 IBM launched the first PC at 1,565 dollars — roughly five thousand dollars in today's money. Ask an ordinary consumer whether they want one and the answer is probably "have you lost your mind?"
So why did anyone pay for that black box at all?
Businesses did.
Enterprises bought the first wave — consumers came years later
Companies did not need the computer to be fun. They needed it to get work done. Accounting, inventory management, order tracking — all of it used to run on paper and people, which was slow and error-prone. However awkward the machine was, if it raised efficiency, cut mistakes and saved labour, the enterprise would pay.
In its first year the IBM PC sold 250,000 units, and the vast majority of buyers were companies. Throughout the 1980s, PC adoption grew almost entirely on enterprise purchasing.
Only after hardware became cheap enough and the software ecosystem matured did the home market rise in the 1990s — first with games, then with the internet.
People hear "the internet" today and immediately think of Google, Facebook, YouTube or Taobao. None of those existed at the start. What the early internet genuinely changed was work that was plain to the point of boredom: corporate email, internal messaging, online forms. The first companies that made real money online were doing unglamorous B2B work — helping businesses manage customers, orders and collaboration.
Enterprise demand carried the installed base first. Only once that base and the user habits existed did consumer applications grow on top of it.
We may be waiting in the wrong place
Three years in, we have ChatGPT, Claude, Codex, all kinds of agents and generation tools, yet ordinary people still have not seen many "must-use" new applications. Everyone is waiting for a consumer super-app — something people open every day the way they open WeChat or TikTok.
But the first explosion is probably happening in the office, not on your phone. On the B-side, not the C-side.
Why does the enterprise scenario mature first? Because companies calculate ROI. A customer-service agent costs over a hundred thousand a year in salary. If AI removes 30% of that workload, the company saves tens of thousands annually — once the math works, they buy. Consumers do not think that way. Free is best, and even a twenty-yuan subscription gets agonised over. The gap in willingness to pay decides which side matures first.
Enterprise AI will spread the way enterprise IT always has: starting at the edge of the process (knowledge-base Q&A, meeting minutes, coding assistants), then moving into core workflows (ticket handling, contract review, financial reconciliation), and only finally cross-system agents that run an entire process on a person's behalf.
Consumer AI will come — after enterprises flatten the road
Consumer applications have three structural disadvantages:
Weak willingness to pay. Individuals do not hand over money as readily as companies.
Tasks are too fragmented. Enterprise processes, however complex, have fixed goals and fixed systems. Personal life differs from person to person, and one standardised product can hardly cover it.
Habits are not formed yet. Most people are still figuring out how to ask an AI anything. Enterprises can push adoption through training and policy; consumers can only be won over when a product becomes good enough to use naturally.
Only once enterprise demand drives model costs down and reliability up can a personal AI assistant turn from "a fun toy" into "a daily gateway you cannot live without."
The real first step of a technology revolution is saving money, not addiction
So back to the opening question: three years on, why hasn't the AI application explosion arrived?
Because we have been staring at the phone home screen while the real first wave is happening quietly inside offices. Code repositories, customer-service consoles, ticketing systems, financial workflows — none of that trends on social media, but it has budgets, ROI and cost-cutting pressure, and someone willing to pay for results.
The first move of any genuine technology revolution is never to make people addicted. It is to make people save money.
FAQ
Why will enterprise AI applications arrive before consumer ones?
Businesses buy on ROI: if AI cuts a measurable share of paid workload, the purchase is justified. Consumers rarely have that discipline, personal tasks are too fragmented to standardise, and prompting habits have not formed yet. The same enterprise-first sequence played out with the IBM PC and the early internet.
Is AI just a bubble like the dot-com crash?
The infrastructure layer can absolutely over-invest and correct — that is the part likely to shake out first. But the application pattern from every prior platform shift is that real revenue arrives on the enterprise side before consumer breakout apps appear.
What is the realistic adoption order for enterprise AI?
Edge workflows first — knowledge-base Q&A, meeting minutes, coding assistants. Then core processes such as ticket handling, contract review and financial reconciliation. Cross-system agents that run an end-to-end process come last.
