Many people think that the AI bubble is something that only exists today.

In fact, AI history has already seen a very typical capital frenzy.
That was the "expert system" bubble of the 1980s.
The backdrop of that AI boom was that after the first AI winter, the industry finally found a route that seemed commercially viable.
Early AI researchers once tried to directly build general intelligence, but directions like machine translation, general reasoning, and robotics consistently fell short of expectations, and funding and confidence cooled rapidly.
So by the 1980s, the AI industry shifted its approach:
Since general intelligence was too hard, let's first build vertical intelligence.
The experience of doctors, engineers, chemists, financial experts, and equipment maintenance specialists was broken down into rules and written into computers.
If A and B occur simultaneously, then judge it as C.
If a certain combination of symptoms appears, then recommend a certain diagnosis.
If there is a configuration conflict in an order, the system automatically fixes it.
That's the expert system.
It is not the large model trained on massive data like today, but a knowledge base manually organized, plus an inference engine.
At that time, it really wasn't just a concept.
DEC's XCON/R1 system was one of the most classic commercial cases. It helped DEC automatically configure complex minicomputer orders, reducing configuration errors, improving delivery efficiency, and later became a landmark case for the commercial success of expert systems.
Once this case emerged, market imagination was completely ignited.
Companies began to believe:
If an expert system can replace part of an engineer's judgment, can it also replace a doctor's judgment?
If it can configure computers, can it also configure factories, manage supply chains, make financial decisions, and perform legal reasoning?
So capital began to pour in.
Large enterprises started setting up AI departments; IBM, DEC, GE, GM and others were investing in expert systems.
Startups also emerged, working on knowledge engineering, expert system software, inference engines, and industry solutions.
More interestingly, the AI bubble quickly spread to hardware.
Because many AI software at the time used the Lisp language, which had high demands on computing resources and development environments, a batch of Lisp machine companies optimized for AI appeared.
Companies like Symbolics, Lisp Machines Inc., and Texas Instruments participated in this wave of specialized AI hardware.
This is very much like a historical reflection of today:
Back then, expert systems drove Lisp machines.
Today, large models drive GPUs, HBM, optical modules, switches, data centers, electricity, and liquid cooling.
In every round of AI boom, the first to make money are often not applications, but the shovel sellers.
But the problems were exposed precisely during large-scale deployment.
The core bottleneck of expert systems was knowledge acquisition.
Real expert knowledge is not a manual.
Much judgment comes from experience, intuition, boundary conditions, and years of trial and error.
Experts themselves may not be able to articulate all tacit knowledge clearly.
Even if they could, it's hard to write it completely into rules.
So companies found that building an expert system was much slower and more expensive than imagined.
The second problem was maintenance cost.
Business processes are not static.
Products change, customers change, regulations change, supply chains change, market environments change.
Every time the real world changes, the rule base must be updated accordingly.
With many rules, the system also faces conflicts, omissions, and overlaps.
In the end, many companies bought not an automatic money-making machine, but a rule maze that could never be fully fixed.
The third problem was brittleness.
Expert systems could perform well within the scope of covered rules.
But once encountering boundary cases, ambiguous information, or incomplete information, the system would easily fail.
It had no real common sense.
It did not actively learn.
It was also hard for it to understand changes in context like humans.
A demo could be impressive, but enterprise systems had to face a real world that changed every day.
This was the huge gap between the lab and the production environment.
The fourth problem was the collapse of hardware economics.
Lisp machines were initially AI infrastructure.
But by the late 1980s, general-purpose workstations and personal computers rapidly became more powerful, cheaper, and had larger ecosystems.
When cheaper general-purpose computers could also run the relevant software, the expensive specialized Lisp machines lost their commercial justification.
Thus, the Lisp machine market collapsed quickly.
The hardware chain was repriced first, then software companies and AI consulting firms also came under pressure.
A large number of expert system companies went bankrupt, were acquired, or pivoted, and AI once again became a word the capital market didn't want to hear.
This was the second AI winter.
But here is a very important detail:
Expert systems were not completely useless.
They didn't disappear; they were absorbed into enterprise software, rule engines, risk control systems, knowledge management systems, customer service scripts, and process automation systems.
The technology survived.
The bubble died.
This is exactly what today's AI investors should chew on repeatedly.
Today's large models are certainly not expert systems.
LLMs are not hand-crafted rule bases; they come from large-scale data, neural networks, Transformers, computing power, and reinforcement learning.
Their generality, language ability, code ability, and multimodal ability far exceed the expert systems of that era.
So it's not rigorous to simply equate today's large models with expert systems.
But where history is truly similar is not the technical route, but the psychological structure of the capital market.
Every round of AI boom goes through a similar three-step process:
Step one: a real technological breakthrough occurs.
Step two: capital believes it can transform all industries.
Step three: companies discover that turning the technology into a stable, controllable, auditable, and profitable system is much harder than imagined.
Today's large models are also entering step three.
On the C-end, there is usage, but the traffic landscape hasn't been completely rewritten.
On the B-end, there are pilot projects, but many remain at the pilot stage and haven't moved into truly large-scale production systems.
Agents are attractive, but in long-process tasks, single-step errors accumulate.
Even if single-step accuracy seems high, as long as the process is long enough, the overall success rate drops significantly.
Fields like industry, finance, healthcare, law, and supply chain do not just want a nice answer; they want end-to-end reliability, exception handling, permission control, accountability, audit trails, and human fallback.
These cannot be solved by a model launch event.
So the biggest risk for AI today is not that the technology has no value.
On the contrary, AI is very valuable.
The real question is:
Is the cash flow it generates sufficient to support the current wave of capital expenditure and valuations?
If enterprises find that AI ROI is still vague, the speed of budget cuts could be very fast.
If cloud providers find that inference revenue cannot cover depreciation, power costs, and data center costs, capital expenditure guidance may be revised downward.
If the application layer does not produce a sufficiently strong paid closed loop, the hardware chain will start to price in downward revisions of demand expectations.
History does not simply repeat.
But the capital market often pushes a real technological breakthrough into overly high profit expectations in a similar way.
The lesson of the 1980s is not that "AI is a scam."
The real lesson is:
Technological revolutions can be real.
The pace of commercialization can be slow.
Infrastructure investment can be ahead of its time.
Stock valuations can be wrong.
These four things can all be true at the same time.
Expert systems did not disappear in the end, but the expert system bubble burst.
Today's large models likely won't disappear either.
The real question is:
In this round of AI infrastructure frenzy, which companies will become the infrastructure of the future, and which companies are just the Lisp machines of this cycle?