What are the scenarios for the collapse of the generative AI bubble? - Predicting collapse scenarios from the industrial structure.

Last time, we analyzed the industrial structure of generative AI and whether the generative AI market is a bubble.

Our analysis concluded that the generative AI market is in a bubble, exceeding the typical mid-stage of a bubble compared to past major bubbles. However,Unfortunately, the bubble economy has historically been weathered without any significant damage.It has never existed in the past.

So, in this article, let's consider the possible scenarios and timing of the collapse of the generative AI bubble from an industrial perspective.

First, we will analyze the general mechanism of bubble collapse by examining historical bubbles.

The mechanism of bubble collapse

A bubble is defined as follows:

To put it more simply, it's a situation where expectations that real estate, stocks, and other assets will increase in value in the future grow, leading to transactions at a value (price, etc.) significantly higher than their actual intrinsic value (price, etc.) at that time.

In other words, it's safe to say that the source of a bubble's expansion is expectations for the future.

Considering the mechanism of a bubble collapse, it begins when expectations for the future crumble. In other words, the trigger is when it becomes clear that the price that was expected to rise in the future is unlikely to be reached.

This can be expressed using a simple inequality as follows:

If this inequality holds true, the bubble will burst all at once.

The actual mechanism of historical bubble collapses is as follows:

In both patterns, the bubble inflates due to expectations for the future, and bursts when it becomes clear that reality does not live up to those expectations.

The most recent bubble collapse, which many people will likely remember, was the Lehman Shock, triggered by the subprime mortgage crisis.

To put it simply, the expectation that fueled the subprime mortgage bubble was that "the population in America will inevitably increase, leading to an increase in housing and consequently, a rise in housing prices."

Fueled by this expectation, loans were forcibly extended to people with low loan repayment capacity (the subprime class) to cover housing costs (subprime mortgages). The collateral for these loans was the belief that "housing prices will definitely rise in the future" (similar to a malicious residual value credit scheme). What was even worse was that subprime mortgages had variable interest rates (loans linked to the policy interest rate).

Those subprime mortgage bonds were sold worldwide as a solid financial product.

However, the expectations that fueled the bubble were thwarted by the oversupply of housing, which stifled the rise in housing prices. To make matters worse, the Federal Reserve (FRB) gradually raised its policy interest rate from 1% to 5.2%.

This will cause some people to default on their mortgages (subprime loans). The foreclosed properties resulting from these loan defaults will enter the market, further driving down housing prices. Subprime loans were originally secured by rising housing prices, so the decline in housing prices will lead to a large number of loan defaults.

This is when subprime mortgage bonds became worthless, and the bubble burst.

Thus, in this recent bubble, just like in past history, the expectations (increase in housing and rising prices in the US) were shattered, causing the bubble to collapse.

Next, let's consider what expectations will be the source of the bubble in the generative AI market.

What is the source (expectation) of the generative AI bubble?

First, as a recap from last time, both Open AI (chatgpt) and Anthropic (claude), leading generative AI companies, are projected to have significant losses in 2025.
*As of February 2026, the date of writing, none of the generative AIs have ever been profitable since their establishment. Currently, due to the industry structure of generative AIs, they are inevitably operating at a loss.

In other words, the actual revenue projected for 2025, which is the subscription and API fees for generative AI, is far from being profitable (a significant discrepancy from reality). In fact, both OpenAI and Anthropic have a $12 deficit hidden behind the $20 monthly subscription fee for their entry-level plans.
*The recently introduced introductory plan is likely to be even more unprofitable, but conversely, it's a good deal.

Despite operating at a loss, further large-scale investments are being made.

These funds are being injected based on the expectation that "generative AI will generate enormous profits in the future."

However, as explained previously, it is quite difficult to make a profit with the current industry structure and revenue models (subscriptions and API fees) for generating AI.
* Achieving profitability is virtually impossible under the current industrial structure. The only way to make profitability possible with the current structure is to revolutionize the computational resources required for inference models.

In that case, investors' expectations will be for some kind of change (paradigm shift) that will generate enormous profits. We will try to decipher that change from the statements of the CEOs of various AI generation companies.

These statements clearly show that they are not simply extending the current subscription/API fee model, but rather intend to integrate it into social infrastructure.

Investors have also made statements that support this, such as:

Considering the common threads in these statements, it's safe to say that the expectation that generative AI will become a vital part of social infrastructure like electricity, gas, water, trains, and roads is the source of the bubble.

In other words, The true nature of the expectations that fuel the AI ​​bubble is "social infrastructure development."It's almost certainly correct to conclude that this is the case.

While it's unclear from statements by CEOs and investors of various AI generation companies what this "social infrastructure" entails, at a minimum, the following conditions are essential for it to be used as infrastructure.

Therefore, the starting point of the scenario for the collapse of the generative AI bubble, which is the subject of this article, isWhen social infrastructure development cannot be achieved, or when social infrastructure development cannot be achieved at a speed that meets expectations.Historical facts reveal that this occurred in that time.

Next, we will explore the starting point of the bubble's collapse by considering the challenges of "making generative AI a social infrastructure."

Based on the previous diagram, we will consider the challenges to achieving social infrastructure status from two perspectives: hardware aspects such as AI computing chips like GPUs and power, and software aspects such as generation AI software and training data.

Hardware challenges (AI computing chips, power, etc.)

In terms of hardware, the three biggest factors influencing the integration of generative AI into social infrastructure are reliability, cost, and scalability. To put it simply, it is essential to have data centers that can always operate reliably, where the usage fees for generative AI are affordable for everyone, and that can withstand a massive number of users without failing.

Therefore, the minimum requirements for a data center are the following capabilities.

Based on these points, regarding data centersThe major bottlenecks are the stable supply of AI computing chips and the stable supply of power.You will see that these two are the key.
*There are many new data center construction projects planned, so I don't think this will become a bottleneck.

Next, let's take a closer look at each of the challenges.

Challenges in ensuring a stable supply of AI computing chips (such as GPUs)

First, let's get a sense of the volume of demand for AI computing chips.

Data centers used for generative AI typically have around 1000 computing racks in a standard-sized facility, and around 5000 in a large-scale data center. A standard setup uses about 8 AI computing chips per computing rack.

Therefore, each data center requires a large number of AI computing chips, approximately 8000 to 40,000.

As of March 2025, there are 5426 of these data centers in the United States alone. Considering that, the scale alone would be in the tens of millions to hundreds of millions of chips. If we extend this to a global scale, although we don't know the exact number, it wouldn't be surprising if the number of AI computing chips required were in the range of over a billion.

Up to this point, we've only calculated the total number of AI computing chips derived from the number of existing facilities, so it doesn't mean there's a demand of over a billion chips every year. However, there is a huge demand every year for new data centers, replacements due to malfunctions and other problems (they actually break down quite often), and replacements due to performance differences between generations as chips evolve rapidly.

In fact, the total number of AI server computing units shipped in 2025 is projected to be between 1000 and 1200 million units, each equipped with 4 to 8 AI computing chips.

The company that is meeting this enormous demand for AI computing chips is NVIDIA, which is said to have an 80-90% market share. And almost all of the production of NVIDIA's AI computing chips (GPUs for AI calculations) is handled by a single company, TSMC in Taiwan.

Naturally, the current situation does not fully meet the demand, which exceeds 1000 million units per year, and there will be a complete supply shortage from 2025 to 2026.

From thisThe economic structure makes it difficult for GPUs used for AI computing to see a significant price decrease.Therefore, reducing the computational cost of AI and expanding its computing power are major challenges in order to make it a part of social infrastructure.

Challenges in ensuring a stable power supply

Just like with AI computing chips, let's start by getting a grasp of the overall electricity demand. You can probably intuitively understand that even just running the billions of AI computing chips mentioned above would require an enormous amount of electricity.

Taking the NVIDIA H100, a standard GPU that appeared in 2022-2023, as an example, the power consumption per card is approximately 700W, so roughly calculating for over a billion cards, it comes to 70GW.
*Newer chips consume more power; the latest and most talked-about chip, Blackwell (B200), consumes 1000-1200W.

This 70 GW represents approximately 25% of Japan's total power generation capacity, which is 300 GW.

Of course, data centers consume power for cooling, communication, and other purposes in addition to AI calculations, so the required power will exceed 70 GW.

According to the International Energy Agency (IEA), data center electricity consumption is projected to exceed 1,000 TWh (terawatt-hours) by 2026, a figure comparable to the total annual electricity consumption of Germany, a leading industrial nation in the EU. Moreover, this figure represents a doubling of pre-AI levels in just a few years.

Because of this, companies operating AI-generating systems are desperately trying to secure a stable power supply.

Thanks to the desperate efforts of each company, the current generation AI is somehow managing to function.

Despite these power supply issues, plans for new massive data centers, such as the Stargate project, Anthropic's own massive data centers, and cloud expansion plans from companies like Microsoft, Google, and Amazon are steadily progressing.

Ultimately, electricity demand far exceeds supply, so, similar to AI computing chips, the economic structure is such thatLower electricity prices and a stable supply are not progressing easily.Therefore, the requirements for making AI a social infrastructure, such as ensuring it always works reliably, reducing the computational cost of AI, and handling a massive number of users, present significant challenges.

Could hardware challenges be the trigger for a bubble collapse?

Based on the explanation so far, regarding whether the problems of AI computing chips and stable power supply, which are representative challenges in the development of hardware-based social infrastructure, will be the starting point for the collapse of the bubble economy, in my opinion, it will not be.

Let me give you a few reasons.

Regarding the issue of a stable supply of AI computing chips, various AI companies have already begun taking steps to address it, and signs of improvement are starting to appear.

The actions of various companies could potentially break the current supply system, which is heavily reliant on NVIDIA.

Furthermore, NVIDIA is not content with the current situation and is looking to secure new production bases to replace TSMC (considering utilizing Intel-affiliated companies), and TSMC itself has announced plans to increase production and expand its number of production bases.

Although there are various challenges, we are making strong progress in developing next-generation AI computing chips that will be overwhelmingly more powerful than current ones.

Similarly, while addressing the power supply problem will take time, efforts are being made to overcome it by restarting suspended nuclear power plants and building new gas-fired power plants. In the worst-case scenario, there is also the option of generating electricity from private gas sources, although this would lead to higher costs.
* Oracle is considering this.

Furthermore, the development of SMRs (next-generation small nuclear power generation units) is steadily progressing, although there are various challenges.
*The underlying technology for SMRs is based on nuclear submarines and nuclear-powered aircraft carriers, so there is a solid foundation of pure technical expertise. The challenges for civilian applications lie in cost, safety, and legal aspects.

None of these challenges will make the integration of AI into social infrastructure impossible; they will only delay the timing. As mentioned in Chapter 1, the speed and timing of the integration into social infrastructure are also important for the bubble economy,If there are clear, logically and physically feasible solutions, it is unlikely to betray the expectations of "social infrastructure development," which is the source of the bubble.I think.

Conversely, the bubble may inflate further to accelerate the transition of AI into a social infrastructure. This issue can be controlled to some extent by investing money, so it can also be used as a pretext to raise funds. In other words, I believe it will be a factor that intensifies competition rather than the trigger for a bubble collapse.
*The success or failure in the industry may depend on the AI ​​computing chips and the ability to secure a stable power supply.

Challenges on the software side (generative AI software, training data)

Here, let's consider the challenges that the software aspects will face in making AI a social infrastructure. Similar to the hardware aspects, when we organize the relationship with the four requirements for AI to become social infrastructure, we find that the software aspects have an even wider impact than the hardware.

To make the above issues more specific, they can be rephrased as follows:

As can be seen from the diagram above, the key software challenges for social infrastructure development can be narrowed down to two items: improving the inference model and providing high-quality, large-scale training data.
*I don't think improving the generation AI app is particularly technically difficult, so I'll omit that part.

In other words, the essential conditions for generative AI to become a part of social infrastructure are a significant reduction in the weight of inference models and ensuring that high-quality training data is not depleted.

Next, I will delve deeper into these two points, but since I come from a background in mechanical design (engine design) and do not have deep knowledge of AI programming or algorithms, I will focus on the problem of depletion of training data.

Training data depletion problem

To consider this problem, let's first think about what happens when the training data is completely depleted. As a premise, current generative AIs operate probabilistically, so it is structurally impossible to eliminate hallucination (the generation of misinformation).
*Current AI generation algorithms are based on probability theory, so a 0% hallucination rate is absolutely impossible.

Imagine a world where training data has been completely depleted. In that world, a generating AI answers questions from users. A certain percentage of these answers will contain hallucinations. These hallucinations are accumulated on the internet, and the AI ​​learns from the information containing these hallucinations, using that information to generate answers for new users. In other words, the more it is used, the more it quotes from hallucinations.
*Uses the exact same mechanism as the much-talked-about echo chamber.

Eventually, hallucination will be amplified unilaterally, resulting in a generated AI that is completely useless. This will inevitably occur regardless of whether the hallucination rate is 5%, 1%, or 0.1%.
*The more users (number of answers) there are, the more exponentially the effect will increase.

When this state occurs, no matter how advanced the inference algorithm becomes, it becomes useless. This is called model collapse.

Incidentally, in traditional information exchanges between humans, both the information sender and receiver judge the true intent, which somewhat suppresses the amplification of lies. However, since AI cannot judge true intent, this kind of thing happens.
*This is a major reason why conspiracy theories and similar ideas tend to remain confined to certain circles.

As can be seen from this, the generation AIHuman judgment and information provision, such as "That's wrong, the real information is this."This will definitely be necessary (information feedback).
* Information gathered through a simple chat is overwhelmingly insufficient.

In other words, a certain percentage of information generated by humans is absolutely necessary for generating AI.
*Personally, I think a 50% ratio is the minimum acceptable level.

While these are extreme examples, let's consider what is realistically likely to happen.

As of the end of February 2026, the use of generative AI is expanding at an accelerating pace and is being used in the creation of all kinds of content.
*This website's content also utilizes AI. We take the utmost care to avoid misinformation and hallucinations.

This extends beyond just website content to include images, music, and videos. A concrete example of videos generated entirely by AI is many of YouTube's short videos, which you are probably already familiar with.

In this way, AI-generated content is gradually permeating all kinds of content. As a result, hallucination inevitably gets included, even unintentionally, and the reality is that hallucination is gradually increasing.
*This will depend on the morals and abilities of the content creator.

Even if the content itself isn't completely replaced, the amount of content generated by humans is decreasing as the amount of AI-generated content increases.

Adding to the problem are the AI ​​modes and AI summarization features of search engines, such as Google's. These summarization features are reducing the number of views of content created by humans. As a result, the incentives for content creators (such as revenue) decrease, which leads to a decline in the number of content creators.
*Zero-click problem

This phenomenon is not limited to websites; even on YouTube, a leading video-sharing site, the rise of short videos, music videos, and short animated videos created entirely by AI is leading to a decrease in the number of content creators.
*Voice actors, actors, and musicians jointly issue a statement regarding the unauthorized use of their voices by AI.

If this situation continues, eventuallyHuman information generation amount < Information generation amount by AIIt's only a matter of time before this happens, and a noticeable decline in the performance of generating AI will inevitably occur in the not-too-distant future.
*This cannot be covered in principle by improving the inference model.

Leaving it as is will lead to the model breakdown explained at the beginning of this chapter.

As a specific timeframe issue, it's not a major topic in the media as of the end of February 2026, when this was written, but according to predictions by the AI ​​research institute Epoch AI, high-quality text data is estimated to be depleted between 2026 and 2028.

They estimate the total amount of publicly available human-generated text data to be approximately 300 trillion tokens, and they predict that this inventory will be depleted if the current AI learning trend continues. Once this training data is exhausted, as explained above, performance improvements will decline and eventually stop. Ultimately, they predict that hallucination will increase, leading to model collapse.

Unfortunately, this is an unavoidable problem no matter how much the inference model is improved, making it a very difficult issue.

This becomes a challenge when training data is depleted.

Could software-related issues (the problem of depletion of learning information) be the trigger for the collapse of the bubble economy?

Based on the explanation so far, if you ask whether the depletion of learning information, a representative challenge in the development of social infrastructure on the software side, will be the starting point for the collapse of the bubble, my answer is a resounding yes, it will be the absolute starting point.

Now, I will explain the reasons.

First, as mentioned above, the amount of information generated by humans is definitely decreasing as of 2026. Furthermore, one specific event that will accelerate this decrease is the introduction of Google's Search AI mode and AI Overview feature. This has caused website click-through rates to drop by about 30-60%. As a result of this event, it is almost certain that the number of creators who have lost incentives (such as revenue) to operate websites will decrease.
*This is a very serious problem in the web industry, and it is highly likely to occur in video content such as YouTube in the near future.

As a result of this decrease in incentives (such as reduced revenue), major news sites and others have begun to take concrete measures, such as making official announcements prohibiting learning.

In another development, a lawsuit is underway between the New York Times and OpenAI regarding unauthorized learning by AI.

Initially, the case was filed by only one newspaper, the New York Times, but since then, it has escalated into 51 separate lawsuits with almost identical content.

This is a historic trial for AI, and the verdict is expected around the summer of 2026. Depending on the outcome of the verdict, it may become impossible for AI to learn from websites without permission (individual licensing fees may be required).
*For some reason, there is little media coverage of this in Japan.

Adding insult to injury is the implementation of the EU AI Act.

This law will come into effect on August 2, 2026. Perhaps as a result of this, even I, who am in Japan, can configure my server (Xserver) to reject AI learning.
*I haven't refused yet, but I'll have to think about it in the future.

This trend is highly likely to spread worldwide, resulting in a significant reduction in the high-quality training data essential for generative AI.

On the other hand, the harsh reality is that there are currently almost no effective solutions for companies involved in generative AI that address the depletion of training data.

The few measures taken so far are limited to a major overhaul of Google's search algorithm (eliminating low-quality content through AI generation) and monetary agreements with several major websites.
*Google search results in a significant decrease in click-through rates, making it largely ineffective.

These contracts are exclusively with the largest companies, which is nowhere near enough to meet the enormous amount of training data required for generating AI; it's like a drop in the ocean.

One proposed solution involves using AI to synthesize existing information and then training the AI ​​with it. However, this approach is largely pointless because AI is incapable of determining the truthfulness of the synthesized data in the first place.
* Is it even possible for humans to manually check every single instance of this massive amount of synthesized data? And how will hallucination be handled when creating this synthesized data in the first place?

While this drastic trend of reducing training data is accelerating significantly, there are currently no fundamental countermeasures or plans in place.

The fundamental problem with the depletion of training data is that, currently, there is no connection between AI-related companies and investors and the diverse creators who generate the information. Therefore, no matter how much investment is raised, this problem cannot be solved.

Therefore, if the current situation continues, the training data will inevitably be depleted, resulting in a decrease or cessation of performance improvement, and eventually leading to the model's collapse.

Currently, there are no logical and effective measures to resolve this situation, so the source of the bubble isChallenges that will surely shatter expectations of "social infrastructure"Therefore, I believe this will be the starting point for the bubble's collapse.

Predictions for the collapse scenario of the generative AI bubble

Here, using the considerations discussed so far, we will try to predict a bubble collapse scenario in chronological order.

First of all, as of the end of February 2026, the generative AI market is booming and thriving.

From a technical standpoint, autonomous AI agents are beginning to become commonplace. Generative AI-related companies are also pushing forward with the construction of massive data centers, securing power sources, developing power grids, and increasing production of new AI computing chips, all aimed at making AI a vital part of social infrastructure. Investors are also actively involved, with a steady stream of large-scale investments being announced.
*The AI ​​agent, by its very nature, is moving in a direction that will accelerate the increase in hallucination.

The EU AI ACT law, which prohibits unauthorized learning of information, will come into effect around the summer of 2026. Furthermore, a ruling in a US lawsuit concerning unauthorized AI learning may be issued as soon as possible.

This will drastically reduce the amount of new training data generated by humans. It's possible that the main source of AI training data will shift to synthetic data. At this point, the issue of training data depletion may start to be discussed in the media.

At the end of 2026, the massive data centers of the Stargate Project and Anthropic's own massive data centers are scheduled to begin operations.
*The start of operations may be delayed.

Despite improvements in inference models and the operation of powerful data centers, I predict that performance improvements will slow down in the first half of 2027. This may lead to the issue of depleted training data becoming a serious concern in the media.

In the second half of 2027, ordinary users will begin to express concerns that "performance improvements have completely stopped." At this point, the intensification of competition in generative AI means that learning cannot be stopped, and the amplification of hallucination can no longer be stopped.

As predicted by AI research institute Epoch AI, high-quality training data will be completely depleted in early 2028, and a significant slowdown in the improvement of AI performance will become apparent. In fact, even if inference models are improved, performance improvements will hardly be seen in various benchmark tests.
*Videos, images, and audio are not very suitable learning materials for generating AI, in addition to being difficult and costly to learn.

HereIs it possible that the integration of generative AI into social infrastructure is impossible?"Doubts will begin to arise, and I think some investors will start to take their investments back. Some users may also start to leave (slowing down of user growth)."

Around the summer of 2028, many investors will enter recovery mode, and assets that have not yet recovered their investments will begin to turn into non-performing loans.

I believe there is a high probability that a wave of bankruptcies and acquisitions of AI-related companies will begin in the latter half of 2028 and the beginning of 2029, causing AI-related financial products scattered around the world to crash (including for individual investors, of course).
*It's possible that the current giant in generative AI might go bankrupt.

The bubble will burst in early 2029, tentatively named the "AI shock."

I have a feeling things will unfold in this way. Coincidentally, the prediction of the bubble burst falling in early 2029 coincides with the projected year for profitability for each AI generation company. Well, each AI generation company's projected profitability year is probably set near the limit of what investors are willing to wait, so this might give the prediction a little more credibility.

Since we're at it, let's illustrate this in chronological order.

What do you think? It's probably difficult to pinpoint the exact timing of each event, but the overall flow seems reasonable.

What I personally fear most about this AI bubble burst is the sheer scale of investment compared to the past. Investment is projected to reach $1.5 trillion in 2025 and $2.5 trillion in 2026, already a massive amount exceeding any previous bubble. My prediction is that the bubble will burst in early 2029, so the amount will have multiplied many times over in the next three years.

I experienced the subprime mortgage crisis firsthand as a working adult, and it was a terrible situation with a huge impact. So, the idea that the AI-related bubble in 2025 could be about four times the size of the subprime crisis is quite frightening.

I don't want my prediction to come true, but what do you all think?

Predictions for the world after the collapse of the generative AI bubble

Since we're on the subject, let's take a moment to consider the environment surrounding generative AI after the bubble burst.

First, let's consider what will remain after the bubble bursts.

In this state, virtually no external funds, that is, funds from investors, are being injected.

In that case, it is likely that what remains after the bubble burst will be in the following state.

In this situation, it's highly probable that the capabilities of generative AI won't improve significantly, and generative AI-related companies will be offered to the market at a price that keeps them from incurring losses.
*Research and development of generative AI should focus on reducing inference costs.

I expect the price to be about the same as now, or slightly higher. This is because, even though data center usage fees and electricity prices are falling, there are still minimum operating costs to cover, and we absolutely cannot afford to operate at a loss, so we cannot lower prices arbitrarily.

Even minimal performance improvements would be economically impossible until the large surplus of AI computing chips is completely utilized.
*Development of next-generation AI computing chips may also be frozen, or at best, significantly delayed.

If that happens, we can hardly expect the kind of explosive increase in users we see now, and I think it will end up being used only by people who need it.

Therefore, I believe that each AI-generating company will focus on developing services for corporate clients as a new source of profit. The services will likely not be dream AI that can do anything, but rather support services that assist with daily operations.

I suspect that the learning models for generative AI will also evolve from learning from all kinds of information worldwide to becoming AI models specifically tailored to each contracted company, learning only from the information of that company.
*Security technology is key.

In other words, it seems that the main source of revenue will be the customized implementation, maintenance, and management tailored to each individual company. Essentially, it's exactly the same as IT tools used in typical businesses today.
*In the author's field, CAE software (computer simulation) followed exactly the same path.

Ultimately, I think it will settle down to a situation where only a select few ordinary people use it, and its primary use is for business purposes.

In other words, While it may not become the "social infrastructure used by everyone" that current AI companies and investors envision, it will achieve "social infrastructure status" as an indispensable tool for our economic activities.I think so.
*Like Microsoft Office, this is a tool that is always used at work, but not often used at home.

As I was writing this, it occurred to me that the whole world seems to be experiencing a massive frenzy, mirroring the Windows 95 fever that occurred in Japan.
*While Windows 95 sold explosively, in reality, many were machines used solely for playing Minesweeper. It did contribute to the spread of PCs in Japan.

Summary

The conclusion will be illustrated with a diagram.

I believe the bubble will burst according to the structure and scenario shown in the diagram above. Personally, I don't want my scenario to come true because the damage from a massive bubble bursting will be immense, but sadly, I'm strangely confident about it.

To reiterate, looking back at past bubbles, there are very few examples of bubbles that didn't burst, so perhaps this is inevitable. This current bubble, in particular, bears a striking resemblance to the railroad bubble of the 1840s and the IT bubble of the late 1990s.

Coincidentally, my collapse scenario ended up being almost identical to the collapse of the railway bubble and the IT bubble. It may be human nature to create a bubble when an extremely useful technology comes to light, as the media and others tend to make an excessive fuss about it.

I suspect that this generational AI bubble, like the railway and IT bubbles, will only truly become a social infrastructure once it bursts by utilizing the tangible assets that remain afterward.
*The railway bubble involved vast railway lines and stations, while the IT bubble involved fiber optic networks and data centers.

The only effective measure I can think of to stop this AI-generated bubble from collapsing is to limit the use of AI-generated content. It seems like the only way to prevent this is to keep the amount of misinformation generated below a certain level. Alternatively, I think we need to control the balance between the amount of content generated by humans and the amount generated by AI-generated content.
*In short, measures to balance the amount of content generated are absolutely essential.

However, in the highly competitive environment created by the bubble, measures such as restricting usage are absolutely impossible. Just declaring restrictions on usage would likely cause the bubble to burst.

Ultimately, for ordinary citizens like myself, if we're going to invest in AI-related things, we should be careful about the timing of the bubble burst. As users, it might be best to make the most of it while it's still a good deal during the bubble.

That concludes our discussion on the collapse scenario of the generative AI bubble: predicting collapse scenarios from the perspective of industrial structure.

P.S
Personally, I think the indicator of an impending bubble collapse is the improvement in the performance of AI generators. However, I don't think accurate data on the extent of performance improvements will ever be made public, so your gut feeling from using them might be surprisingly accurate.
* You probably won't get a good sense of it from benchmarks.

If you notice an increase in hallucination after using it, or if you feel that your condition is somehow worse than before, it might be a sign of something.