The AI boom appears to have already reached a turning point.
Until now, when the term “AI bubble” was used, it often reflected a somewhat superficial skepticism. Questions included: Is AI really useful? Isn’t generative AI just a passing fad? Isn’t it just a chatbot?
However, the issues now coming to light are slightly different. It is no longer reasonable to doubt that AI technology itself is genuine. Generative AI, large language models (LLMs), image generation, code generation, speech recognition, and agent technology have already begun to permeate many aspects of work and daily life. The question is not whether AI is useful.
The question is whether the massive investments being poured into AI will generate the returns investors expect.
In other words, it is not AI technology that is at risk of collapsing, but the AI investment bubble. An inconvenient truth is coming to light: excessive AI investment is structurally unprofitable.
1. The success of AI technology and the success of AI investment are two separate issues.
First and foremost, we must recognize that technological innovation and return on investment are two separate things.
The Internet was a real innvation. However, many companies disappeared during the dot-com bubble. Solar power was also a real innovation. However, many solar panel manufacturers got caught up in price wars. LCD panels were also a real phenomenon. However, Japanese companies struggled to compete with South Korean, Taiwanese, and Chinese rivals after making massive investments. The widespread adoption of a technology by society is not the same as companies that invested heavily in that technology early on reaping sufficient profits.
The same thing could happen with AI. AI will become deeply embedded in society. The use of AI will increase.
AI will transform work and daily life. Nevertheless, much of the capital invested in AI infrastructure, AI models, AI startups, and AI-related stocks may fail to yield the expected returns. This is the basic scenario for the collapse of the AI bubble that is now coming into view.
2. Data center investments are too costly
At the heart of the AI boom lies not software, but massive physical infrastructure: GPUs, servers, data centers, electricity, cooling systems, power grids, land, and water resources. Behind generative AI lies an enormous amount of capital investment in infrastructure.
This differs from traditional SaaS. With software, once it’s built, marginal costs can be kept relatively low. However, LLM inference consumes computational resources every time it is used. As usage increases, more servers, electricity, and cooling are required.
In other words, while AI companies may appear to be software companies at first glance, in reality, they share many characteristics with infrastructure companies. Furthermore, data centers cannot be completed overnight. Construction takes time, and power contracts and regulatory compliance are required. Just because there is demand does not mean supply can be increased starting next month. There is a significant time lag here.
Investment decisions are being made based on the current frenzy. The facilities will likely be completed in a few years. However, by that time, market prices and the competitive landscape may have changed. This is the primary structural risk behind the collapse of the AI bubble.
3. LLM prices are falling rapidly
The second risk is the decline in LLM prices. Currently, the cost of using LLMs is falling rapidly. Thanks to model optimization, improvements in inference technology, advancements in semiconductor performance, intensifying competition, and the widespread adoption of open models, the cost of achieving the same level of performance continues to drop.
This is good news for users. It is also desirable for companies if they can use AI more affordably. However, from the perspective of those who have made massive investments in AI infrastructure, this poses a serious problem.
If a data center was built based on the assumption of high usage rates, but by the time it is completed, LLM usage fees have dropped significantly, the facility may be operational but unable to generate the revenue needed to recoup the investment. In other words, the essence of the AI bubble bursting may not be that “AI isn’t being used.”
Rather, the opposite is true: AI will be used on a massive scale. However, it will be used too cheaply. As a result, providers will not turn a profit. In other words, AI will become something akin to a public good. This is the most likely scenario.
4. Chinese Companies and Open Models Are Accelerating Price Erosion
Furthermore, the presence of Chinese LLM companies and open-source models is accelerating the decline in prices.
In the past, Japanese, European, and American companies were the first to advance technological development in LCD panels and solar panels. However, South Korean, Taiwanese, and Chinese companies subsequently made large-scale investments and established mass production systems, significantly reducing the profit margins of the early-mover companies through price competition.
A similar scenario could unfold in the AI sector.
Of course, LLMs are different from physical panels. However, once models meeting a certain performance threshold become commoditized, API prices fall, and open-source models become widespread, it will become difficult to differentiate the models themselves.
Top-performing models will likely retain their value, but for many everyday tasks, anything less than top performance will suffice—such as summarization, translation, FAQs, document search, meeting minute organization, simple code assistance, and support for routine tasks. These tasks do not necessarily require state-of-the-art models.
As a result, the AI market will become polarized. On one side will be expensive frontier models,
and on the other, inexpensive mid-range models, open models, and local LLMs.
As this polarization progresses, the profitability of AI infrastructure built with massive investments may become even lower than initially anticipated.
5. Many AI startups will be weeded out
If the AI bubble bursts, AI startups will likely be the first to feel the brunt of it. As the saying goes, “When the parent turtle falls, the baby turtle falls too.”
The following types of companies are particularly at risk:
- Companies that have merely wrapped a thin layer around an LLM API.
- Companies with many proof-of-concepts but few recurring revenue streams.
- Companies unable to absorb GPU costs.
- Companies that can easily be replaced by major models as new features are added.
- Companies that secured high valuations based solely on the “AI” label.
In the midst of a boom, even these companies can raise funds. Since investors are betting on growth potential, it becomes easier to secure funding if a company can demonstrate—by any means necessary—that it is rapidly scaling its operations. It’s well known that during the dot-com bubble, startups engaged in sham transactions with one another to inflate their apparent revenue. In a boom, such practices are often overlooked. However, once investors begin to realize they cannot recoup their investments, the situation reverses. Valuation criteria inevitably change.
- Do they actually have customers?
- Is the product being used consistently?
- Are they generating gross profit?
- Can they pass on AI costs to customers?
- Do they have a unique selling point that will prevent them from being swallowed up by major players?
Companies that cannot answer these questions will find it difficult to raise funds and will likely be forced to withdraw from the market, merge, or go out of business. It’s fine to operate on a small scale and build steadily, but companies that have expanded their scale after receiving investment will suddenly find themselves unable to survive in this phase.
6. Even major companies are not immune
Even major corporations are not necessarily safe.
Companies such as NVIDIA, Microsoft, Google, Amazon, Meta, and Oracle are unlikely to disappear even if the AI bubble bursts. They are highly likely to remain at the core of AI infrastructure.
However, the fact that a company will survive is distinct from whether its current stock price or investment plans are justified. A bubble burst does not simply mean that a company will go bankrupt. It means that overly optimistic expectations of future profits will be adjusted to more realistic levels.
Therefore, even for major companies, if the profitability of AI-related investments is called into question, we could see significant stock price corrections, scaled-back investment plans, delays in data center construction, and renegotiations of related contracts. In particular, companies whose business models rely on massive data center investments—and the investors funding them—will find themselves at the epicenter of the AI bubble’s collapse.
7. Masayoshi Son and SoftBank will likely be among the main culprits as well.
In this context, the role of Masayoshi Son and SoftBank is extremely important. Son has consistently made major bets on massive technological trends—such as the internet, mobile, Alibaba, the Vision Fund, and Arm. If he succeeds, he becomes a historic winner; if he fails, he incurs massive paper losses. It is precisely this wide swing that defines Son’s management style.
The same is happening with AI. SoftBank is strengthening its involvement with OpenAI and AI infrastructure, and is also participating in massive initiatives like Stargate. If investments in AI infrastructure generate the expected returns, Son will likely be reevaluated as “the man who foresaw the AI infrastructure era.”
However, if the AI bubble bursts, OpenAI’s valuation plummets, the profitability of data center investments is called into question, and the cost per unit of AI usage crashes, SoftBank could be severely impacted.
In this scenario, Son would not be merely a victim. Rather, he would likely be viewed as one of the key players in the AI bubble. If he succeeds, he will be the “infrastructure king of the AI era.” If he fails, he will be the figure who symbolizes the AI bubble. He is in a position where either outcome is possible.
8. Cloud AI will become more expensive, and local AI will become more widespread
What is likely to happen after the AI bubble bursts is a division of roles between cloud AI and on-premises AI.
Currently, the practice of offloading everything to high-performance models in the cloud is becoming widespread, but this situation cannot continue indefinitely. As the cost of cloud AI rises, free tiers shrink, enterprise pricing increases, and usage restrictions tighten, businesses and individuals will have no choice but to seek alternative solutions.
Local LLMs and in-house LLMs will likely fill this gap. Of course, local LLMs cannot immediately and completely replace top-performing cloud models. For advanced reasoning, long-text processing, multimodal generation, and large-scale agent execution, the cloud will likely retain its advantage for the time being.
However, there will be an increasing number of situations where local LLMs are sufficient for day-to-day tasks.
- Text generation
- Summarization
- Translation
- FAQ support
- Meeting minutes organization
- Internal document search
- Simple code assistance
For these uses, the model does not need to be the highest-performing one. Rather, it is more important that it be affordable, stable, and capable of handling internal data securely. As a result, AI usage is likely to shift toward the following model:
- Local LLMs for day-to-day tasks.
- Internal servers or private clouds for processing internal knowledge.
- High-performance cloud models reserved only for complex problems and advanced decision-making.
AI may shift from a highly centralized cloud service model toward a more distributed hybrid structure.
9. What Remains After the Bubble Bursts
Even if the AI bubble bursts, not everything will disappear. In fact, it is only after the bubble bursts that we will see what was truly valuable. What will likely remain are the following, more understated elements:
- Systems for organizing company-specific knowledge.
- AI deeply integrated into business processes.
- Knowledge bases that leverage internal documents and tacit knowledge.
- Mechanisms for verifying AI responses.
- A knowledge base that remains intact even when models are replaced.
- Operational designs that distinguish between cloud AI and on-premises AI.
- Systems that accumulate human judgment criteria and experience.
Conversely, what will disappear are services that claimed value based solely on the term “AI.”
- “AI can do anything.”
- “Generative AI will revolutionize business.”
- “AI agents will automate everything.”
Such claims may hold water at the height of the boom, but sooner or later it will become clear that they were nothing but illusions. As the bubble bursts, these claims will rapidly lose credibility.
10. The collapse of the AI bubble does not spell the end of AI adoption
The most important thing here is not to confuse the bursting of the AI bubble with the failure of AI itself.
- AI is here to stay.
- LLMs are here to stay.
- Local LLMs will continue to evolve.
- Companies will continue to adopt AI.
However, the frenzy in the capital markets will be corrected. Overvalued companies will see their valuations drop. Unprofitable services will disappear. Massive infrastructure investments will be reevaluated. The term “AI” alone will no longer be enough to sell products or raise funds. What will remain afterward is a more low-key, more practical, and more business-oriented use of AI.
This isn’t all bad for society. In fact, for AI to enter a stage where it is truly useful, the bubble may need to burst first.
Conclusion: It is not AI that will collapse, but rather the excessive expectations placed on it.
The scenario for the collapse of the AI bubble can be summarized as follows.
- AI technology is real.
- However, investment in AI may be excessive.
- Investment in data centers is costly and time-consuming.
- LLM prices are falling rapidly.
- Chinese companies and open models are accelerating price competition.
- Many AI startups will be weeded out.
- Even major corporations will be forced to revise their stock prices and investment plans.
- Cloud AI will become more high-end, while on-premises AI and hybrid operations will become more widespread.
Ultimately, what will remain is not the models themselves, but the practical application of AI rooted in knowledge, operations, and judgment.
Therefore, the bursting of the AI bubble does not mark the end of AI. It marks the end of the illusions surrounding AI. And once those illusions are dispelled, companies will finally be forced to confront fundamental questions.
- Is our company’s knowledge properly organized?
- What should we entrust to AI, and what should humans decide?
- Are there knowledge assets that will remain even if the models change?
- What kind of AI is truly usable in our day-to-day operations?
What remains after the AI bubble bursts is not flashy slogans. What remains will likely be down-to-earth AI utilization deeply rooted in knowledge, judgment, experience, and business processes.