Are you all using generative AI? I use it all the time, including OpenAI's Chat-Gpt (Windows' Copilot), Google's Gemini, and Anthropic's Claude.
I also tried local LLM, but unfortunately, my PC didn't have enough specs to make it very user-friendly.
*My PC is an ASUS Vivobook S14 M5406WA Copilot PC.
In the market, both the number of users and subscribers to generative AI services are steadily increasing. Generative AI-related topics now account for a significant portion of tech news in the mass media.
The environment surrounding AI is booming to such an extent that it almost feels like a bubble.
This time, we will use systems engineering techniques to analyze whether this booming market, which could be described as a generative AI bubble, is actually a bubble.
Let's start by reviewing the basic industrial structure of generative AI and then consider the bubble in generative AI.
Basic Industrial Structure of Generative AI
The typical environment for using a generative AI is to launch the generative AI app on your device, input information (prompts) into the app, and then some kind of information will be generated and displayed.
Recently, advanced AI agents and similar technologies have become available to ordinary people like us, but the basic principle is as shown in the diagram above.
Next, let's broaden our scope a bit. High-performance software like generative AI basically cannot operate on the devices we have access to, so each company's AI will primarily utilize data centers (cloud).
*Regarding local LLMs that run solely on a terminal, as of February 2026, I believe even a PC costing around 300,000 yen would be difficult to make practical. I hope for improvements in the near future.
The key point is that information is generated in the data center, not on the terminal. To use an analogy, the terminal is like the reception desk, and the data center is like the brain.
Although the diagram depicts a data center as a single entity, in reality, numerous data centers exist and are used in a distributed manner.
*The system determines which data center to use for each user and instruction set.
The basic configuration of a data center is as follows:
As you can see from the data center's configuration, it's almost identical to a PC. However, data centers house a vast number of servers, which are connected by a network, and are equipped with massive cooling and power supply systems to support that enormous number of servers and network.
*As a general guideline for scale, a medium-sized data center has around 1000 servers, while a large-scale data center has around 5000.
GPUs, which are often discussed in the media, are the core components of a server's computing power. For AI applications, a single server typically uses 4 to 10 GPU cards (8 is a common standard configuration).
*Unlike CPUs, GPUs are arithmetic units that excel at executing specific instructions (matrix calculations) and are good at parallel processing.
To operate a single data center, you would need approximately 8000 GPU cards in a standard setup (assuming 1000 servers), and up to 40,000 cards in some cases (assuming 5000 servers).
The sheer number of GPUs required alone clearly illustrates the enormous resources needed to maintain and manage such a massive facility.
Furthermore, GPUs are components that require very frequent replacement due to the following issues:
- Performance advancements are rapid, and for AI applications, replacement is often necessary every 1-2 years.
- Operating multiple GPUs simultaneously can lead to overheating and a short lifespan, resulting in frequent failures.
- Since multiple GPUs are running continuously in the data center, one of them is always going to fail.
etc.
To summarize, you should now understand that the following elements are necessary to operate a data center for AI.
Unfortunately, this doesn't mean the data center is operational and the generative AI is ready to use. We still need the crucial software (application) for the generative AI.
Even with these changes, the generation AI is still not operational.
Generative AI doesn't simply generate new information based on prompts using software alone; it takes existing information created by humans as a base and reshapes it to be appropriate for the prompt. Therefore, information is necessary for generative AI to operate (primarily via the internet).
*Learning will also be done using existing information.
Now we finally have all the basic elements needed to use generative AI. Let's write out the whole picture.
As you can see from the diagram above, the interactions we usually have with generative AI, which are limited to just our own devices, involve this much complexity and a multitude of hidden elements.
However, there is still a hidden important element. That is, the information exchange between each element and the data center is supported by a massive network (the internet).
As you can see from what we've discussed so far, it's clear that behind the generation AI we use on our devices, there are not only generation AI companies but also numerous other players lurking.
While generative AI is probably strongly associated with the cutting edge of the information industry (IT industry), it's actually a broad industry encompassing information, physical objects, and more.
*The media often reports on this in the context of the latest IT industry.
The important point here is that if even one of the above players is missing, the AI generation industry cannot function. Therefore, when viewed as a whole, the industry has a structure that is far removed from the typical image of IT and resembles an infrastructure industry.
This concludes the basic industrial structure of generative AI.
Revenue of leading generative AI companies
Now, let's take a closer look at the revenue of each AI company.
We will look at the revenue from AI-related businesses of leading generative AI companies. However, since Google and Microsoft have many businesses and it is difficult to determine their pure revenue from generative AI alone, we will focus on OpenAI and Anthropic (we will omit market capitalization as it is not very relevant to the real economy).
*Since neither company is a fully incorporated company, the exact figures are unknown.
Both companies have sales exceeding 1 trillion yen and are earning considerable profits. Looking at this alone, they appear to be highly successful, up-and-coming venture companies.
However, the reality is quite harsh, and both companies are projected to be operating at a loss in 2025.
As expected of the trendy AI industry, the losses are enormous. It's an unthinkable situation, and coming from the stable automotive industry within the manufacturing sector, I'm completely baffled.
The main reasons for the deficit are the fierce competition in the service and development of generation AI, as well as the enormous payments related to data centers, that is, to hidden players, as explained in Chapter 1 of this article.
Another difficult aspect of this business is that as the number of users, who are a source of revenue, increases, the amount of equipment needed also increases, and consequently, the expenses also increase, which is extremely tough.
In other words, simply increasing the number of subscribers in the current situation will not solve the problem of operating at a loss, which puts us in a very difficult position.
Despite these enormous losses and a structure that makes it difficult to generate profits, the two AI generation companies are able to survive because they are subsidized by massive amounts of money from investors.
The main reason this enormous amount of capital is being raised is that investors are competing to pour in money based on the expectation that "generative AI will become a social infrastructure in the near future and generate enormous profits."
This is exactly it.The generative AI market is in a bubble.That's the main reason I think so.
Therefore, each company constantly updates and publishes its business plan, outlining the timing of profitability, to ensure that funding from these investors does not cease and to prevent investors' expectations from waning.
Investors are currently investing in the hopes of enormous profits in 2028, 2029, and beyond, but a major point of contention will likely be how they react if the actual timing of profitability is delayed.
Even based on the general earnings situation so far, it's fair to say that a bubble is already expanding.
Incidentally, the definition of a bubble is as follows, and it is safe to say that the generative AI market will definitely be far removed from reality at least by 2025 and 2026.
The price of assets such as real estate and stocks, driven by speculation, rises at a pace exceeding economic growth, becoming significantly detached from the real economy.
→Since it is surviving while running massive deficits, it can be said that it is in a state far removed from reality, and therefore can be called a bubble.
Next, let's take a closer look at the flow of funds and resources for OPEN AI and Anthropic.
*The reason for featuring OPEN AI and Anthropic is simply that they are almost exclusively focused on generative AI business and their structure is easy to understand; there is absolutely no other intention.
The industrial structure surrounding Open AI
First, let's look at OPEN AI's relationship with Microsoft, its largest partner.
The main components that Microsoft is providing to Open AI are its massive cloud services (data centers) and funding (totaling $130 billion: approximately 2 trillion yen).
*The general name for cloud services is Azure.
In return, you will receive the exclusive right to use OpenAI's generated AI, a commercial license, a share of OpenAI's revenue, and cloud fees.
From the surface, it appears that OpenAI is responsible for the software, while Microsoft handles data center-related matters, creating a division of labor.
Microsoft's cloud system (Azure) requires high-performance servers to deliver the necessary performance for using OpenAI's generative AI. The core component of this system is the much-discussed GPU.
The key player that emerges here is NVIDIA. NVIDIA develops GPUs based on technical requirements from OpenAI and feedback from real-world operational data, and Microsoft pays for and purchases these GPUs to implement them in its own cloud system.
*In 2026, it is planned to be a joint development project between NVIDIA and OpenAI.
The unique position of NVIDIA, a key player in this context, lies in its deep relationship with Open AI in GPU development. From another perspective, this could be seen as Microsoft successfully leveraging Open AI to significantly enhance the AI capabilities of its own cloud platform.
Of course, other data center-related companies besides NVIDIA also do business with Microsoft and Open AI in a similar way, but their relationship with Open AI is not as direct and deep as that of GPUs, and their primary trading partner is Microsoft.
Up to this point, all the necessary players for generative AI, as explained in Chapter 1 of this article, were in place. However, in 2024, Microsoft's cloud service (Azure) was not large enough to support the demand for generative AI (this could be considered the beginning of the AI bubble).
That's where Oracle, famous for its System Greta, comes in. Originally strong in database management software and infrastructure development, its late entry into the AI industry allowed it to set up data centers optimized for the latest AI (the timing was just right).
*First of all, Oracle is a gigantic corporation with plenty of money and a powerful founder, so its decision-making is unique. In my area of expertise, it's similar to irreplaceable manufacturers like DuPont or 3M.
This completes the basic key players surrounding Open AI.
From here on, players emerge who seek to amplify the generative AI industry. These are investors such as venture capitalists and banking syndicates.
This player invests capital (money) with future growth as collateral, accelerating and further expanding the already expanding AI generation bubble.
Further complicating matters is the emergence of a movement by various companies, including OpenAI, to establish their own massive data centers dedicated to AI.
*The goal is for Open AI to escape various constraints and operate freely using its own data center.
In addition to building its own data center, Open AI has also embarked on developing its own proprietary GPU chips.
*The goal is to break away from NVIDIA dependency.
This concludes the basic structure surrounding Open AI in early 2026.
Let's focus solely on the flow of funds and isolate OpenAI.
As can be seen from the diagram above, Open AI's income and expenses are spread across many different areas. However, the income shown in the diagram above is basically from investments, so it will eventually need to be repaid. In other words, the only source of funding for these various payments is the subscription and API usage fees for chatgpt, which is a rather alarming situation.
Based on these points, Open AI's funding inputs and outputs can be summarized as follows:
The cash flow results show that the company's only source of revenue, chatgpt subscription fees and total API usage fees, are nowhere near enough to cover its expenses, resulting in a deficit of over $120 billion (more than 2 trillion yen) in 2025. Moreover, in the future, the company will need to repay the loans it has received in some form.
This deficit of over $12 billion is so far removed from reality that it can be seen as a direct reflection of the inflated bubble. To put this into perspective, the $20 (approximately 3000 yen) subscription fee for CAHTGPT's most popular plan, Plus, is essentially being covered by $120 (approximately 1800 yen), mainly from investor funds.
What's even more frustrating is that when they try to increase the number of users to boost their revenue, the increased users also lead to higher data center-related expenses, making it difficult to translate that into profits.
Even in these challenging times, the fact that OpenAI is attracting such massive amounts of capital from investors is definitely indicative of a bubble.
This is the structure of the bubble surrounding Open AI.
For reference, let's take a look at NVIDIA, which is considered the overwhelming winner in the generative AI market. Similar to OpenAI, we'll focus solely on NVIDIA.
As you can see from the diagram, all of NVIDIA's funding is based on actual costs, as they simply receive payment for the GPUs they supply.
*NVIDIA has no known source of funding outflow.
In other words, the key to overwhelming success lies in steadily converting the demand inflated by the AI generation bubble into tangible funds through core business operations.
However, there are rumors that NVIDIA will invest in the Stargate concept (which doesn't yet exist) after 2026, so this might be one of the turning points.
As of February 21, 2026, when this article was being written, NVIDIA has decided to invest $30 billion in OpenAI. This could potentially cause another bubble to inflate and mark a turning point.
Industrial structure surrounding Anthropic
Similar to OpenAI, let's examine the structure of Anthropic.
First, Anthropic's biggest partners are Amazon and Google. These companies provide funding and cloud services (data centers). In return, Anthropic provides inference models, cloud usage fees, a portion of its revenue, and operational data.
*Anthropic was founded due to an internal split over the philosophy of Open AI, and therefore, from one perspective, it is also a Microsoft vs. Amazon/Google alliance.
It's basically similar to the relationship between OpenAI and Microsoft. The biggest difference is that the work is divided between the two companies.
The next player to emerge is NVIDIA, which supplies GPUs, similar to OpenAI's analysis. However, the structure surrounding Anthropic makes this GPU relationship quite complex, as Google and Amazon supply their own AI computing chips to reduce their dependence on NVIDIA alone.
* AI processing chip, not GPU
This is a slight digression from the main topic, but Anthropic, wary of OpenAI's over-reliance on Microsoft, has partnered with two IT industry giants, Google and Amazon.
Similarly, the AI computing devices in servers are divided into three systems—NVIDIA chips, Google chips, and Amazon chips—as a way to diversify risk against NVIDIA's dominance.
At first glance, this seems like a rational strategy for risk diversification, but since the data center would be split into three separate systems, there are disadvantages such as technical difficulty and difficulty in achieving economies of scale, so it's hard to say whether it's a good idea or not.
*NVIDIA uses GPUs for image processing, while Google and Amazon use TPUs for matrix and tensor operations; they are all different.
Next, we'll look at data center-related companies other than those offering AI computing chips.
This completes the list of essential key players surrounding Anthropic.
From this point on, players will emerge who, like OpenAI, are trying to amplify the generative AI industry. These include venture capital firms, banking consortia, and government-backed investment funds.
Just like with OpenAI, these investors are further inflating the Anthropic bubble.
At this point, the basic framework of the Anthropic bubble structure is complete.
Up until this point, the situation had been Microsoft vs. Amazon and Google, but in November 2025, Anthropic and Microsoft joined forces, further complicating matters.
This means Anthropic has secured new data centers and funding from Microsoft, accelerating the bubble. In the end, Anthropic has teamed up with three giants of the IT industry.
*This could be seen as Microsoft attempting to move away from relying solely on OpenAI, resulting in an increase in GPU orders to NVIDIA via Microsoft.
While newly partnering with Microsoft, Anthropic is also showing signs of moving towards building its own data centers. In November 2025, the same month that it announced its partnership with Microsoft, it also announced its decision to build its own data centers.
Interestingly, Anthropic's data center construction is a completely separate project from OpenAI's Stargate initiative. While slightly off-topic, in the US, data centers are experiencing a kind of bubble, with giant manufacturers like OpenAI (with its Stargate project), Anthropic (with its own project), Oracle, Microsoft, Google, and Amazon all fiercely competing to build new ones.
Furthermore, like OpenAI, Anthropic is also embarking on the development of its own proprietary AI computing chip.
*Interestingly, their partner is Broadcom, the same as OpenAI.
With this setup, Google and its own proprietary chips (Broadcom) will supply the AI computing chips for the new data center.
*It is unclear whether Amazon's chip will be adopted as of February 2026, but it is likely that it will be adopted in the future.
The strategy surrounding the procurement of these chips goes beyond this; in November 2025, they will also be forming a strategic partnership with NVIDIA. This will result in a joint GPU development system with NVIDIA.
This is the basic structure surrounding Anthropic in early 2026. The situation is considerably more complex than that of OpenAI due to the risk diversification measures.
This is slightly off-topic, but I'm worried about whether Anthropic's new data center's AI computing chips, which come from four different systems—NVIDIA, Google, Amazon, and Broadcom—will be able to function properly.
From here, similar to the Open AI analysis, we will focus solely on the flow of funds and isolate only Anthropic.
As can be seen from the diagram above, Anthropic's revenue and expenses are diverse, similar to OpenAI. In fact, because its structure is more complex than OpenAI's, they are even more diverse.
However, the revenue shown in the diagram above is basically from investments, so it will eventually need to be repaid. This creates a terrifying structure, similar to Open AI, where the only source of funding for these multifaceted payments is the subscription and API usage fees for Claude.
Based on these points, Anthropic's funding inputs and outputs can be summarized as follows:
Like OpenAI, Anthropic is facing massive losses because its sole source of funding—subscription and API fees for Claude—is nowhere near enough to cover these wide-ranging and enormous payments. Moreover, in the future, they will need to repay the loans they received in some form.
This deficit of up to $52 billion can be considered the exact amount that was inflated by the Anthropic bubble in 2025. To put this into more relatable terms, behind the $20 (approximately 3000 yen) per month price for Anthropic's Claude Pro plan, $11.56 (approximately 1800 yen) is being subsidized in some way.
This is precisely the structure of an anthropic bubble.
Summary
After this lengthy explanation, I think it's clear that it's almost certainly safe to conclude that the generative AI market is in a bubble.
I believe the revenue and loss data for 2025 will serve as perfect evidence.
In a normal market, no industry could continue operating with this level of deficit.
An analysis of the industry structure reveals that two leading companies, OpenAI and Anthropic, have remarkably similar structures.
In other words, it's fair to say that the bubble is inflating not just because of expectations for a specific company, but because of expectations for the entire industry. Furthermore, the typical pattern of capital repatriation during a bubble is also occurring.
I think Microsoft and Google, which are integrated IT companies that provide generative AI services, are in a better situation, but I believe their basic structure is similar.
Furthermore, what fundamentally distinguishes this AI bubble from previous IT industry bubbles is that investment in massive infrastructure (data centers) is a major driving force behind the bubble (including AI computing chips).
The situation appears remarkably similar in structure to historical events such as the railroad bubble (Britain in the 1840s) and the now-nostalgic .com bubble (the IT bubble of the late 1990s).
Based on the above, I believe it's fair to say that the AI generation bubble is not just a bubble, but has entered a stage where the bubble is structurally expanding rapidly.
From a historical perspective, the current generative AI bubble market can be seen as having entered a state of capital investment competition.I think it's safe to say we've entered a typical mid-to-late stage of the bubble economy..
The troublesome thing about this bubble is,Historically, there are no examples of companies surviving a bubble without suffering any damage. Moreover, structurally, the AI-generated bubble has gone beyond the mid-stage due to a construction boom in equipment.
The next major topic of interest, I think, will be the scenario of a bubble collapse. The key clue to this is, "When a bubble collapses, stocks and bonds become worthless, and what remains in reality?"
I'd like to delve deeper into that bubble collapse scenario in my next article.
P.S
From the user's perspective, the "bubble" can be seen as the best time for customers. In fact, even if each AI generation company is operating at a loss, they are still providing us with services at low prices.
- OpenAI's chatgpt Plus service costs $20 per month, resulting in a $12 loss; it would normally cost $32 to break even. In other words, it's a service that would typically cost over $50 in the market.
- Anthropic's Claude Pro service costs $20 per month, resulting in a loss of $11.56; it would normally break even at $32. Similar to OpenAI, it's a service that should ideally cost around $50.
It's probably safe to say that the period from around 2026 to 2027 will be the cheapest and most powerful generation AI in history to be used to its fullest potential. I think it's wise to make the most of it before the bubble bursts and prices return to a reasonable level (several times the current usage fee).
What's also interesting is that, coincidentally, both Opn AI and Anthropic have similar compensation figures, suggesting that the actual cost might be at least $32 or more, which could be the industry standard.
*Considering the development costs, equipment maintenance, capital investment, and profit margins of a for-profit company, it's certain to be at least $50 or more.
Comment: