Title

Large Corporations Hoard AI Compute, Leaving 90% Unused—Meanwhile, a ¥30,000 Monthly API Fee Becomes the Optimal Solution for Small Businesses

Main Text

Large corporations are hoarding GPUs, with 90% of them going unused.

This is the current reality of AI infrastructure investment. “I don’t want to miss out”—excessive investments driven by FOMO (Fear of Missing Out) are creating enormous waste. On the other hand, there are small businesses achieving significant results with a monthly API charge of ¥30,000. It’s essential to calmly observe this asymmetry in structure.

—

90% of GPUs are “asleep”—The Reality of FOMO Investments

As the AI competition intensifies, large corporations are scrambling to secure computing resources. The latest GPUs from Nvidia, such as the H100 and H200, cost several million yen each. They procure thousands to tens of thousands of units and stack them in data centers.

However, what is the reality? According to reports from research firms, about 90% of the AI compute resources purchased by large corporations are not operational. In other words, investments amounting to billions of yen are merely “secured for now” and lying dormant.

Why does this happen? The reasons are simple:

• The herd mentality of “If our competitors are buying, we should too”
• Concerns about supply: “We might not be able to get it next year”
• The reverse phenomenon of buying “AI infrastructure” before deciding “what to do with AI”

In short, infrastructure is being prioritized without a clear purpose. This mirrors the scenario where expensive SaaS contracts were signed under the banner of “promoting DX” but went unused.

Utah’s 9GW Campus—A Symbol of “AI That Consumes Power”

There is a symbolic case of this excessive investment: a 9GW AI data center campus approved in Utah, USA.

How large is 9GW? The entire power supply capacity of Utah is about 4GW. This means that a single AI data center would consume more than double the state’s electricity. In terms of nuclear power plants, it equates to about nine reactors. The fact that this has been approved as “infrastructure” for AI vividly illustrates the overheating of investments.

Even Nvidia executives acknowledge that “AI is more costly than human labor.” This is not a sarcastic remark or an act of humility; it is a fact at this point. The costs of GPUs, electricity, cooling, and data center construction—owning AI is more expensive than hiring humans for many operations.

Consider this: Is the logic of building a 9GW data center the same as that of your company?

Of course not. The reasons why large corporations must “own” infrastructure and what small businesses need to “use” AI are fundamentally different.

¥30,000 is Enough—The “Reverse Calculation” of Small Businesses

So, how much does it actually cost for small businesses to use AI in their operations?

To cut to the chase, with ¥30,000 a month, a lot can be achieved.

Let’s do some specific calculations.

OpenAI GPT-4o API Pricing (as of 2024):

• Input: $2.50 / 1 million tokens
• Output: $10.00 / 1 million tokens

For example, let’s assume AI handles 100 inquiries a day. If we assume 500 tokens for input and 1,000 tokens for output per inquiry:

• Daily input: 50,000 tokens → about $0.125
• Daily output: 100,000 tokens → about $1.00
• Daily total: about $1.125
• Monthly (30 days): about $33.75 ≒ about ¥5,200

For ¥5,200 a month, you can automate the handling of 100 inquiries per day. In terms of labor costs, this translates to savings equivalent to one part-time employee’s monthly salary of ¥150,000 to ¥200,000.

Even if you add slightly heavier tasks—such as automatic summarization of daily sales reports, generating analysis reports on customer data, or creating Q&A bots for internal manuals—it’s unlikely to exceed ¥30,000 a month.

There’s no need to buy a ¥3 million GPU server. A ¥30,000 API is sufficient.

The ability to choose this “non-ownership” option is, in fact, the greatest weapon for small businesses.

A Structure for Overcoming Large Corporations

Let’s organize the structure here.

Looking at this table, it’s clear. The amount invested in AI and the value derived from AI are not proportional at all.

Large corporations “own” but “cannot use.” Small businesses “do not own” but “can fully utilize.” This is the structure of reversal.

Moreover, this structure becomes increasingly advantageous for small businesses over time. Why?

Because API prices continue to drop.

The API pricing for GPT-4 has halved in a year. With the introduction of GPT-4o, it has dropped even further. As long as competition continues with Claude and Gemini, this price decline will not stop. In other words, as long as small businesses choose the “pay for what you use” model, costs will automatically decrease.

On the other hand, the GPUs hoarded by large corporations do not decrease in price. In fact, they become obsolete with each new generation of chips. An H100 bought for hundreds of millions of yen will become outdated in two years. This is the very risk of “ownership.”

“So, what should we start with?”

Let’s leave the abstract discussions behind. There are only three things that small businesses should do starting tomorrow.

1. Choose one task that “people are doing but can be patterned”

Inquiry handling, summarizing daily reports, drafting estimates, organizing meeting minutes—anything is fine. The first target should be a task that feels repetitive.

2. Experiment with the API. Try it for ¥5,000 a month

OpenAI, Claude, Gemini—any of them will do. Obtain an API key and automate one task first. The initial experiment won’t cost more than ¥5,000. If it doesn’t work out, just stop. You’ll only lose ¥5,000.

3. Distinguish between “tasks for humans” and “tasks for AI”

There’s no need to replace everything with AI. Building relationships with customers, making on-the-spot judgments, and delivering the final word in complaint handling—these should be done by humans. Delegate the preliminary “prep work” to AI. If you can make this distinction, you can create a system that works even with a small number of people.

The Real Risk is “Not Doing Anything”

The FOMO investments of large corporations may seem ridiculous. Hoarding GPUs that are 90% unused, right?

However, the real risk for small businesses lies elsewhere. It’s not using a weapon that can be obtained for ¥30,000 simply because they are unaware of it.

While competitors are shortening the time to create estimates from 30 minutes to 3 minutes using AI, your company continues to rely on manual processes. While neighboring companies are using AI for customer analysis to improve proposal accuracy, your company is relying on intuition and experience.

This is not a question of whether to invest in AI. It’s about ¥30,000. It’s not even an amount that qualifies as an investment.

While large corporations are burning hundreds of millions of yen to build infrastructure, small businesses can access the same technology for a monthly API fee of ¥30,000. This asymmetry is historically unprecedentedly favorable for small businesses.

There’s no need to hoard GPUs. “Pay for what you use.” That alone can make AI a friend to small businesses.