Conclusion

To put it simply, the money that large corporations have poured into AI is starting to burn.

AI inference startup Fal has doubled its valuation to $16 billion (approximately ¥2.4 trillion) and raised $130 million. Meanwhile, Anthropic is facing deteriorating computational costs, revealing a structure where infrastructure expenses outpace revenue. Oracle is betting on AI expansion with a $20 billion fundraising effort.

When these three news items are lined up, a single structure becomes apparent.

“The players who have invested heavily in AI are being dragged down by the gravity of computational costs.”

This is not just a concern for small and medium enterprises (SMEs); in fact, it is the opposite. The moment large corporations become immobilized by this gravitational pull is precisely when SMEs have the opportunity to acquire weapons that are “cheap and fast.”

—

What is Happening — The “Bomb” of Computational Costs

First, let’s clarify the facts.

Anthropic’s Claude 3.5 Sonnet is one of the highest-performing LLMs, on par with GPT-4o. However, as performance increases, the GPU costs associated with inference (the process of answering user questions) skyrocket. While Anthropic’s annual revenue for 2024 is reported to be around $850 million, investments in computational infrastructure are increasing at a pace that significantly exceeds this revenue. Even if revenue grows, costs are growing even more. This is the essence of the “deterioration of computational costs.”

On the other hand, Fal is a company attempting to solve this computational cost problem through “inference optimization.” It provides infrastructure that lowers the costs of running AI models. The reason its valuation has doubled to $16 billion is simple: investors believe that “lowering the cost of using AI” itself will become a massive market.

In other words:

• Those creating AI models (like Anthropic and OpenAI) → Struggling as costs balloon
• Those providing affordable AI model operation (like Fal) → There is enormous demand in this area
• Those betting on AI infrastructure (like Oracle) → Investing $20 billion in capital expenditures

What this structure implies is that the “center of gravity” in the AI industry is shifting from model development to cost optimization.

—

Why This is the “Sound of the Bubble Bursting”

Looking back at past bubbles, the pattern is the same.

The dot-com bubble of 2000 saw massive inflows of capital into internet companies, inflating valuations even without revenue. After the collapse, servers, domains, and engineers were released onto the market in large quantities. Amazon and Google grew by picking up resources cheaply from this “burned-out landscape.”

What is happening now in the AI industry resembles this structure.

• 2023-2024: Investment in AI-related startups exceeds $100 billion annually worldwide (according to PitchBook)
• Overheating GPU demand: NVIDIA’s H100 costs about $30,000 each, with investments in large-scale clusters reaching hundreds of millions becoming the norm
• However, many AI services still operate under a “revenue < costs" deficit structure

This state cannot be sustainable. The moment investment money dries up, startups lacking stamina will begin to collapse. What will happen then?

Talent will flood the market. Technologies will become open source. Tool prices will plummet.

This is just like the aftermath of the dot-com bubble burst.

—

Three Things SMEs Should Do “Now”

Now we get to the main point. It would be meaningless to end with “That’s an interesting story.” Local SMEs need to focus on three actions they should take to actually pick up weapons from this structural change.

1. Inventory “Tasks in Your Business That Can Be Replaced by AI”

Even if the bubble bursts and tools become cheaper, you cannot pick them up if you haven’t decided “what to use them for.”

What you need to do is simple. Write down all the tasks in your company and ask each one, “Is this something that needs to be done by a human?”

• Creating estimates → By templating and using AI generation, a task that takes 30 minutes could be reduced to 5 minutes
• Responding to inquiries → If you let an AI chatbot handle FAQs, 40 hours of phone support a month could be reduced to 10 hours
• Daily reports and documentation → Using voice input and AI summarization, a task that takes 30 minutes a day could be almost eliminated

The important question is not “How do we implement AI?” but rather “Can we reduce the cost of this task to one-tenth?” The tools can be chosen later. Having a list of “where to use it” will be the condition for moving quickly after the bubble bursts.

2. “Familiarize Yourself with Open Source LLMs”

Open-source AI models like Meta’s Llama 3.1, Mistral’s Mixtral, and Google’s Gemma are rapidly reaching practical levels of usability.

API-based models like GPT-4 and Claude incur monthly costs ranging from tens of thousands to hundreds of thousands of yen. However, if you run open-source models on your own servers or in the cloud, the running costs can drop to just a few thousand to tens of thousands of yen per month, excluding initial setup costs.

I can hear some voices saying, “But we don’t have engineers.”

That’s precisely why you should familiarize yourself with them now. Using a tool like Ollama, you can install and run Llama 3.1 on a laptop in less than 30 minutes. It’s free. First, create a state where at least one person in your company has “touched it.” This way, when you hire AI talent after the bubble bursts, you’ll know “what to ask for.”

3. Start Creating an “AI Talent Recruitment List” Now

After the bubble bursts, the most valuable asset will be “affordable talent.”

Currently, the salary range for AI-related engineers is between ¥10 million and ¥20 million. Large corporations and startups are competing fiercely for them. However, if startups collapse, this talent will flood the market.

Local SMEs do not need to offer ¥20 million salaries. It would be sufficient to build a relationship that starts with remote work or contract work at ¥200,000 to ¥300,000 per month. The important thing is to understand “where this talent is” from now on.

Specifically, follow engineers who are sharing AI-related content on X (formerly Twitter). Look at participants in Kaggle competitions. Find people hosting AI study groups on platforms like connpass or ATND. You don’t need to hire them immediately. Just creating a list of “people I want to reach out to” will allow you to act when the timing is right.

—

When Large Corporations Become Heavier, SMEs Can Become Faster

Oracle is raising $20 billion to invest in AI infrastructure. Anthropic is grappling with computational costs. These are battles of large corporations.

However, as a “byproduct” of this battle, the prices of AI tools will undoubtedly decrease. The performance of open-source models will improve. Technologies and talent will flow into the market from the startups that have been eliminated.

Large corporations must recoup their massive investments, making them slow to pivot. SMEs can move faster because their investment amounts are smaller.

When cloud computing became widespread, there are numerous examples of SMEs that began using AWS for just a few tens of thousands of yen a month advancing digitalization faster than large corporations that maintained server rooms. The same will happen with AI.

The question is simple: “When the bubble bursts, is your company ready to pick up the pieces?”

Preparation does not mean massive investments. It means inventorying your operations, having experience with tools, and maintaining a list of talent you want to reach out to. With these three, SMEs can move faster than large corporations.

The collapse of the AI bubble is not something to fear. It is something to prepare for. And only those who are prepared will acquire weapons in the next stage.