Conclusion First

The “producers” of AI are becoming less profitable, while the “users” are gaining an advantage.

Google is pouring $75 billion (about 11 trillion yen) annually into AI infrastructure. Anthropic is burning through $8 billion (about 1.2 trillion yen) each year. Yet, both are being criticized by the market for being “unprofitable.”

This is a tragedy for large corporations, but it is good news for small businesses.

The reason is simple: the AI infrastructure and models created at great cost by these giants are now available at incredibly low prices. What used to cost 3 million yen can now be done for just 50,000 yen — such a structural change is happening right now.

Understanding the Struggles of Large Corporations

First, let’s look at the numbers.

Google’s parent company, Alphabet, announced a capital investment plan of $75 billion for 2025, a 50% increase from the previous year’s approximately $50 billion. Almost the entire amount is allocated for AI-related data centers and GPUs.

Despite this, the growth rate of search advertising revenue has slowed from 17% in the previous quarter to 9.8% in the most recent quarter. The cloud division (Google Cloud) appears to be growing steadily at 28%, but operating profit margins have plateaued. The market has judged that “the returns on investment are not sufficient,” leading to a temporary drop in stock prices of 7-8% following the earnings announcement.

Anthropic is facing an even tougher situation. Its annual revenue for 2024 is reported to be around $850 million, while its computing costs alone are estimated to exceed $2 billion that same year. More than double its revenue is consumed by computational costs. Training Claude 3.5 costs hundreds of millions of dollars, and inference costs are several cents per query. The structure is such that the more users it gains, the larger the losses become.

OpenAI is in a similar situation, with projected revenues of about $3.7 billion for 2024 against annual expenditures of approximately $8.5 billion.

Why is this happening? There are three significant costs associated with developing AI models:

1. Training Costs: It is estimated to cost over $100 million to train a model of the GPT-4 class once.
2. Inference Costs: Every time a user accesses the API, GPUs are utilized. The more users there are, the higher the costs.
3. Labor Costs: The annual salary for AI researchers typically ranges from $500,000 to $1 million, with top-tier researchers earning several million dollars.

In other words, those on the “production side” of AI are being crushed by both enormous fixed and variable costs.

So, What Changes for Small Businesses?

Now we get to the main point.

Thanks to large corporations competing at a loss, the costs for those on the “usage side” of AI have dramatically decreased. Let’s summarize what has happened over the past year.

The Reality of Cost Collapse

Dramatic drop in API costs.

When GPT-4 was released in March 2023, the cost per million input tokens for the API was $30. As of May 2025, GPT-4o is priced at $2.5 per million tokens. In just two years, this has become 1/12th of the original price.

Anthropic’s Claude 3.5 Sonnet charges $3 for input and $15 for output per million tokens. Furthermore, Claude 3.5 Haiku charges $0.8 for input and $4 for output. Practical-level AI is now available for a few thousand yen a month.

Google’s Gemini 2.0 Flash even offers input at $0.1 and output at $0.4 per million tokens. It is practically free.

In other words, with just 5,000 yen a month, small businesses can sufficiently integrate AI APIs into their operations.

Open Source Models Have Caught Up

Another structural change is that open-source AI models have reached a level comparable to commercial models.

Meta’s Llama 3.1 (405B parameters) has achieved benchmark scores comparable to GPT-4. This is available for free. There are also Google’s Gemma 2, Mistral’s Mixtral, and models like ELYZA and CyberAgent for Japanese.

The important factor is the cost of running these models.

• The 70B model of Llama 3.1 can be operated on a cloud GPU server (1 A100) for about 30,000 to 50,000 yen per month.
• A lightweight 8B model can run locally on a PC costing around 150,000 yen with an RTX 4060.
• Furthermore, using quantization (a model lightweightening technique), it can run on a gaming PC costing 80,000 yen.

In the past, running AI in-house required hundreds of thousands of yen for servers and specialized engineers. Now, a 150,000 yen PC and publicly available setup instructions are all that is needed.

A Hidden Gem in the Used GPU Market

As large corporations switch to the latest GPUs like the H100 and H200, older generation GPUs are starting to flow into the used market.

• NVIDIA A100 (80GB): Approximately 2.5 million yen new → About 800,000 to 1 million yen used
• NVIDIA RTX 3090: Approximately 250,000 yen new → About 80,000 to 100,000 yen used
• NVIDIA RTX 4090: Approximately 300,000 yen new → Around 200,000 yen used

The RTX 3090 comes with 24GB of VRAM and can run 70B class models when quantized. It’s an AI inference machine that can be obtained for under 100,000 yen.

“So, What Should We Do?”

Let’s move beyond abstract discussions. Here are three things that local small businesses can start doing today.

1. Experiment with “using” APIs (Initial Investment: Zero to 5,000 yen per month)

OpenAI, Anthropic, and Google’s APIs are pay-as-you-go. If you have a credit card, you can start using them today.

Concrete examples:

• Automatic generation of draft estimates (Input: past estimate data → Output: new project estimate)
• Automatic classification of inquiry emails and drafting responses
• Q&A chatbot based on internal manuals

All of these can be built without engineers by combining clever prompts with Google Apps Script or Make (formerly Integromat). Monthly API costs range from a few hundred to a few thousand yen.

2. Create “Unique Weapons That Cannot Be Imitated by Competitors” Using Company Data and Open Source Models

The greatest asset of small businesses is their unique data accumulated over many years.

• For manufacturing: inspection image data, trends in defective products
• For retail: customer purchase histories, seasonal fluctuation patterns
• For construction: past construction records, cost data

By connecting this data to open-source models using RAG (Retrieval-Augmented Generation), you can create an AI that “knows your business.” A system that would cost 1 to 3 million yen to outsource can be built for about 100,000 to 200,000 yen (including PC costs) using LangChain and Llama 3.1.

While large corporations are desperate to create generic AI, small businesses can focus on creating “AI specifically tailored to their operations.” There’s no need to compete with generic solutions. By specializing, you can create something more useful than the AI of large corporations.

3. Implement AI as a Mechanism to Break Dependency on Key Individuals

The most serious issue for small businesses in rural areas is the dependency on key individuals.

By feeding the tacit knowledge of veteran employees into AI, it can be made accessible to anyone. This alone dramatically reduces the risk of business continuity.

As a real example, a local manufacturing company (with 30 employees) organized the processing know-how (about 200 decision criteria) of veteran craftsmen into a ChatGPT-based chatbot. The construction period was two weeks, and the only cost was the monthly API fee of 2,000 yen. The onboarding period for new employees was reduced from three months to one month.

The Real Risk is “Not Using AI”

Finally, one last point.

“AI is for large corporations” and “it’s too early for us” — this mindset is the most dangerous.

What is happening now is not the democratization of AI, but rather the commoditization of AI. We are in a process where AI is becoming a “standard infrastructure,” just like water and electricity.

What large corporations have burned trillions of yen to create is now available for just a few thousand yen a month. This presents an unprecedented opportunity for small businesses.

Conversely, when competitors in the same industry start using this tool, companies that do not will find themselves in the same position as those “fighting with handwritten ledgers.”

While Google laments its lack of profitability, let’s seize the opportunity.

AI that is created by large corporations, discounted by large corporations, and provided at a loss by large corporations.

There’s no reason not to use it.