Conclusion

To put it simply, small and medium enterprises (SMEs) no longer have a reason to continue using “big AI.”

In the fiscal year 2024, the combined CO₂ emissions from Microsoft, Amazon, and Google are expected to reach approximately 119 million tons, exceeding one-third of France’s annual emissions (around 300 million tons). This marks a roughly 20% increase from the previous year, primarily due to the explosive expansion of data centers powering AI.

Many people hear this figure and conclude, “Environmental issues are important.” However, the key point for SME owners lies elsewhere.

The rising costs of the infrastructure required to run “big AI” will eventually be passed on to users in the form of API fees and electricity bills.

In other words, this is not an environmental issue; it’s a matter of bills.

—

The Cost Structure of “Big AI” Is Starting to Break Down

It is said that the power required for a single inference of a large language model (LLM) like GPT-4 is about ten times that of a single Google search. OpenAI’s ChatGPT (GPT-4) is estimated to consume about 0.01 kWh per query. For SMEs that make tens of thousands of API calls each month, the server-side electricity consumption can reach several hundred kWh.

What does this translate to in terms of API costs?

The API pricing for GPT-4o is $2.50 per million input tokens and $10.00 per million output tokens (as of June 2025). In contrast, GPT-4 Turbo costs $10.00 for input and $30.00 for output. Although prices have decreased compared to the previous generation, if SMEs use it extensively for daily operations, monthly costs can easily reach tens of thousands of yen.

Handling inquiries, summarizing meeting minutes, generating reports, drafting emails — using GPT-4 class models for these “mundane but high-volume” tasks is like sending a trailer for deliveries that could be handled by a light truck.

—

The Turning Point Where Small Models Have Become “Sufficient”

The most significant change over the past year is that the performance of small models has surpassed practical thresholds.

Here are some specific names:

• Phi-3 Mini (Microsoft): 3.8 billion parameters. Runs on smartphones. Benchmarks show it is equal to or better than GPT-3.5 for simple summarization and classification tasks.
• Gemma 2 (Google): Even the 9 billion parameter version can run inference on a single laptop GPU.
• Llama 3.1 8B (Meta): 8 billion parameters. Open source. With fine-tuning, it can achieve accuracy close to GPT-4 for specific tasks.
• Qwen2.5 (Alibaba): The 7 billion parameter version performs highly accurately in Japanese. Commercial use is allowed.

These models can run on local PCs (with around 16GB of memory). This means that cloud API costs can be reduced to zero, with electricity costs only reflecting the power consumption of the PC. When calculated monthly, this could be just a few hundred to a few thousand yen.

A business that was paying 100,000 yen per month for the GPT-4 API could switch to a local small model for just 3,000 yen per month.

Such cases are already beginning to emerge, showing a cost reduction rate of 97%. This is not an exaggeration; it is a structurally occurring change.

—

A Practical Comparison of “AI Electricity Costs” for SMEs

Let’s consider a hypothetical local SME with 20 employees and estimate costs.

Use Case: Daily inquiry handling (50 inquiries per day), summarizing meeting minutes (5 per week), drafting internal reports (20 per month)

Pattern A: Using GPT-4o API

• Inquiry handling: Approximately 1,500 tokens per inquiry (input and output combined) × 50 inquiries × 20 business days = 1.5 million tokens per month
• Meeting minutes summarization: Approximately 4,000 tokens per document × 20 documents = 80,000 tokens per month
• Report drafting: Approximately 3,000 tokens per document × 20 documents = 60,000 tokens per month
• Total: Approximately 1.64 million tokens/month
• API cost: Approximately 2,000 to 3,000 yen/month (for GPT-4o)
• However, for GPT-4 Turbo, it would be around 15,000 to 20,000 yen/month

Pattern B: Local Small Model (e.g., Llama 3.1 8B)

• Initial cost: GPU-equipped PC approximately 150,000 to 250,000 yen (class with RTX 4060)
• Monthly electricity cost: Approximately 1,500 to 3,000 yen for PC operation
• API cost: 0 yen
• Running cost: Approximately 2,000 to 3,000 yen per month

You might think, “Wait, isn’t the monthly cost for GPT-4o not that different?” That’s correct; the latest APIs have become cheaper.

However, here’s the crucial point:

API prices are determined by providers, while local electricity costs are governed by physical laws.

API fees can increase. In fact, OpenAI has changed its pricing structure multiple times in the past. In contrast, the electricity costs for running small models locally continue to decrease due to hardware energy efficiency and model optimization.

For SMEs, whether they can control their “costs” is more important than the amount itself.

—

The Real Issue Is Not “Environment” but “Dependency Structure”

The fact that CO₂ emissions from the three big tech companies amount to one-third of France’s emissions — the essence of this news is not about environmental issues.

The structural problem is that AI infrastructure is concentrated in a few giant companies, creating costs, risks, and dependencies.

In response to the construction boom of data centers, local residents have begun opposing movements in various places. In the U.S., Virginia has seen issues with power supply, while in Ireland, power shortages have become a social problem. In Japan, data center clusters are advancing in Inzai City and Chitose City, and discussions about their impact on power infrastructure have begun.

What this trend signifies is the risk that the costs of using cloud AI may rise in the medium to long term. Rising electricity costs, the introduction of carbon taxes, and increased regulations — all of these will reflect back on API fees.

What SMEs should do is not merely a moral discussion of “let’s consider the environment,” but rather to have options to distance themselves from this dependency structure.

—

So, What Should Be Done Specifically?

I propose three steps:

1. First, “visualize” your company’s AI usage costs

Most companies do not know how much they are paying for which APIs or how many tokens they are using. Start here.

2. Sort out “tasks that don’t need GPT-4”

Standard responses to inquiries, summarizing meeting minutes, classifying data — these tasks often can be handled sufficiently by small models. There’s no need to replace everything. Just by distinguishing between “this task requires GPT-4 and this task can use a small model,” costs can be halved.

3. Try running one local model

Using tools like Ollama or LM Studio, you can run small models on your existing PC in less than 30 minutes. Start with one task. If the accuracy isn’t sufficient, you can revert to the cloud. The risk is zero.

—

“Small AI” Can Be a Weapon for SMEs

Large corporations invest trillions of yen in massive data centers and develop cutting-edge models in-house. Their way of competing is irrelevant to SMEs.

In fact, it’s the opposite.

While large companies become immobilized by the enormity of their infrastructure, SMEs can run small models locally and manage AI operations for just a few thousand yen a month. Data doesn’t leave their PCs, and they aren’t swayed by rising API prices. They can apply the necessary size of AI only to the tasks they need.

The era where bigger is better is coming to an end.

The news that CO₂ emissions from big tech have reached one-third of France’s emissions is not a story about “being bad for the environment” but a signal that “the cost structure of big AI is nearing its limits.”

The ones who can leverage this signal first will be the agile SMEs.

I hope they start by looking at this month’s API bill.