The Choice Between Paying 30,000 Yen Monthly or Owning “Your Own AI” for 100,000 Yen

How much are you paying monthly for the ChatGPT API?

Even for a company with 10 employees, if you properly integrate cloud AI into your operations, you can expect to spend around 30,000 to 50,000 yen per month. That amounts to 360,000 to 600,000 yen annually, or 1.8 to 3 million yen over five years. Moreover, during that time, the pricing structure can change multiple times, leaving you at the mercy of sudden price hikes. Your company’s data is continuously being uploaded to the cloud every day.

On the other hand, what about the current local AI — that is, AI that operates on your own PC without using the cloud?

To cut to the chase, an AI can run for zero yen per month with a 100,000 yen PC and a free open-source model. It’s already happening. This is the reality of 2025.

This article will not just conclude with “local AI sounds fun.” We will delve into what changes for small and medium enterprises, and what steps they should take.

—

Facing the “Invisible Costs” of Cloud AI Dependence

The issues with cloud AI go beyond just monthly fees. There are three “invisible costs” to consider.

1. Risk of Data Leakage
In 2024, Microsoft CEO Satya Nadella stated, “All AI models should be considered compromised.” By using cloud AI, your company’s customer information, estimates, and know-how are sent to external servers. For small and medium enterprises, information leakage can be a fatal blow to their credibility. While large corporations might manage with a public apology, local SMEs cannot afford such luxuries.

2. Vendor Lock-in
OpenAI changed its API specifications. Google revised its pricing structure. Each time, there are costs associated with adapting. When your business processes depend on another company’s service design, it’s akin to relinquishing control over your management.

3. A Structure That Becomes More Expensive the More You Use It
Cloud AI is fundamentally pay-per-use. The more deeply it is integrated into operations, and the more results it produces, the higher the costs grow. This creates a contradictory structure for small and medium enterprises where “the more successful you are, the more burdensome it becomes.”

Local AI flips all three of these on their head.

—

What Can 100,000 Yen Achieve? — Specific Specs and Performance

“Local AI is ultimately just inferior, right?”

That was the common belief until a year ago. It’s no longer the case.

As of 2025, open-source large language models (LLMs) such as Meta’s Llama 3.1, Google’s Gemma 2, and Mistral’s Mixtral are available for free. By running these with free tools like Ollama or LM Studio, you can have an AI on your company’s PC that performs similarly to ChatGPT.

Here’s a guideline for the necessary specs:

Even without a GPU, a model with 7B to 8B parameters can operate at a practical level with 32GB of RAM. Tasks like summarizing meeting minutes, drafting emails, creating manuals, and handling FAQs — 80% of daily operations can be sufficiently managed with this class of AI.

In other words, an initial investment of 100,000 to 150,000 yen and nearly zero monthly costs. Compared to paying 500,000 yen annually for cloud AI, you can recoup your investment in the first year.

—

“Learning by Doing” is the Ultimate Implementation Strategy

The Verge described the experience of local AI as “fun, overwhelming, and frustrating.” This is actually a remarkably accurate depiction.

Using cloud AI is an experience of “utilizing a service.” Local AI is an experience of “creating your own tools.” This difference is significant.

Let me share an experiment conducted by our team.

We placed a used PC (purchased for 60,000 yen) with Ollama installed in a local construction company (15 employees). For the first week, they used it solely for “summarizing daily reports.”

The results were as follows:

• Time spent summarizing daily reports: reduced from 15 minutes to 2 minutes per report
• Time saved on administrative tasks for site supervisors: approximately 40 minutes per day
• Monthly reduction of about 13 hours of labor
• Calculating at 2,000 yen per hour, this translates to a monthly saving of 26,000 yen in labor costs

The 60,000 yen PC recouped its investment by the third month. Moreover, no data was sent outside. Daily reports often contain client names and amounts, making this a significant advantage.

The key was starting with the small task of “summarizing daily reports.” If you jump straight into “AI-driven management reform,” the field staff may be hesitant. Start by eliminating one tedious task in front of you. That’s all it takes.

—

The Structural Advantage for Small and Medium Enterprises

Now, let’s get to the main point. Here are three reasons why local AI is not just a cost-cutting tool for small and medium enterprises, but a weapon for structural reversal.

1. The Multiplication of “Company Data × AI” is Faster for SMEs

For large corporations to implement AI, they must go through security reviews, internal approvals, proofs of concept, vendor selection… it takes at least six months. For small and medium enterprises, if the president says, “Let’s do it,” they can be up and running by next week.

Moreover, local AI can directly utilize your company’s operational data. Data from 20 years of estimates, past customer interaction histories, and unique know-how. These cannot be shared externally, but can be safely utilized locally.

While large corporations may produce an 80-point response using generic cloud AI, small and medium enterprises can generate a “95-point response tailored to their operations” using local AI. This is the structure of reversal.

2. Eliminating Dependency Becomes a “System”

The biggest challenge for local small and medium enterprises is the issue of “the business cannot operate without that person.” When veteran employees retire or fall ill, operations can come to a halt.

By feeding local AI with internal manuals, past case studies, and the judgment criteria of veterans through RAG (Retrieval-Augmented Generation), you can somewhat recreate the “knowledge in the heads of veterans.” It doesn’t have to be perfect. Just changing the state from “new employees don’t know who to ask first” to “they can generally ask the AI” can drastically change how the field operates.

Trying to achieve this with cloud AI would mean sending confidential internal information to external servers. With local AI, you can safely input the know-how of veterans. This difference is decisive.

3. Companies That Can Use AI Win in Recruitment

For local small and medium enterprises, recruitment is a matter of survival. Just being able to say, “Our company is improving efficiency with AI” can change the response from young talent. In fact, I have heard multiple reports of increased applications from local companies that have implemented AI tools.

Moreover, with local AI, you can create a narrative of “we are running our own AI.” This is fundamentally different from simply logging into a cloud service.

—

Realistic Considerations for Implementation

It would be irresponsible to only hype this up without addressing the risks honestly.

There are performance limitations. There are physical constraints on the size of models that can run locally. Reproducing the performance of a GPT-4o class model locally is currently difficult with hardware available to individuals or small businesses. However, most daily operations do not require the performance of GPT-4o. It’s crucial to determine whether you aim for a perfect score or if an 80-point performance is sufficient.

Some technical knowledge is required for initial setup. While Ollama is relatively straightforward, there will still be instances where you need to interact with the command line. Having someone in-house who is IT-savvy or an external supporter can be reassuring. However, the user interface has dramatically improved compared to a year ago. This barrier is rapidly decreasing.

Decide what to use it for before implementation. “Let’s just install AI for now” will lead to failure. First, identify a specific task that is considered tedious and target that precisely. If successful, you can then expand to other areas. Following this order is essential.

—

So, What Should You Do?

1. This week’s task: Install Ollama on one PC. It’s free and takes about 30 minutes.
2. This month’s task: Choose one document task that is considered the most “tedious” in your company and let AI handle it.
3. Goal for three months from now: Quantify the time and cost saved. This will serve as the basis for your next investment decision.

The essence of local AI is not about technology. It’s about the management decision of “having our own tools that we control ourselves.”

Will you continue to pay the subscription fees for cloud AI every month, constantly sending your data out and being at the mercy of vendors? Or will you invest 100,000 yen to acquire your own AI, optimizing your operations with your own data?

I believe the answer is clear.

Start with one PC. Start with one task. I encourage you to begin this week.