Conclusion First: The “Cost of AI” is Starting to Break Down

A year ago, using a GPT-4 class model via API for business purposes would easily cost tens of thousands of yen per month. Now, the assumptions are crumbling on a weekly basis.

Let’s list what has happened just in the past week:

• Meta has released the 816B parameter “Llama 4 Maverick” as open-weight and for free.
• DeepSeek has launched its latest inference model “R1-0528” and is offering the API at a low cost.
• Ollama has implemented a proxy-compatible update, lowering the barriers for local operation within corporate networks.

If there were only one of these developments, it might be dismissed with a shrug. However, all three have arrived simultaneously. This is not a coincidence. The price collapse of AI models is progressing structurally and irreversibly.

The question is simple: “What value is increasing, and what value is decreasing due to this change?”

—

Meta’s Free Release of the 816B Model—The End of the Assumption that “High-Performance AI Equals High Costs”

Meta’s released “Llama 4 Maverick” adopts a Mixture-of-Experts (MoE) architecture with 816B parameters. Not all parameters are active at all times; approximately 42B become active per token. This means it’s a massive model designed to keep inference costs low.

Let’s summarize the noteworthy specifications:

• Context Window: 1 Million Tokens (equivalent to approximately 500,000 to 700,000 characters in Japanese, allowing for the reading of 3 to 4 paperback books at once)
• Multimodal Input Support: Accepts text, images, video, and audio, and responds in text
• License: Released under the Llama 4 Community License, allowing commercial use (no additional license required for companies with fewer than 700 million monthly active users)

This is free. You can download it and run it on your own servers or cloud GPU instances.

Just six months ago, using a model of comparable performance via API would require a budget of 300,000 to 500,000 yen per month. Now, it’s possible to operate it with just the costs of GPU electricity and instance fees—around 30,000 to 50,000 yen per month. Costs have decreased by a factor of ten. This isn’t just a story of “becoming cheaper”; it’s about a structural change.

The implications for small businesses are even greater. Dependency on APIs has created a structure where costs balloon with usage, under the guise of pay-as-you-go. With local operation, costs can be fixed. Monthly budgets become predictable. This alone lowers the barriers to adoption.

—

DeepSeek’s Latest Model—”Cheap and Poor Quality” No Longer Applies

DeepSeek’s movements are also noteworthy. The latest inference model has achieved performance improvements of 2.0 to 11.1 points on major benchmarks compared to the previous version, particularly showing significant enhancements in mathematical reasoning and coding tasks.

However, the real impact lies not in performance but in price.

DeepSeek’s API pricing is approximately 1/10 to 1/30 of comparable models from OpenAI. It operates in a world where the cost for 1 million tokens of input is just a few dozen to a few hundred yen. For monthly business use, it’s not uncommon to see costs in the thousands of yen.

Moreover, DeepSeek has also released the model itself as open-weight. Not only is the API inexpensive, but it can also be run locally. There are double the options available.

The essence of what is happening here is that the value of high-quality AI models is rapidly decreasing. We are entering an era where models cannot be differentiated on their own.

So, what value is increasing? The answer is clear: the value of “your own business data” and “the mechanisms to feed it to AI” is rising. Models are accessible to everyone. The differentiation lies in what data is fed to the model, what outputs are generated, and how it is integrated into business workflows. This is where small businesses can move faster than large corporations, as decision-makers are closer to the front lines.

—

Ollama’s Proxy Support—Unassuming, but “This is What We Wanted” in the Field

Ollama is a tool that makes it easy to run LLMs in local environments. The recent update has added support for HTTP/HTTPS proxies.

You might think, “Proxy support? Is that all?” However, for those familiar with corporate environments, the significance of this change is immense.

Many small and medium-sized enterprises in Japan route their internet access through proxy servers due to internal network security policies. In many environments, communications that do not go through a proxy are blocked. Until now, attempting to download models with Ollama often resulted in being blocked by proxy barriers, leading to frequent cases where installation was simply impossible.

With this new support, it is now possible to download and operate models simply by setting the proxy in the environment variables. Even in small businesses with only one IT staff member, a local LLM environment can be set up in just 30 minutes.

Technically, it’s a minor change. But it directly impacts the adoption rate.

The barriers to AI adoption often lie not in model performance or costs, but in operational friction such as “not working on the internal network” or “not receiving approval from the security department.” Ollama’s support has reduced one of these frictions.

—

Overlapping the Three Movements Reveals a “Structural Change”

Looking at these three news items individually has little meaning. When viewed together, a structure emerges.

Costs are collapsing simultaneously across all three layers: models, APIs, and infrastructure.

This is not just a story of “AI becoming cheaper.” It’s about “the reasons not to use AI disappearing.”

—

What Should Small Businesses Do Now?

No need for abstract discussions. Let’s be specific with three points.

1. First, set up one local LLM environment.

Install Ollama and try running a 7B to 14B class model. No expensive GPUs are necessary. A gaming PC level (around 16GB VRAM) is sufficient. The investment can be as low as 50,000 to 100,000 yen for a used PC. This will provide the experience of having AI running in-house. You can’t formulate a strategy without experience.

2. Choose one of your company’s “tedious routine tasks” and feed it to AI.

Summarizing meeting minutes, drafting estimates, classifying inquiry emails—anything will do. The important thing is to run AI with “your own business data.” It’s a waste to just play around with a generic chatbot and say, “That’s impressive.”

3. Don’t spend too much time on “selecting AI models.”

New models are released weekly. While you’re comparing which one is the strongest, the next model will come out. Right now, deciding “which business to integrate AI into” is 100 times more important than “which model to use.” Models can be swapped out later. The design of business workflows is a race where those who start early will win.

—

What Happens After the Cost of AI Crumbles?

As costs decrease, barriers to entry lower. When barriers to entry lower, the axis of competition changes.

“Simply using AI” is no longer a differentiator. Anyone can use it. The differentiation lies in “how well you can systematize your business operations using AI.”

Reproducing personalized tasks with AI. Translating the judgment criteria of veteran employees into prompts. Turning tasks that took two hours manually into a state where they are “completed automatically.”

This is something small businesses can do faster than large corporations. There are fewer approvals, the distance to the front lines is shorter, and the mentality of “let’s try it first” is more prevalent.

With the cost of AI crumbling, this is a tailwind for small businesses. The question is whether they will recognize this wind and take action.

Tectonic shifts are happening weekly. By next month, today’s common knowledge will likely be outdated. If you’re going to act, now is the time.