Conclusion

To put it simply, the “cost of using AI” has begun to structurally collapse.

Not just from one direction, but from three directions simultaneously.

• Hardware: A model with 1.6 trillion parameters has been trained without Nvidia.
• Pricing Model: An API with unlimited tokens for $6 a month has emerged.
• Supply Chain: Western companies like Proton have started adopting 100% domestically produced LLMs from China.

All three of these developments are happening at the same time. This is not merely a story of “AI becoming cheaper.” The very cost structure of AI is on the verge of being replaced by something entirely different.

What does this mean for small and medium-sized enterprises? Let’s examine each aspect in turn.

—

1. 1.6 Trillion Parameters Without Nvidia—The “Sanctuary” of Hardware Has Collapsed

China’s Meituan has announced “LongCat-2.0,” a large model trained with over 35 trillion tokens and 1.6 trillion parameters.

What’s noteworthy is not the specifications, but the fact that it was trained without using any Nvidia GPUs.

Instead, it utilized domestically produced AI ASICs (application-specific integrated circuits) in a super pod. Faced with U.S. export restrictions preventing access to Nvidia’s cutting-edge chips, China has taken serious steps to create its own solutions, resulting in this development.

The key consideration here is not whether the technology is impressive. What happens when the assumption that “Nvidia is the only option” collapses?

Until now, the majority of the costs associated with AI training and inference have been tied to the procurement costs of Nvidia GPUs (like the H100 and A100). The H100 costs around 4 million yen each, and thousands are needed for large-scale training, leading to a hardware cost in the billions of yen.

When this monopoly structure collapses, what happens? A competition for GPU procurement emerges, training costs decrease, and the benefits ripple through to API pricing. This is a basic market principle. However, until now, that “basic principle” had been halted by Nvidia’s monopoly.

The arrival of LongCat-2.0 signifies that the dam has burst.

—

2. Unlimited API for $6 a Month—The Conventional Pricing Model Has Been Blown Away

The second collapse is more direct.

An LLM API has emerged with a monthly fee of $6 and no token limits.

To illustrate just how unusual this is, let’s compare the numbers.

Using OpenAI’s GPT-4o via API costs about $2.5 for every 1 million input tokens and about $10 for every 1 million output tokens. If you were to use 5 million input tokens and 2 million output tokens in a month, that would amount to approximately $32.5 per month. While this is already quite affordable, depending on usage, it can easily exceed $100 a month.

Now, it’s $6 a month for unlimited use.

This is 1/5 to 1/50 of the traditional costs.

Of course, it’s necessary to scrutinize whether the model’s performance is on par with GPT-4o. We should also verify the realities of latency and rate limits. However, what’s crucial here is not whether the performance is identical.

“There are plenty of use cases in small and medium-sized enterprises that can be sufficiently met with a $6 monthly fee.”

Drafting inquiry emails, summarizing daily reports, formatting quotes, organizing meeting minutes—these tasks do not require GPT-4o-level performance. A reasonably capable model that can be used without limits is sufficient.

Previously, even a “reasonably capable model” could cost tens of thousands of yen per month on a pay-per-use basis. This is why many small and medium-sized enterprises hesitated, thinking, “I want to try AI, but I can’t predict the costs.” The flat-rate $6 model completely removes that psychological barrier.

The fear of “What if I use too much?” is eliminated. This may seem subtle, but it represents a decisive change in the decision-making process for implementation in the field.

—

3. Proton Adopts Chinese-Made LLM—The Diversification of the Supply Chain Has Begun

Another significant development is that the Swiss privacy company Proton has adopted a 100% Chinese-made LLM for its services.

This is not simply a matter of “Chinese AI being excellent.” It signals that “options other than American-made solutions are beginning to become viable at a practical level.”

Until now, the LLM API market has been effectively dominated by three companies: OpenAI, Anthropic, and Google—all American firms. When options are limited, price competition is hard to come by.

Now, Chinese companies are entering the fray. DeepSeek, Qwen, and the recently launched LongCat-2.0 are starting to catch up in terms of performance and are overwhelmingly cheaper. The API price of DeepSeek-V3 is said to be about 1/10 that of GPT-4o.

As the supply chain diversifies, the power to set prices shifts to the users. This represents a highly advantageous structural change for small and medium-sized enterprises.

—

So, What Should Small and Medium-Sized Enterprises Do?

It’s not enough to simply observe these three collapses and say, “That’s impressive.”

Here are three actionable steps you can take starting today.

1. Consider Whether API Costs Can Be Converted to Fixed Costs

With services like the unlimited API for $6 a month emerging, the reasons for being bound to pay-per-use pricing are diminishing. By inventorying your company’s use cases and categorizing them as “This can be handled by a low-cost model’s flat-rate API” or “This absolutely requires GPT-4o,” you could potentially reduce your monthly API costs by more than half.

2. Check Whether You Are Overly Dependent on Vendors That Assume Nvidia

If the vendors providing AI tools and services are 100% reliant on Nvidia infrastructure, their cost structures may change in the future. Choosing vendors that support multiple infrastructures can lead to future cost savings.

3. Experiment Now That the Cost of “Trying It Out” Has Dramatically Decreased

In the past, experimenting with AI required a commitment of tens of thousands to hundreds of thousands of yen per month. Now, you can start for just $6 a month. That’s $72 a year. About 10,000 yen. Just the cost of one night out.

Instead of spending time debating in meetings whether to implement AI, try it out in one area of your business first. Whether it’s automatic summarization of meeting minutes or drafting email replies, it doesn’t matter. In an era where you can experiment for 10,000 yen a month for a year, being “in the consideration phase” is nothing but an opportunity loss.

—

The Essence Is Not “Costs Have Decreased” but “The Cost Structure Has Changed.”

Let’s summarize once more.

• Collapse of Hardware Monopoly → Decrease in Training Costs → Decrease in API Prices
• Emergence of Flat-Rate APIs → Liberation from Pay-Per-Use Constraints → Elimination of Barriers to Adoption for Small and Medium-Sized Enterprises
• Diversification of the Supply Chain → Intensification of Price Competition → Shift of Pricing Power to Users

These three are not independent phenomena; they are accelerating each other. Models trained cheaply on chips other than Nvidia are being offered as flat-rate APIs, which companies around the world are beginning to adopt as options. This cycle has begun.

By the end of 2025, it would not be surprising if the AI usage costs for small and medium-sized enterprises are reduced to 1/3 to 1/5 of current levels.

The issue is not whether AI will become cheaper; it already is.

The question is whether you will recognize this change and take action or continue to say, “It’s still too early.” In a world where there is a $6 API, saying, “We will consider implementing AI next year” no longer holds water.