Conclusion

Let’s get straight to the point: “The cost of AI APIs will decrease further.”

Chinese AI companies have begun to transition from a phase of burning cash in investments to a profitable business model. This structural change brings direct benefits to Japanese SMEs, increasing API options and further lowering prices.

SenseTime is set to achieve its first profitable year in the fiscal year 2024. Revenue from generative AI has surged approximately 3.6 times year-on-year, reaching a total revenue of about 3.6 billion yuan (approximately 72 billion yen). Meanwhile, Kimi (Moonshot AI) aims for annual sales of $2 billion (approximately 300 billion yen). Both companies have shifted from a phase of “burning cash on R&D” to one of “earning and reinvesting.”

Let’s break down what this change means from the perspective of local SMEs.

—

Reasons Chinese AI Companies Have Become Profitable — “Low Prices” as a Competitive Advantage

Why have Chinese AI companies become profitable? The reason is simple: a cycle has begun where they are used because they are overwhelmingly cheap.

DeepSeek’s API is offered at a price point that is less than one-tenth of the input token cost compared to OpenAI’s GPT-4o. SenseTime’s “SenseNova” series has also been deployed at extremely low prices for the domestic Chinese market. Moonshot AI’s Kimi is experiencing a surge in monthly active users in China, leveraging its strength in processing long-context texts (up to 2 million tokens).

They are used because they are cheap. They gather data because they are used. The models improve because data is collected. They are used even more because the models improve. This positive cycle has finally led to profitability.

It is noteworthy that profitable companies are likely to lower prices further rather than raise them to capture market share. When they were operating at a loss, there was a risk of “not knowing when prices would increase.” Profitable companies can strategically maintain low prices. This means that the risk of API selection has decreased for SMEs.

—

The “Three Options” Facing Japanese SMEs Today

So, how can SMEs looking to integrate AI through APIs organize their options? There are three main choices.

1. OpenAI / Anthropic (Major US Players)

• Approximately $2.5 per million input tokens and about $10 for output with GPT-4o
• Claude 3.5 Sonnet is in the same price range
• High performance in Japanese and abundant documentation
• Choose here for peace of mind and quality

2. Chinese Model APIs (DeepSeek, Moonshot AI, SenseTime, etc.)

• Approximately $0.27 per million input tokens with DeepSeek V3 (about one-tenth of OpenAI)
• Kimi excels in long-text processing, suitable for summarizing manuals and contracts
• Choose here if cost is the top priority. However, confirming data handling policies is essential.

3. Open Weight Models (Llama, Mistral, Qwen, etc.)

• No API usage fees. Operate on in-house servers or cloud GPUs
• Y Combinator is supporting the establishment of labs specializing in open-weight AI, rapidly developing the ecosystem
• Initial setup incurs engineering costs, but monthly variable costs can be fixed
• Consider if monthly API costs exceed 100,000 yen.

Importantly, there is no need to “decide on just one” option.

For example, use DeepSeek’s inexpensive API for internal meeting minutes summarization, GPT-4o for generating client reports, and run an open-weight model locally for highly confidential contract reviews. Such “usage-based differentiation” is the most cost-effective for SMEs.

In fact, for a local manufacturing company we support (with 30 employees), the API costs previously incurred by relying solely on GPT-4, which were about 150,000 yen per month, have been reduced to about 40,000 yen per month by using DeepSeek in conjunction. They use GPT-4o only for processes requiring quality, while leaving drafts and summaries to DeepSeek. This alone results in annual cost savings of over 1.3 million yen.

—

Will Profitable Chinese AI Companies Fully Enter the Japanese Market?

Now, let’s look ahead to the main topic.

Both SenseTime and Moonshot AI currently derive most of their revenue from the domestic Chinese market. However, what do profitable companies do next? They expand overseas.

DeepSeek has already made its API globally available, and it can be used easily from Japan. Alibaba’s Qwen is also spreading worldwide as an open-weight model. There is a strong possibility that SenseTime will follow the same path.

If this happens, it will lead to a positive cycle for Japanese SMEs, where “options increase further and prices decrease even more.” On the other hand, there is also the problem of having too many choices and not knowing what to select.

That’s why it’s crucial to organize “where in your business to use AI, what level of quality is needed, and how confidential the data is” now. If this is done, it will be easier to make switching decisions when new models or APIs emerge.

—

What SMEs Should Do Now — Just Three Things

No need for abstract discussions. What should be done specifically?

1. Understand the API costs of the AI you are currently using (or planning to use)
How much are you paying per month, and how much does it cost per request? Too many companies claim to have “adopted AI” without knowing this.

2. Divide usage into three categories

• High confidentiality → Operate open-weight models locally
• Quality is important → OpenAI / Anthropic
• Cost priority → Chinese models like DeepSeek

Just this categorization can often reduce costs by more than half.

3. Design to minimize “switching costs”
If you abstract the API calling parts, you can switch to new models by simply changing the prompts and endpoints when they come out. A design that avoids vendor lock-in is vital for SMEs.

—

Summary — The Decrease in AI Costs is Just the Beginning

SenseTime’s profitability indicates that Chinese AI companies have transitioned from a phase of “burning cash on investments” to one of “earning and operating.” Moonshot AI’s target of $2 billion annually is in the same context.

This structural change will lead to further declines in AI API prices and diversification of options. This is more beneficial for SMEs than for large corporations. Large companies already have their own AI infrastructure, but as API prices decrease, SMEs will be able to achieve the same results with less investment.

What used to cost 3 million yen for outsourced business analysis can now be automated with API costs of 50,000 yen per month. Such a world is already within reach.

The question is whether to continue paying high costs without realizing “it has become cheaper” or to stop thinking and assume “cheap equals suspicious.”

The answer is simple: First, try it. Use it on a small scale. Compare the numbers.

SMEs that can do this will benefit the most from this structural change.