Same Week, Clear Divide

In July 2026, three significant news items emerged almost simultaneously in the AI industry:

• SenseTime announced its first annual profit since going public. The revenue for the first half of 2026 was approximately 3.8 billion yuan (about 76 billion yen), a 42% increase year-on-year, with an operating profit of about 230 million yuan (about 4.6 billion yen).
• Anthropic CEO Dario Amodei declared a “slowdown in AI development” and unveiled a three-stage safety plan that includes external audits.
• OpenAI‘s Sam Altman retracted the IPO plans for 2026, citing safety risks and uncertainties in the market environment.

Despite all being “AI companies,” one is profitable, another is slowing down, and the last has even postponed its IPO.

What accounts for this disparity? And what does this structure mean for local small and medium enterprises?

“General-purpose” Consumes Money, “Specialized” Generates It

To get straight to the point, the difference between profitable AI and unprofitable AI is not about being ‘general-purpose’ or ‘specialized.’ It hinges on whether it directly addresses customer challenges.

SenseTime’s profitability is not supported by large language models (LLMs). Instead, it focuses on AI that precisely targets specific industries and tasks, such as facial recognition, image recognition for autonomous driving, and medical image analysis.

Looking at the specifics, SenseTime’s revenue composition is as follows:

• Smart Business Division (for urban management and security): about 45% of revenue
• Smart Life Division (image processing for smartphones, etc.): about 25%
• Smart Car Division (autonomous driving support): about 20%
• Others (medical AI, etc.): about 10%

Each of these has a clear “use case.” Implementing companies do not hesitate about “what to use it for.” This is why it sells, and this is why it is profitable.

On the other hand, Anthropic and OpenAI are developing “general-purpose models” that can be used for anything. Claude and GPT—indeed impressive. However, “general-purpose” implies, conversely, “customers need to figure out what to use it for.”

OpenAI’s annual revenue is estimated to exceed $6 billion (about 900 billion yen), but its estimated costs exceed $8 billion. Even with massive revenue, the electricity costs for GPUs and computational resources surpass the income. Anthropic faces a similar situation, having received up to $8 billion in funding from Amazon while still not seeing a path to profitability.

$900 billion in revenue yet in the red. $76 billion in revenue and in the black.

These numbers alone illustrate the structural differences.

Behind the Slowdown Declaration Lies “We Can’t Sustain This”

Reading Anthropic’s slowdown declaration merely as a “consideration for safety” is too superficial.

Looking at the details of the three-stage plan proposed by Dario Amodei reveals the following:

1. Mandatory safety audits by external evaluators before releasing new models
2. Models that exceed a certain capability threshold will not be deployed without consultation with government agencies
3. Cooperative restraint on the development speed across the industry

This should not be interpreted as “slowing down for safety’s sake,” but rather as “having to slow down, with safety as a convenient justification.”

The reason is simple. The competition to develop general-purpose models has become a battle of endurance. The estimated computational costs required for training a model like GPT-5 range from $500 million to $1 billion. Each training run can cost hundreds of millions of yen. Moreover, as models grow larger, the performance improvements diminish (the limits of scaling laws). The returns on investment are worsening.

OpenAI’s withdrawal from the IPO can also be understood in this context. If they go public, they will be questioned about profitability every quarter. However, under the current cost structure, profitability is not in sight. Hence, the decision is made to “not go public for now.”

In other words, the game of “developing general-purpose AI” itself is nearing the limits of financial capacity.

Lessons for Small and Medium Enterprises

Now, let’s get to the main point.

SenseTime is a company with a market capitalization of several hundred billion yen, which is a different scale from local small and medium enterprises. However, the “structure of profitable AI” remains the same regardless of scale.

There are three key points.

1. Use General-purpose AI as “Components.” Don’t Build It Yourself.

APIs for GPT-4o and Claude 3.5 can be used for just a few yen per 1,000 tokens. This means that the “development costs of general-purpose AI” are being borne by other companies. What small and medium enterprises should do is to use these inexpensive components to create systems specialized for their own operations.

For example, a local manufacturing company did the following:

• Converted 10 years’ worth of defective product reports (about 3,000 cases) from PDF to text
• Designed a prompt to automatically classify defect causes using the GPT-4o API
• Automatically output the classification results to a spreadsheet
• Development costs: 1 in-house engineer for 2 weeks + API costs of about 8,000 yen per month

Previously, a quality control staff member spent 20 hours a month on this task, which has now been reduced to 30 minutes of verification work per month. A system that was estimated to cost 3 million yen to outsource is now running for less than 50,000 yen.

This is what it means to “use general-purpose AI as components.”

2. Decide “What to Use It For” First. Technical Selection Comes Later.

SenseTime was able to achieve profitability because there were specific challenges such as “the need for facial recognition on-site” and “the need for image recognition in autonomous driving.”

A common mistake among small and medium enterprises is to start with “I want to implement AI.” This is the wrong order.

The correct order is:

1. What is the most time-consuming task right now?
2. Within that task, which of judgment, classification, summarization, or generation is needed?
3. What APIs or tools fit that need?

Start with “business improvement” rather than “AI implementation.” AI is merely a means to that end.

3. Now That Large Companies are Slowing Down, It’s a Chance for SMEs.

With Anthropic slowing down and OpenAI postponing its IPO, what does this mean?

The pace of evolution for general-purpose AI will temporarily slow down.

This means that the performance of existing APIs (GPT-4o, Claude 3.5, Gemini, etc.) will stabilize at a level that is “sufficiently usable” for a while.

This is good news for small and medium enterprises. Because when technology is stable, they can focus on systematization.

During periods when models change dramatically every six months, there was a risk of anything created becoming obsolete. But now, it’s different. If you systematize your operations with the current APIs, you can use them for at least 1 to 2 years. This allows for a clearer outlook on return on investment.

While large companies are treading water, small and medium enterprises can move ahead as the “users.”

So, What Should Be Done?

Here are three actionable steps to take starting tomorrow:

1. Identify the most personalized task within your company. This could be know-how that exists only in the minds of veteran employees or manual judgment tasks. This is the top priority for AI implementation.

1. Test that task first with the free version of ChatGPT. Input the task details into the prompt and see how far the AI can go. Even if the accuracy is 70%, there are many cases where the remaining 30% can be supplemented by humans to make it practically usable.

1. Consider API integration that can be started for under 10,000 yen per month. A combination of Google Apps Script and the GPT-4o API can create a system for automating operations with zero initial costs and monthly fees of just a few thousand yen.

What SenseTime’s profitability teaches us is that “whether AI is profitable or not is determined not by the technology’s prowess but by its proximity to the challenges.”

Place AI as close as possible to your company’s challenges. Just by doing that, the use of AI in small and medium enterprises can shift from a “cost” to a “weapon.”

Utilize general-purpose AI, developed at a cost of several hundred billion yen by large companies, as components for just 8,000 yen a month. This is the correct way for small and medium enterprises to compete in 2026.