Conclusion

Let’s get straight to the point: “Don’t fall behind in AI” is the most dangerous mindset.

TSMC has raised its sales forecast for 2025 due to the relentless demand for AI semiconductors. They expect a 30% year-on-year increase in AI-related sales by 2026.

However, the performance improvements of AI models are slowing down. The evolution from GPT-4 to GPT-4o was dramatic, but beyond that, the differences are only marginally smarter. The growth in benchmark scores has clearly decelerated.

In other words, the structure looks like this:

Chips continue to sell. Infrastructure investments are ballooning. But performance growth is sluggish.

For large corporations, this might just be a matter of “preemptive investment.” However, for small and medium enterprises (SMEs), it could become a fatal trap. The moment they get swept up in the atmosphere of “don’t fall behind in AI” and make investments beyond their means, they are left with “GPUs they can’t utilize” and “ballooning monthly bills.”

The Reality of AI Investment in Large Corporations: “Starting Without a Clear Plan”

A survey by Diginoctica involving 107 companies clearly illustrates the distortion in this structure:

• Only 21% of companies are fully operational with AI.
• The majority have GPU utilization rates of less than 50%.
• Yet, spending on AI infrastructure is accelerating year-on-year.

In short, even large corporations are in a state of “buying infrastructure without knowing what to use it for.” A GPU utilization rate of less than 50% means that more than half of the purchased computing resources are idle. This translates to hundreds of millions of yen in infrastructure investment, with hundreds of millions going to waste.

This “calculation gap” signifies that the economics of AI have yet to be proven. Large corporations can endure this, but if SMEs attempt the same, they will run out of cash.

This is the turning point: Do they imitate large corporations or turn the structure to their advantage?

The Slowdown in Performance Improvement is a “Tailwind” for SMEs

Paradoxically, the slowdown in AI performance improvement presents an opportunity for SMEs. There are three reasons for this.

1. The period during which “existing models” can compete effectively has lengthened.

In an era where performance changes monthly, companies needed the stamina to chase the latest models. However, the gradual evolution means that workflows built on current models like GPT-4o or Claude 3.5 are likely to remain usable six months or a year from now.

For SMEs, this means that “once a system is established, it will last for a while.” The payback period for investments has become more predictable.

2. The points of differentiation have shifted from “model performance” to “data and business design.”

As the performance differences between models narrow, the deciding factors will be “what data is available” and “how it is integrated into business processes.” This is an area where SMEs have an advantage over large corporations.

Large corporations can take six months to coordinate between departments. In an SME, if the president says, “Let’s do it,” they can start moving as early as next week. Those who know the on-ground workflows intimately can directly design how to use AI. This speed and resolution are something large corporations cannot replicate.

3. API costs continue to decline.

While demand for semiconductors is increasing, API usage fees are actually decreasing. The API pricing for OpenAI’s GPT-4 has dropped significantly compared to a year ago. For GPT-4o mini, the cost is about $0.15 per million tokens of input, which is approximately 22 yen. Considering that GPT-4 was $30 (about 4,500 yen) per million tokens a year ago, this represents a reduction to 1/200th.

In other words, TSMC is thriving because large corporations are “stockpiling infrastructure,” while the costs for SMEs to “use” APIs have actually decreased dramatically. Whether one recognizes this asymmetry can completely change their strategy.

Three Things SMEs Should Do Now

1. Don’t buy your own GPUs. Stick to “pay-per-use” with APIs.

While large corporations are buying GPUs and struggling with utilization rates below 50%, SMEs can simply use what they need on an API basis.

For a monthly fee of a few thousand to tens of thousands of yen, they can access functionalities like natural language processing and image recognition that once cost hundreds of thousands of yen. There’s no need to own servers. Utilize cloud AI services on a pay-as-you-go basis. This alone reduces the risk of infrastructure investment to nearly zero.

Specifically, major models like OpenAI API, Claude API, and Google Gemini API can all be accessed via APIs. You can start on a scale where the monthly usage fee is around 10,000 yen.

2. Start with “automating one business process” rather than “company-wide AI implementation.”

A common mistake SMEs make is launching an “AI implementation project” aiming for company-wide transformation. This is a mindset suited for large corporations and is not appropriate for SMEs.

What they should do is to automate one business process at a time.

For example:

• Automate the daily summary of reports that takes 30 minutes → Save 10 hours a month.
• Generate drafts for estimates based on past data using AI → Reduce from 20 minutes per case to 5 minutes.
• Automatically classify and draft responses for inquiry emails using AI → Tripling response speed.

These “small automations” should be built up one by one. Once one is successful, move on to the next. Even if there’s a failure, the loss is small. This cycle of “experiment → validate → expand” is the greatest weapon for SMEs.

3. Focus on “tasks that are dependent on specific individuals” for AI implementation.

In SMEs, there are always tasks that can only be performed by certain individuals. The decision-making criteria that exist only in the minds of veteran employees, the know-how cultivated over years of experience—this is the essence of dependency on specific individuals.

AI can be used to eliminate this dependency. By translating the veteran’s decision-making criteria into prompts and feeding past response histories as data, anyone can replicate a “70-point decision.” While achieving a perfect score may be impossible, if a 70-point decision can be generated automatically, humans can supplement the remaining 30 points.

This is not about “reducing personnel.” It’s about ensuring that “the operation continues even if that person is absent” and “new hires can become productive in three days.” For SMEs, this addresses a more pressing issue than hiring difficulties.

“So, what should we do in the end?”

The answer is simple.

Now that large corporations are pouring money into infrastructure, SMEs should focus on being the “users.”

Large corporations are buying GPUs, building data centers, and hiring AI talent with salaries of 20 million yen. Thanks to this, API performance has improved, costs have decreased, and usability has increased. SMEs should simply reap the benefits of this.

The slowdown in performance improvement means that “the systems built with current AI won’t become obsolete for a while.” In other words, now is the best time to invest in “systematization.”

With a monthly API cost of 10,000 yen and someone in-house spending a few hours a week experimenting with its use, that’s all it takes to get started. Before commissioning a 3 million yen system development, first, try it out yourself.

While large corporations struggle with the “calculation gap,” SMEs can “test small and iterate quickly.” This is the survival strategy in an era of performance ceilings and rising costs.

I want to ask: What routine tasks in your company take more than 30 minutes every day? Why not let AI take care of that starting next week?