Conclusion

Let’s get straight to the point. The seismic shifts in the AI industry represent a “chance” for small and medium-sized enterprises (SMEs).

Anthropic’s valuation is nearing $900 billion (approximately ¥135 trillion), more than three times that of Toyota’s market capitalization. Investors are reportedly being urged to finalize their funding commitments within 48 hours. The pace is astonishing.

On the other hand, DeepSeek has opened its API at a nearly free price point, while Meta has tightened the terms of use for Llama, steering towards what can be described as “closed-source under the guise of open source.”

AI vendors are polarizing into three camps: Anthropic, which pursues performance with massive funding; DeepSeek, which aims to capture the market through price disruption; and Meta, which seeks to dominate the platform.

If any SME executives think, “This is a war among large corporations, so it doesn’t concern us,” they should reconsider. This polarization is precisely the structure that benefits SMEs the most.

What’s Happening—Three Key Facts

Fact 1: Anthropic’s $900 Billion is the Pinnacle of “Performance-Based Pricing”

Anthropic’s Claude 4 is delivering top-tier performance in coding, analysis, and long-form processing. The soaring valuation is due to enterprise companies starting to pay tens of thousands of dollars monthly for this performance.

However, there is a number that should bring some calm. The actual cost of the Claude API is $3 for input and $15 for output per million tokens (Claude 3.5 Sonnet). For the amount SMEs typically use in daily operations—such as handling 50 inquiries or document creation per day—the monthly cost would be around $50 to $200. There’s no need to be intimidated by that $900 billion valuation; it pertains to the world of investors.

Fact 2: Is DeepSeek’s “Free” Offer Genuine?

DeepSeek-V3 and DeepSeek-R1 have API usage fees that are extraordinarily low—just $0.27 per million tokens for input. That’s less than one-tenth of Claude’s cost. Moreover, since it is publicly available with open weights, running it on your own server means API costs can be zero.

You might think, “Free sounds suspicious.” That’s a valid concern. DeepSeek is a Chinese company, and there are risks regarding data handling and continuity. However, as a technical fact, since the model weights are publicly available, if you download and run it in your own environment, your data won’t be sent outside. Understanding this can significantly change your judgment.

Fact 3: The True Nature of Meta’s “Open Source Killing”

Meta’s Llama 4 is high-performing, but attention should be paid to its terms of use. A license is required if used in services with over 700 million monthly active users. Furthermore, using outputs from Llama to train competing models is also prohibited.

You might think, “We don’t have 700 million users, so it doesn’t concern us.” But that’s not the essence of the issue. Meta is gathering developers at the “entry point” of open source, making them dependent on its ecosystem, and then blocking the exit with its terms. This is a classic platform strategy. If SMEs build systems based on Llama and the terms change, the switching costs will skyrocket.

The True Nature of “Switching Costs”—They Can Actually Be Brought to Nearly Zero

Now we get to the main point.

When hearing “vendor lock-in” or “switching costs,” you might imagine transition costs in the hundreds of thousands of yen. That was true for SaaS transitions five years ago. However, in the context of AI use in 2025, the switching costs for SMEs can be designed to be nearly zero.

Why is that? There are three reasons.

1. API Input and Output Formats Are Almost Standardized

OpenAI-compatible API formats have effectively become the standard. Anthropic, DeepSeek, and Llama-based hosting services all provide OpenAI-compatible endpoints or can be adapted with thin wrappers. This means that rewriting code can be done in just a few lines.

2. Prompts Are Assets That Are Vendor-Independent

The true assets that SMEs should accumulate with AI are prompts (instructions) and workflows optimized for their business. These are text files that can be reused across any model. Even if you switch models, minor adjustments to the prompts can accommodate the change.

3. “Mixing and Matching” Has Become the Optimal Solution

The era of relying on a single vendor is over. The configurations we are actually implementing for our clients look like this:

In total, that’s ¥20,000 to ¥40,000 per month. This covers most of the tasks that previously cost ¥300,000 to ¥500,000 in outsourcing.

If switching costs are close to zero, there’s no need to fear lock-in. What’s truly scary is not being locked in, but “not using anything at all.”

Three Actions SMEs Should Take Today

The term “multi-vendor strategy” is correct but too abstract. What should you specifically do?

1. Try Hitting an API for One Task (This Week)

Whether it’s sorting invoices, drafting inquiry emails, or summarizing meeting minutes—anything will do. Automate just one task using either the Claude API or DeepSeek API. You can start for just a few thousand yen per month. Think of it as “removing one task” rather than “implementing AI.”

2. Manage Prompts with Git or Notion

Save the prompts used for each task in text format and manage versions. This will become an “asset that is independent of the vendor.” It will also eliminate dependency on specific individuals. If a person leaves, as long as the prompts remain, anyone can maintain the same quality of work.

3. Compare and Test Models Quarterly

Throw the same prompts at three different models and compare quality and cost. This can be done in half a day. The AI industry can change dramatically in just three months. It’s common for last year’s best to become this year’s worst. Stay flexible and continuously update for optimal solutions.

The Real Risk Is Not “Becoming More Expensive” but “Becoming Too Cheap to Make Judgments”

Finally, let’s look a little further ahead.

AI costs continue to decline. Two years ago, GPT-4 class performance was $30 per million tokens. Now it can be obtained for less than $1. This trend is not going to stop.

What happens when costs go down? The only differentiating factor will be “what to do with AI” rather than “which AI to use.”

Large corporations, due to their size, tend to have slower decision-making processes. “We will form a task force on AI utilization and consider it in next year’s budget”—that’s the speed they operate at. In contrast, SMEs can move as soon as the owner says, “Let’s do it.” This speed difference is the greatest weapon for SMEs.

Whether Anthropic is valued at $900 billion, DeepSeek is free, or Meta is killing open source, it doesn’t matter. The tools have become cheaper. The question is, what will you change by using them?

The answer lies only in your field.