Conclusion

Let’s get straight to the point. The variable that divides them is whether it is clear “whose work is being replaced, what work it is, and at what cost.”

OpenAI has an annual revenue of $3.7 billion, while its expenses are estimated to exceed $8.7 billion. It is burning approximately $5 billion each year. Plaud, on the other hand, automates meeting minutes with a small AI recorder and has surpassed an ARR (Annual Recurring Revenue) of $100 million. Probably has raised $9 million with technology that detects “AI lies.”

Despite both being “AI startups,” why is there such a stark contrast in their fortunes?

The answer is simple. It boils down to whether a company can succinctly state “whose work is being replaced, what work it is, and at what cost.” Companies that can articulate this earn revenue, while those that cannot burn cash. This single line serves as a critical criterion for small and medium-sized enterprises when choosing AI vendors.

—

OpenAI’s Structural Problem—”General Purpose” Is Not Profitable

Analyzing OpenAI’s financial documents leaked in 2024, the numbers are as follows:

• 2024 Revenue: Approximately $3.7 billion (more than double year-on-year)
• 2024 Expenses: Estimated to exceed $8.7 billion (GPU infrastructure, personnel costs, inference costs)
• Net Loss: Approximately $5 billion
• 2025 Revenue Forecast: Approximately $13 billion

Revenue is growing rapidly. However, as the number of users increases, inference costs (the GPU costs incurred each time the AI generates a response) also rise. When revenue doubles, costs balloon nearly as much.

Why does this happen? OpenAI’s ChatGPT is designed as an “AI that can do anything.” It can write code, compose poetry, and provide cooking recipes. However, “being able to do anything” means it is not designed as a tool to solve specific problems for specific individuals. With a subscription of $20 per month, inference costs range from a few to several dollars per user each month, resulting in thin margins.

This is an important insight for small and medium-sized enterprises as well. Implementing an “AI that can do anything” does not dramatically reduce the costs of specific tasks in the workplace. Ultimately, employees end up using it merely as a “convenient toy.”

—

Why Plaud Can Earn $100 Million—A Clear Replacement for “Meeting Minutes”

Plaud’s business model is extremely straightforward.

• Whose: Business professionals’
• What work: Replacing the process of recording meetings → transcribing → summarizing → creating meeting minutes
• At what cost: Replacing it with a subscription of a few thousand yen per month plus a dedicated device (about 20,000 yen)

It can be stated in one line.

The time taken to create meeting minutes averages 30 to 60 minutes for a one-hour meeting. For an employee earning 3,000 yen per hour, the cost per meeting ranges from 1,500 to 3,000 yen. For someone with three meetings a week, that amounts to 20,000 to 40,000 yen per month. With Plaud’s subscription of a few thousand yen, this task can be completed “automatically.”

The cost-saving effect can be calculated concretely. That’s why it sells.

Moreover, Plaud sells a dedicated small device. Hardware sales come first, followed by recurring revenue from subscriptions. This two-tier structure of initial sales plus ongoing revenue is effective. Although the breakdown of the $100 million ARR is not disclosed, it is believed that device sales recoup customer acquisition costs while subscriptions build profit.

The rise of remote work has exacerbated the problem of “too many meetings.” Plaud has hit that pain point directly.

—

Probably’s $9 Million—A Clear Challenge to Stop “AI Lies”

Probably is developing technology to detect and prevent factual inaccuracies (so-called hallucinations) generated by AI. They successfully raised $9 million in seed funding.

This can also be stated in one line.

• Whose: Companies wanting to incorporate AI into their operations
• What work: The task of “checking whether AI outputs are correct”
• At what cost: Replacing it with an automated verification tool

The first hurdle companies face when they start using AI in their operations is “output reliability.” Can they send a response generated by ChatGPT directly to a customer? Can they use a draft of a contract as is? The answer is No. Therefore, humans must check it. This checking cost is not negligible.

Probably reduces this checking cost. By inserting another layer of AI into the output, it automatically verifies consistency with facts. As AI becomes more widespread, the demand for this “reliability layer” will increase. It is a structurally growing position.

—

The Structure That Divides “Burners” and “Earners”

To summarize, it looks like this:

I do not mean to say that OpenAI is a bad company. It is in a phase that requires massive investment as a platform. However, when small and medium-sized enterprises choose AI vendors, relying on “OpenAI-type” vendors is a high-risk proposition.

Why? A vendor that continues to burn cash may raise prices at any time. They may discontinue services unexpectedly. They may pivot their strategy without notice. In fact, OpenAI has changed its pricing structure multiple times over the past year.

—

The “5-Second Check” for Small and Medium-Sized Enterprises When Choosing AI Vendors

There’s no need for a complicated checklist. Just ask three questions.

1. “What specific task in our operations will be replaced, and at what cost?”

Vendors that cannot answer this with concrete numbers can be eliminated from consideration. Vendors that respond with buzzwords like “operational efficiency,” “DX promotion,” or “productivity improvement” do not understand what they are selling themselves.

2. “Is that vendor able to sustain itself with its own revenue?”

Do not be swayed by the size of the funding. “We raised 1 billion yen!” is synonymous with “We plan to burn 1 billion yen.” What matters is the structure of revenue and profit. For publicly traded companies, check their financial statements. For private companies, ask about the “retention rate of existing customers.” A retention rate of over 90% is a pass; below 70% is a red flag.

3. “If we want to stop, can we?”

Can you export your data? Are you locked into a proprietary format? Can you switch to other tools via API integration? Will your business operations stop if the vendor goes under? Too many small and medium-sized enterprises fail to verify this.

—

How This Relates to Local Small and Medium-Sized Enterprises

You might think, “This doesn’t concern us; it’s about OpenAI or Plaud.” But the structure is the same.

There are also vendors in local areas claiming, “We will transform your business with AI.” It is essential to discern whether that vendor is a “burner” or an “earner.”

The way to discern is simple. Just ask, “Our accounting month-end process currently takes 30 hours; how many hours will it take with your solution?” Vendors that can respond with specific numbers can be trusted. Vendors that can only say, “Let’s start with a PoC (Proof of Concept)” just want to experiment in your company.

AI costs are dramatically decreasing. The inference cost of GPT-4 has dropped to less than one-tenth in a year. This means that an AI implementation that cost 3 million yen a year ago could now potentially be done for 300,000 yen. As costs decrease, opportunities are also arising for small and medium-sized enterprises.

That’s why you don’t want to make mistakes in vendor selection. The criterion is one.

“Whose work is being replaced, what work is it, and at what cost?”—Can this one line be articulated?

Partner with vendors that can articulate it. Run away from those that cannot. That’s all there is to it.