A $399 Robot, $500 Monthly Edge AI, and $6,900 LLM Training—Three Costly Mistakes Small Businesses Make in an Era Where They Don’t Need to Buy ‘Expensive AI’

AI costs are collapsing.

You can buy a bipedal robot for $399. For $500 a month, you can run an LLM locally. With $6,900 for GPUs, you can train a decent model.

Three years ago, this would have been a conversation with a price difference of two orders of magnitude.

Nevertheless, local small businesses are repeatedly making the mistake of choosing the “more expensive” option. Why? Because of a sense of security. “It’s from a well-known vendor,” “It comes with support,” “Everyone is using it.” They are paying hundreds of thousands of yen annually for that sense of security.

Is that really okay?

This time, we will dissect the common mistakes small businesses make through specific cost comparisons using three examples: Hugging Face’s $399 robot “Microduck,” local LLM operations, and low-cost model training.

—

Mistake 1: Not Trying the $399 Robot Before Buying a $50,000 One

The “Microduck” announced by Hugging Face is a small bipedal robot priced at $399 (about 60,000 yen). It can be trained by users through reinforcement learning and is controlled by a neural policy that learns actions such as walking, sitting, kicking, and skating in a physics simulation.

On the other hand, what about commercial robots for industrial or educational use? Excluding Boston Dynamics, even bipedal robots for research and education typically range from $5,000 to $30,000. Including customization and prototyping with 3D printing, it’s not uncommon for the total to reach tens of thousands of dollars.

$399 vs. tens of thousands of dollars. A price difference of over 50 times.

Of course, the Microduck is not something you can directly deploy on an industrial line. But that’s not the issue.

When small businesses start experimenting with robotics or AI, they often try to buy a “production-ready” solution right from the start. This is the essence of the mistake. Understanding the mechanism of reinforcement learning for $399 allows them to verify whether robotics fits their business. If it doesn’t, they only incur a tuition fee of 60,000 yen. If it does, they can then make a substantial investment.

“Experiment for 60,000 yen first” or “Implement for 3 million yen right away.” The difference in this decision can determine the survival of small businesses. Large companies can absorb a 3 million yen failure. A company with 20 employees cannot. Therefore, small businesses should be the ones benefiting the most from an era where low-cost experimentation is possible.

Yet, due to the misconception that “cheap options are unreliable,” they jump at high-priced solutions. They feel reassured by vendor sales materials. This is mistake 1.

—

Mistake 2: Continuing to Pay Tens of Thousands of Yen Monthly for External APIs and Handing Over Data

“We use the ChatGPT API,” “We have an external AI chatbot service.”

There’s nothing wrong with that in itself. However, I want to ask: How much are you paying per month, and do you know where your data is going?

The subscription costs for external AI platforms typically range from 50,000 to 500,000 yen per month, depending on usage. For a company with dozens of employees using it across multiple departments, it’s easy to exceed 200,000 to 300,000 yen per month. That amounts to 2.4 to 3.6 million yen annually. After three years, it can reach 10 million yen.

On the other hand, the option of running an LLM locally has become realistic. Recent evaluations show that if you have a high-spec personal PC (with a GPU costing around 300,000 to 500,000 yen), you can run open-source LLMs (like Llama 3, Mistral, Phi-3, etc.) locally, and even including monthly electricity and maintenance costs, the running cost can be kept around 50,000 yen per month.

300,000 yen per month vs. 50,000 yen per month. A difference of 3 million yen annually.

Moreover, with local operations, customer data does not leave the company. If you send it to an external API, customer information, internal know-how, and sales data included in prompts go through external servers. Even if the terms of service state that “it will not be used for training,” the fact remains that data is leaving the company.

For local small businesses, trust with customers is the foundation of their business. Is there a CEO who can say, “We were sending customer information to external AI”?

Of course, local LLMs are not omnipotent. There are performance differences compared to GPT-4o or Claude 3.5. However, for tasks like internal FAQ responses, meeting minutes summarization, template email generation, and manual searches—these “80% sufficient tasks” can be adequately handled by local LLMs. The choice for small businesses is clear: pay 300,000 yen for 100% or keep it at 50,000 yen for 80%.

“It’s easier with external APIs,” leading to a halt in thinking. This is mistake 2.

—

Mistake 3: Assuming Only Large Companies Can Train Models

“Training an LLM? That’s impossible for a company like ours.”

This might be the biggest misconception.

Indeed, training a model at the level of GPT-4 from scratch can cost tens of millions to hundreds of millions of dollars. However, there’s no need to compete on the same stage as OpenAI or Google. Training or fine-tuning a small model specialized for your business is now within reach for small businesses.

Let me provide an example from academic research. The model “Puro-2B” reportedly achieved practical performance with training on an NVIDIA RTX 5090 (costing about 300,000 yen) for only $6,900 (about 1 million yen).

1 million yen. While not a small amount for small businesses, it is certainly a different scale from what one might imagine when hearing about “AI model training.”

Moreover, there are paths that do not require full scratch training. If you fine-tune existing open-source models with your company data, you only need to rent cloud GPUs for a few hours. The cost ranges from a few thousand to tens of thousands of yen.

What once seemed to be a world of hundreds of millions of yen is now within reach for 1 million yen. Fine-tuning can be done for tens of thousands of yen.

Without knowing this price disruption, they conclude that “AI is for large companies” and do nothing. Alternatively, they continue using generic external services and lament that they do not fit their business. This is mistake 3.

Adjusting models with unique data—such as your own estimates, past customer interaction logs, and manufacturing process records—can become a weapon that large companies do not possess. You can achieve precision specialized for your business that generic AI cannot provide.

The strength of small businesses lies in having “narrow and deep data.” Large companies create generic models with broad and shallow data. Small businesses create specialized models with narrow and deep data. This asymmetry can lead to a reversal of fortunes.

—

So, What Should We Do?

The common thread among the three mistakes is “not knowing that costs have decreased.”

Because they are unaware, they make decisions based on the market conditions from three years ago. Because they are unaware, they buy at the vendor’s asking price. Because they are unaware, they give up, thinking, “It’s impossible for us.”

The specific actions are simple.

1. Start experiments at minimal cost. A $399 robot, free open-source LLMs, cloud GPUs for a few thousand yen. First, get hands-on experience. If it doesn’t fit, you can stop. The damage from a 60,000 yen experiment versus a 3 million yen implementation is 50 times different when it fails.

2. Take stock of “how much you are paying per month.” Add up the monthly costs of external AI services. How much does it amount to annually? Are there any services that can be switched to local operations? You don’t need to switch everything. Even internalizing just core operations could save 1 to 2 million yen annually.

3. Reassess the value of your own data. Customer interaction data from the past 10 years, trends in estimates, complaint histories—these are data that large companies cannot buy even with money. Fine-tuning a model with this data can yield a dedicated AI for your company. Costs start from tens of thousands of yen.

—

The Era of “Expensive AI” is Over

AI costs continue to decrease even at this very moment.

What seemed “expensive” six months ago may now be half the price. It will likely decrease further in a year. This trend will not stop.

The problem is continuing to make decisions based on outdated perceptions without realizing that costs are decreasing.

A $399 robot. A $500 monthly local LLM. $6,900 model training.

Looking at these numbers, can you think, “Maybe we can do this too”? If you can, then try one thing first.

The way for small businesses to succeed with AI is not to imitate large companies. It is to start “cheaply, small, and with their own data.”

Those options are now finally available at realistic prices. There’s no reason not to use them.