Conclusion

Let’s get straight to the point: the break-even point for owning GPUs is reached in just six months.

For small and medium-sized enterprises (SMEs) looking to leverage AI, the biggest bottleneck has been the cost of GPUs. Should they continue paying 100,000 yen per month for cloud APIs, or invest in their own infrastructure? The results of an experiment using two NVIDIA DGX Sparks connected via Tailscale for distributed learning provide a clear answer.

Let’s look at the numbers:

• Cloud API usage: 100,000 yen per month → 1,200,000 yen per year
• Self-built DGX Spark: Initial investment of about 1,000,000 yen (for two units) + annual electricity and maintenance costs of about 200,000 yen → 1,200,000 yen in the first year, and 200,000 yen per year thereafter.

In six months, the cost amounts to 600,000 yen. If the 100,000 yen monthly cloud API fee disappears against the initial investment of 1,000,000 yen, then the investment is recouped in 10 months. From the second year onward, the annual savings widen to 1,000,000 yen. This is the route out of the “GPU scarcity” dilemma.

What Happened: Distributed Learning on Tailscale Worked “Normally”

Let’s break down the noteworthy aspects of this experiment.

The DGX Spark is a desktop AI workstation that NVIDIA launched in 2025. Equipped with Grace Blackwell, each unit has 128GB of unified memory. The reported price starts at around $3,999 (approximately 600,000 yen) per unit.

In this experiment, two DGX Sparks were connected via Tailscale (a VPN mesh network) to perform distributed learning in a remote environment. The results were as follows:

• Processed 653 million tokens with a two-node configuration
• Step time was approximately 69.4 seconds
• Operated over a VPN connection on a standard internet line, not a dedicated data center network.

This is crucial. Traditionally, distributed learning was assumed to require clusters connected via InfiniBand within data centers. However, it operated seamlessly over a regular internet connection using the free VPN tool Tailscale.

What does this mean? It signifies that the conventional wisdom that “distributed learning requires a data center” has been shattered.

What Changes for SMEs?

Let’s translate these experimental results into the context of SMEs.

1. Elimination of Location Constraints

If connected via Tailscale, one unit can be at the headquarters while the other is at the CEO’s home. It can also function with distributed setups between Tokyo and regional offices. There’s no need to prepare a server room. The DGX Spark is desktop-sized and can operate in a corner of an air-conditioned office.

2. Gradual Investment

Start with one unit and add another when necessary. There’s no need to make an immediate investment of several million yen. You can start with just 600,000 yen. For SMEs, this structure of being able to “start small” is critically important.

3. Data Remains In-House

The data handled by regional SMEs often includes customer information, supplier details, and manufacturing know-how—information they do not want to expose. Using a cloud API means that this data passes through external servers. With an in-house environment, data never leaves the company’s machines. This offers value beyond mere cost savings.

Knowledge Distillation Creates a Structure Where “Small GPUs Can Compete”

Another technology that should be considered in conjunction with this discussion is Knowledge Distillation.

Knowledge distillation is a technique that allows a smaller model (student model) to learn the behavior of a larger model (teacher model). In essence, it means “compressing the intelligence of a GPT-4-level model into a smaller model that can run locally.”

Recent reports from Hugging Face indicate that knowledge distillation can be practically executed even on consumer-grade GPUs (like the RTX 4090). With the Grace Blackwell in the DGX Spark, there’s even more headroom.

What this implies is as follows:

• Large Enterprises: Spend hundreds of millions of yen to train massive models on H100 clusters.
• SMEs: Use those massive models as teachers via APIs and run distilled smaller models optimized for their own data on in-house GPUs.

This allows SMEs to utilize the “knowledge highway” created by large enterprises almost for free while having their own optimized small models. By riding on the investments of large companies, SMEs can possess their own dedicated AI. This is a strategy uniquely available to SMEs.

A More Realistic Look at the Break-even Point

In the draft, the initial investment was stated as 600,000 yen, but let’s calculate this more honestly.

For a configuration of two DGX Sparks:

For cloud API usage:

With a monthly API usage of 100,000 yen, the break-even point is reached in 14 months. If the monthly fee is 150,000 yen, it flips in 10 months. From the second year onward, the annual difference continues to exceed 1,000,000 yen.

Calculating over three years:

• Cloud API (100,000 yen per month): 3,600,000 yen
• In-house DGX Sparks (2 units): 1,740,000 yen
• Difference: 1,860,000 yen

With 1,860,000 yen, you could even add another DGX Spark and still have money left over.

“So, what should we do in the end?”

Let’s break it down into stages.

Immediate Actions:

• Understand your company’s AI-related costs (API usage fees, outsourcing costs) on a monthly basis. If you’re paying more than 50,000 yen per month, it’s worth considering an in-house environment.

Within Three Months:

• Introduce one DGX Spark and test inference and fine-tuning with your own data. This is a 600,000 yen investment. Even if it fails, it won’t be a fatal blow.

Decision After Six Months:

• If results are achieved with one unit, add a second unit and set up a distributed configuration with Tailscale. This will complete a full-fledged in-house learning environment.

What Not to Do:

• Aim for a perfect environment from the start. The strength of SMEs lies in their ability to “test small and make quick decisions.” While large companies are stuck in a three-month approval process, technology will have already moved on to the next phase.

This is the Heart of “GPU Democratization”

To be honest, there are limits to what can be achieved with two DGX Sparks. It is impossible to train a GPT-4-level model from scratch. However, fine-tuning and inference of a small model specialized for your company can yield more than sufficient performance.

And for many SMEs, what they need is not a “huge AI that can do anything generically,” but rather “a reasonably intelligent small AI specialized for their business.” Automating the creation of estimates, classifying inquiries, predicting inventory, recognizing parts from drawings—these types of AI that solve “on-site issues” can operate sufficiently with small models.

A 3,000,000 yen cloud contract can be replaced with a 1,200,000 yen in-house environment. Data remains in-house. Know-how accumulates within the company. Decisions that were previously outsourced can now be handled by the company’s AI, effectively “finishing on its own.”

With two DGX Sparks and Tailscale, totaling 1,200,000 yen and a network connection, this is the minimal configuration for “SMEs to possess their own AI” by 2025.

While large enterprises line up thousands of H100s, SMEs can compete with just two small boxes on their desks. This asymmetry represents the true value that technology brings to regional areas.