What Changes When AI’s Electricity Costs Become “Almost Zero”

The former AI chief of Databricks stated that “we can reduce the power costs of AI to one thousandth.”

This is not just a technical discussion. Changing the cost structure means that the players who can win will change.

Currently, when trying to fully utilize AI, the burden of electricity costs weighs heavily. The power consumption for a single inference of large language models is about ten times that of a Google search. Data centers account for 2-3% of global electricity consumption, and this is further inflated by the growing demand for AI.

This “electricity cost” has become an invisible barrier for small and medium enterprises (SMEs) looking to fully leverage AI. If we break down the API usage fees, about 30-40% of that is attributed to electricity costs. In other words, if electricity costs drop to one thousandth, the cost of using AI itself will dramatically decrease.

So, what happens to SMEs if this becomes a reality? This is where the main topic begins.

How the Cost of “Using AI” Will Change

Let’s consider some specific numbers.

Currently, integrating an API like OpenAI’s GPT-4 class into business operations and calling it several thousand times a month costs about 50,000 to 150,000 yen monthly, or 600,000 to 1,800,000 yen annually. For local SMEs, this is the borderline of what can be considered “let’s give it a try,” or it may already exceed that amount.

If 30-40% of this cost comes from electricity, then with electricity costs reduced to one thousandth, API usage fees could theoretically drop by 30-40%. A monthly fee of 150,000 yen could become 90,000 to 100,000 yen. This alone is significant.

However, the real impact lies beyond that.

With lower electricity costs, AI models can be run more casually and in larger volumes. Instead of thinking, “Let’s limit the number of calls because of cost concerns,” it can shift to “Let’s check everything” or “Let’s run it continuously in real-time.”

For example, in manufacturing inspections. Currently, if one tries to conduct a full inspection using image recognition AI, the inference costs pile up, leading to the conclusion that “maybe a sample inspection is sufficient.” If electricity costs drop to one thousandth, running full inspections 24/7 would incur almost negligible electricity costs. What was previously “not feasible due to cost” transforms into “there’s no reason not to do it.”

Three Structural Changes for SMEs

1. “Owning AI In-House” Becomes Realistic

Currently, for SMEs to use AI, tapping into cloud APIs is the practical choice. Owning GPU servers in-house can easily cost tens of thousands of yen just in electricity.

But what if electricity costs drop to one thousandth? A small AI server could be placed in-house, running solely on the company’s data for an on-premises AI that operates with an electricity cost of just a few thousand yen a month.

What does this mean? There’s no need to send data outside. Customer information, partner information, know-how. The anxiety that local SMEs have about “Is it safe to send our data outside?” would disappear. Security concerns would be resolved alongside cost issues.

2. The Breakeven Point Between “Hiring More People or Relying on AI” Changes Dramatically

The biggest struggle for local SMEs is labor shortages. Even when they post job openings, no one applies. Even if they do, they often don’t stay. The hiring cost per person is around 500,000 to 1,000,000 yen annually, and the total labor costs, including salaries and social insurance, range from 4,000,000 to 5,000,000 yen.

On the other hand, what happens if the cost of relying on AI for operations drops to below 100,000 yen annually?

Tasks such as responding to inquiries, creating estimates, summarizing daily reports, and making inventory ordering decisions—these “tasks currently done by people but can be patterned” could be handled by AI for just a few thousand yen a month, 24/7.

“Not enough people” would change to “it runs even without people.” This impact is greater for SMEs than for large corporations. Large companies already have dedicated IT departments and advanced division of labor. SMEs often have one person wearing multiple hats. If AI takes over just 2-3 of those roles, the burden on the ground can change dramatically.

3. Access to the “Same Weapons as Large Corporations”

Currently, the degree of AI utilization is almost proportional to capital strength. Large corporations can invest tens of millions to hundreds of millions of yen annually in AI. SMEs typically manage only several hundred thousand to a few million yen. This 10 to 100 times difference directly translates into a disparity in competitiveness.

If electricity costs drop to one thousandth, this gap could close rapidly. It may become possible for SMEs to achieve the same level of AI analytical infrastructure that large corporations built for 100 million yen, for just several hundred thousand yen annually.

When the price difference of weapons disappears, what will determine the outcome is “how to use those weapons.” In other words, the person who knows the challenges on the ground best will be able to use AI most effectively. The president or team leader of an SME, who stands on the front lines every day, can more accurately determine what AI should be tasked with than a consultant in a large corporation’s headquarters.

This is the structure of reversal that allows SMEs to win.

So, What Should We Do Now?

“Let’s wait until it’s one thousandth” is the worst decision.

The reason is simple. While AI costs will decrease if we wait, the skills to effectively utilize AI will not improve just by waiting.

When the era arrives where anyone can use AI cheaply due to lower electricity costs, the difference will be between “companies that know how to use AI” and “companies that don’t know what to do with AI even though it’s cheap.”

There are three things that should be done now:

1. Start Using AI on a Small Scale
There are already plenty of AI tools that can be used for under 10,000 yen a month. Integrate tools like ChatGPT, Claude, and NotebookLM into your operations and identify “tasks that can be delegated to AI.” It doesn’t have to be perfect. Just start experimenting.

2. Inventory “Tasks That Are Person-Dependent”
Know-how that exists only in the minds of veteran employees or tasks that only specific individuals can perform. Create a list of these. When AI costs drop, these should be the first tasks to automate.

3. Start Accumulating Data
Your company’s operational data, interactions with customers, and historical decision-making records. This will serve as fuel for AI utilization. By organizing and accumulating data now, you can leverage it immediately when costs drop.

When Will Electricity Costs Drop to One Thousandth?

To be honest, it is uncertain when “one thousandth” will be realized. The statement from the former Databricks executive is a long-term outlook combining new chip architectures, cooling technologies, and software optimizations, not something that will happen overnight.

However, the direction is certain. NVIDIA has announced that its next-generation GPU (Blackwell) has improved inference power efficiency by up to 25 times compared to the previous generation. Google is also advancing similar efficiency improvements with its in-house developed TPUs. Innovations in semiconductor-level efficiency, software-level model lightweighting, and data center-level cooling and power management are all progressing simultaneously.

Whether “one thousandth” will come in five years or ten years is unknown. However, “one tenth” could feasibly arrive within 3-5 years. Even at one tenth, the landscape of AI utilization for SMEs would change dramatically.

Conclusion: When Costs Disappear, What Remains is the “Ability to Utilize”

A future where AI’s electricity costs drop dramatically is a tailwind for SMEs. The difference in capital strength will no longer directly translate into differences in weaponry.

However, tailwinds blow for everyone. The difference will be whether you had your sails up before the wind started blowing.

Now is the time to engage with AI, inventory your operations, and accumulate data. While it may seem mundane, this is the most reliable preparation for winning in a world where “AI’s electricity costs have disappeared.”

Do not wait for technological evolution. Build a system now that can fully utilize technology when it becomes cheaper.