AI’s Electricity Costs Are Not “Someone Else’s Problem”

Microsoft, Amazon, and Google are estimated to emit a total of 119 million tons of CO₂ by the fiscal year 2026. This is roughly one-third of France’s annual emissions. Even looking at Microsoft alone, recent reports indicate a 25% increase in emissions compared to the previous year. The primary cause? The rapid expansion of AI data centers.

Can we dismiss this as “a problem for big corporations”? We cannot.

As electricity demand skyrockets, wholesale electricity prices will rise. If competition for renewable energy procurement intensifies, those costs will be passed on to electricity bills. In other words, regardless of whether they use AI or not, small businesses will face structural pressure to pay higher electricity costs. The benefits of AI will be enjoyed by Big Tech, while the electricity bills will be distributed across society. It is essential to recognize this asymmetry.

Understanding AI’s “Electricity Gluttony” Through Numbers

Just how much electricity are we talking about?

According to estimates from the IEA (International Energy Agency), global data centers consumed approximately 460 TWh of electricity as of 2022. This could reach as high as 1,050 TWh by 2026. This figure is nearly equivalent to Japan’s total annual electricity generation (around 1,000 TWh). In just four years, consumption could more than double, with most of that increase attributed to AI workloads.

The power required to train a model like GPT-4 is estimated to be in the tens of GWh. Given that the average household consumes about 4,500 kWh annually, one training session could use the equivalent of several thousand households’ yearly electricity consumption. Even during inference (when users actually utilize the service), it is said to consume about ten times the power of a single Google search.

An infrastructure that consumes this much electricity is currently undergoing a construction boom worldwide. In the United States alone, the investment in new data center projects announced in 2024 amounts to several billion dollars. The competition for power has already begun.

The Proposal of “Turning Homes into Data Centers”

In response to this concentration of power issue, an interesting development has emerged.

Sunrun, a U.S. solar panel leasing company, has announced a “distributed AI compute” program. Essentially, it aims to utilize solar panels and batteries installed on residential rooftops by setting up small computing nodes within homes to provide AI computational resources. Participating households will receive compensation.

The concept is logical. Instead of concentrating power in massive data centers, it harnesses the already distributed renewable energy sources in homes. This reduces transmission losses and potentially lowers cooling costs compared to large facilities.

However, at this stage, it remains in the “conceptual phase.” The processing power of computing nodes that can be placed in homes is limited and not suitable for large-scale model training. Realistically, it may only be feasible to distribute some inference tasks. Still, the direction of “democratizing power” is noteworthy. The mere increase in options that do not rely on centralized data centers holds significance for small businesses.

The “Real Risk” for Small Businesses Is Not Electricity Costs

Now, let’s shift our perspective.

When small businesses use AI, the direct impact of electricity costs on their operations is actually not that significant. This is because most small businesses utilize APIs from OpenAI or Google. Very few small businesses set up their own GPU servers for training.

In other words, the AI electricity costs for small businesses come indirectly in the form of “API usage fees.” This is a crucial point.

Currently, the API usage fee for GPT-4 is $2.50 per million input tokens and $10 per million output tokens. Compared to the GPT-3.5 era, the cost per performance has dramatically decreased. Competition is fierce, with Google’s Gemini, Anthropic’s Claude, and open-source Llama models engaging in price competition.

Therefore, in the short term, the risk of API usage fees skyrocketing is low. In fact, they continue to decrease.

The real risk lies elsewhere: service limitations due to power shortages and dependency on specific providers. If OpenAI were to announce, “Due to rising electricity costs, we are tripling API fees,” would you have alternatives? How dependent is your business on that API? Many small businesses lack awareness of this.

Three Things Small Businesses Should Do Now

So, what should be done?

1. Experiment with Local Small Models

Open-source small models like Llama 3, Phi-3, and Gemma 2 can run on laptops. Not all tasks require a GPT-4 class model. For internal FAQ responses, meeting minutes summarization, or drafting standard emails, an 8B parameter model is sufficiently practical.

Create at least one task that can be completed entirely on your own PC without paying a cent to cloud APIs. This will serve as “insurance against power risks.” Electricity costs? A laptop consumes about 50W. That amounts to a few hundred yen a month.

2. Visualize API Dependency

List which processes in your business workflow depend on which APIs. How much do you pay monthly, and how many hours would operations halt if that API went down? Simply understanding this will enhance your decision-making speed in critical situations.

3. Be Aware of “Edge AI” Options

Edge AI refers to technology that completes AI processing on devices without sending data to the cloud. Facial recognition on smartphones and object detection in cameras already operate at the edge. In industrial applications, areas like factory appearance inspections and inventory counting are also well-suited for edge processing.

By processing at the edge, you incur zero communication costs and cloud usage fees. Since data does not leave the device, security is also enhanced. In some cases, power consumption can be a fraction of that required for cloud processing. Questioning the assumption of “doing everything in the cloud” can change the cost structure.

Understand the Structure and Avoid Being Tossed Around

In summary, here’s what we have:

• Big Tech’s AI data centers are consuming vast amounts of electricity. By 2026, these three companies alone will emit CO₂ equivalent to one country.
• This electricity demand creates structural pressure that will push up electricity prices for society as a whole.
• However, the direct impact on small businesses’ AI usage costs comes not from “electricity bills” but from “API usage fees” and “provider dependency risks.”
• Home data centers and distributed computing are interesting directions, but they remain in the conceptual phase.
• What can be done right now includes experimenting with local models, visualizing API dependency, and considering edge AI.

The electricity issue surrounding AI is not a theme that small businesses can directly solve. However, understanding its structure can determine whether they are swept away by cost waves or can ride those waves to their advantage.

While large corporations invest trillions in infrastructure, small businesses can advance their AI utilization in a “small, light, and independent” manner. This is not a weakness but a structural strength. A time is coming when not having massive data centers will actually provide agility.