The Shocking Cost of AI Inference Reduced to One-Tenth: Simultaneous Waves of GPU Reduction, API Halving, and Power Cost Halving

The cost of AI inference, which used to be 500,000 yen per month, is now down to 50,000 yen.

This is not a distant future scenario. Right now, three technological trends are moving simultaneously: an 80% reduction in GPU usage by an Indian startup, a halving of OpenAI’s API prices, and a 50% cut in data center power costs due to solid-state transformers. When these three factors converge, what will happen?

The common belief that “AI is for large corporations” is quietly coming to an end.

—

80% Reduction in GPU Usage — Simplismart Breaks the “Computational Resource Barrier”

The Indian startup “Simplismart” has developed technology that reduces the GPU usage required for AI inference by up to 80%.

What does this mean? For example, if a business was using 10 GPUs, it can now get by with just 2. If the monthly cloud GPU usage fee was 1 million yen, it would now be 200,000 yen.

The company’s method combines model quantization, optimization of batch processing, and redesigning the inference pipeline, dramatically lowering GPU load without significantly compromising model accuracy. In short, it’s technology that allows “the same work to be done with fewer computational resources.”

What’s noteworthy is the impressive growth figures. Simplismart’s annual recurring revenue (ARR) has grown 14 times in just six months, with more than half of its sales coming from the U.S. market. An Indian startup is cutting AI infrastructure costs for American companies. This is not a case of “cheap but poor quality”; it demonstrates a technological advantage.

At this point, we should consider the fundamental implications of “reducing GPU usage.”

Until now, the bottleneck for AI inference has been “the ability to procure GPUs.” How many NVIDIA H100s can you secure? How much can you afford to pay for cloud GPUs monthly? Financial strength has directly determined whether one can utilize AI.

Simplismart’s technology breaks this structure. If fewer GPUs are needed, the gap between large corporations and small to medium-sized enterprises shrinks. This can be described as the democratization of computational resources.

—

OpenAI GPT-6.1 Terra — The Fact That API Prices Have Halved

Another significant development is OpenAI’s announcement of GPT-6.1 Terra, which has reduced API prices to $1 per million tokens.

In traditional models, the cost was about $5 per million tokens. Now it’s $1. That’s one-fifth the price.

Let’s translate this figure into practical terms. For instance, if AI is used to automatically read and classify invoices or contracts, and if 1,000 items are processed daily with an average of 2,000 tokens consumed per item, that amounts to 2 million tokens per day, or about 60 million tokens per month.

Traditional pricing: 60 million tokens × $5/1 million = $300 per month (about 45,000 yen)
GPT-6.1 Terra: 60 million tokens × $1/1 million = $60 per month (about 9,000 yen)

What was 45,000 yen per month is now 9,000 yen. That’s an annual reduction of about 430,000 yen. This is just for one business. If AI is integrated into multiple operations, the savings will increase even further.

Moreover, it’s important to note that it’s not just about being “cheaper.” When prices drop, usage patterns change.

Operations that were previously deemed “not cost-effective”—such as internal inquiry responses, automatic summarization of daily reports, and sentiment analysis of customer emails—can now be enhanced with AI in areas that were previously considered “nice to have but not worth the investment.”

As costs decrease, applications expand. When applications expand, the benefits of AI implementation ripple throughout the organization. This is the essence of the change.

—

Solid-State Transformers — Data Center Electricity Costs Halved

The third wave is somewhat different: the practical application of solid-state transformers.

In traditional data centers, voltage conversion is done using iron-core transformers. They are heavy, large, and have significant energy losses. Solid-state transformers use advanced power semiconductors (SiC and GaN) to dramatically improve conversion efficiency.

Specifically, their size is one-tenth of traditional transformers, and energy losses are halved. This could potentially reduce the overall power costs of data centers by up to 50%.

Electricity accounts for 30-40% of the operating costs of data centers. If power costs are halved, this results in a 15-20% reduction in overall operating costs. This reduction will be reflected in the fees for cloud services. In other words, the costs for companies utilizing AI will decrease even further.

Although this technology is still in the early stages of large-scale implementation, the direction is clear: a structural change is occurring where “the cost of the infrastructure that powers AI is decreasing.”

—

What Happens When These Three Waves Converge

Let’s summarize:

• Efficiency of Inference (Simplismart): 80% reduction in GPU usage → inference costs reduced to one-fifth
• API Price Disruption (OpenAI GPT-6.1 Terra): token price reduced to one-fifth
• Power Cost Reduction in Infrastructure (Solid-State Transformers): 50% reduction in data center power costs

These three trends are occurring simultaneously at different layers: inference engines, model providers, and physical infrastructure. All costs are decreasing at the same time.

Let’s think of it in multiplicative terms. Inference costs are reduced to one-fifth, API costs are one-fifth, and infrastructure costs are halved. In simple calculations, the total cost of utilizing AI could potentially be “less than one-tenth” of what it was 2-3 years ago.

What used to cost 3 million yen could now be 300,000 yen. What was 300,000 yen could now be 30,000 yen.

Such a scale of cost change is not just about being “cheaper.” The very criteria for deciding “whether to use it or not” will change.

—

What This Means for Local SMEs

Now we get to the main point.

Until now, the adoption of AI was something that “only companies with surplus could do.” The initial costs were high, specialized personnel were needed, and the monthly fees for cloud GPUs were burdensome. For local SMEs, AI was a presence that was “interesting but out of reach.”

However, when inference costs drop to one-tenth, this premise crumbles.

For example, consider a manufacturing company with 30 employees. They wanted to implement AI image recognition in their inspection process but had given up due to a monthly GPU server fee of 200,000 yen. What if that fee dropped to 20,000 yen? That’s less than the cost of one part-time employee. There would be no reason not to proceed.

For instance, consider a professional services office with 10 employees. They wanted to delegate contract review to AI but had put it on hold due to an anticipated API cost of 50,000 yen per month. What if that cost dropped to 5,000 yen? That’s cheaper than the lease for a photocopier.

As costs decrease, the debate over “whether to implement it or not” disappears. The discussion shifts to “where to start using it.”

And here lies the unique strength of SMEs. Large corporations may take six months for approvals and validations. If a small business owner decides to “go for it,” they can start as early as next week. The speed of decision-making becomes a competitive advantage.

In an era when AI costs were high, the “financial strength” of large corporations was the winning strategy. In an era when AI costs have dramatically decreased, the “speed” of small businesses becomes the winning strategy.

—

So, What Should We Do?

There are three actions to take:

1. First, decide on one task to automate with AI.

You don’t need to think about full-scale implementation. Choose one task that takes “a certain number of hours each month” or “would fall apart if someone left.” It could be invoice processing, inquiry responses, meeting minutes creation, or inspections. Anything will do.

2. Test on a small scale and measure the costs.

Hit the API and measure how much it actually costs. You will likely realize it’s “cheaper than expected.” With GPT-6.1 Terra, you can handle a considerable amount of processing for less than 10,000 yen per month. Once you have actual measurements, it will serve as persuasive evidence within your organization.

3. Prioritize tasks that are currently dependent on specific individuals.

The greatest value of AI is breaking down dependence on individuals. If you can systematize tasks that “only that person can do” with AI, the operation can continue even if that person leaves. For local companies facing hiring difficulties, this is a survival strategy in itself.

—

Conclusion

AI inference costs are set to decrease to one-tenth. This is not just a technological news story; it represents a structural change in the business environment.

The GPU reductions by Simplismart, the API price disruption by OpenAI, and the infrastructure cost reductions from solid-state transformers are three waves occurring simultaneously.

To ride this wave, significant investment is not required. What’s needed is simply the decision to “try one thing first.”

With an API cost of 5,000 yen per month, you might be able to automate a task that has become dependent on specific individuals. For 20,000 yen per month, you might eliminate human errors from your inspection process.

The day when the notion that “AI is for large corporations” comes to an end is already upon us.