AI’s ‘Location’ Has Changed. So, What Happens Next?

In the past, using AI meant paying monthly API fees of tens of thousands of yen, sending data to the cloud, and waiting for responses. That era is coming to an end.

A 290MB AI model runs solely in the browser. AI is embedded in Casio watches. Google’s Gemma 4 performs offline inference on iPhones.

What these three pieces of news have in common is that “AI has come down from the cloud.” This structural change is particularly significant for small and medium-sized enterprises (SMEs) rather than large corporations.

A 290MB AI Runs in the Browser—What Changes?

The Bonsai model, with 1.7 billion parameters, operates in the browser at just 290MB in size. Using WebGPU, it executes real-time inference solely with local GPU resources.

Let’s break it down with numbers:

• Model Size: 290MB. Equivalent to dozens of smartphone photos. Easily fits on a USB drive.
• Required Infrastructure: Just a Browser. No server needed, no API contracts required.
• Running Costs: Zero. Not a single request sent to the cloud.

Previously, when SMEs tried to incorporate AI into their operations, they faced significant obstacles. Using the ChatGPT API could cost between 30,000 to 100,000 yen per month. Running a custom model in the cloud could cost several hundred thousand yen monthly. Additionally, there was resistance to sending data to the cloud. A single comment from a CEO, “Can we really send our customer data outside?” could halt a project.

The 290MB browser AI fundamentally changes this structure. No data leaves the premises. Monthly costs are zero. It can run on any device that can open a browser.

Of course, the performance of a 1.7B model is far from that of GPT-4o. It is not suitable for complex reasoning or long text generation. However, consider what tasks are typically required of AI in the context of SMEs:

• Classifying and drafting inquiry emails
• Summarizing daily reports and documents
• Generating standard text
• Extracting and organizing simple data

These “80% of tasks” can be adequately handled by a 1.7B class model. Only the remaining 20% requires GPT-4o. In other words, there is now a possibility to automate 80% of daily operations with AI at zero monthly cost.

Casio F91W with Voice AI—A 2,000 Yen Watch Becomes a ‘Device’

The Casio F91W. One of the world’s best-selling watches, available for around 2,000 yen. An engineer has equipped this F91W with voice AI.

By integrating the nRF52840 chip and combining OpenAI’s Whisper (a speech recognition model) with BLE (Bluetooth Low Energy), voice recognition is achieved on the watch.

What’s noteworthy here is not the technical prowess but the fact that “the concept of where AI operates has been shattered.”

Until now, AI devices were assumed to be high-performance smartphones or workstations costing tens of thousands of yen. Now, it runs on a 2,000 yen watch. Just one chip is enough.

What lies ahead? A factory worker could speak to their watch, asking for “today’s production instructions,” and receive a schedule in response. A craftsman on a construction site, with their hands full, could ask their wrist for “the next process” and receive a voice response.

In SMEs, there are often more employees not sitting in front of computers. Factories, warehouses, stores, and construction sites. The ability to “use AI without opening a computer” resonates more with SMEs, which have more fieldwork than large corporations that primarily focus on desk jobs.

Moreover, the device cost is overwhelmingly low. Distributing smartwatches to all employees could cost hundreds of thousands of yen at 30,000 to 50,000 yen per unit, but devices like the F91W are in a different price range.

Offline Inference with Gemma 4—AI Reaches ‘Areas Without Connectivity’

Google’s Gemma 4 allows for offline inference on iPhones. AI can operate entirely within the device without needing a connection to the cloud.

This is particularly effective in areas with unstable communication environments. Factories in rural mountainous areas. Warehouses underground. Vessels at sea. There are many locations across Japan where communication is weak or non-existent.

Cloud AI becomes just an app the moment the connection drops. Offline AI continues to operate regardless of communication conditions.

More importantly, there is the issue of data sovereignty.

What concerns SME owners most is whether “our data will go outside.” Customer lists, client information, and estimated amounts. Business owners who resist sending this data to the cloud are estimated to be around 70-80% based on experience.

With offline inference, data never leaves the device. This “sense of security” becomes a decisive factor for implementation, more so than technical specifications.

The Essence of Structural Change—The ‘Cost Structure of AI’ is Upended

When we line up these three pieces of news, a single structural change becomes apparent.

The row to note is the one for “Scaling.” Cloud AI incurs higher API fees the more it is used. If ten employees use it 100 times a day, the monthly costs skyrocket.

Edge AI is different. Once a model is installed on a device, the cost remains unchanged no matter how many times it is used. It becomes a cost structure based on “unlimited use.”

This is critically important for SMEs. Large corporations can process monthly API costs of 1 million yen as expenses. SMEs cannot do that. However, if running costs are zero, the situation changes.

“So, What Should We Do?”

Here are three actions you can take starting today.

1. First, Try the 290MB Model in Your Browser

Apply a browser-based model like Bonsai to your company’s routine tasks. Email classification, daily report summaries, FAQ responses. You will surely find tasks that can be accomplished without sending data to the cloud.

2. Consider AI Utilization for ‘Employees Not in Front of Computers’

For field workers to use AI, they will need smartphones or wearables. Experiment with a combination of voice input and a small device with one team first. The cost will be just a few tens of thousands of yen.

3. Propose ‘AI That Doesn’t Send Data Outside’ to Business Owners

The biggest barrier to AI implementation is not technology but the anxiety of business owners. “No data will leave the device. Everything will be completed within this device.” This phrase will become the strongest argument for getting approval.

A New Era Where AI’s ‘Location’ Determines the Outcome

In the era of cloud AI, companies with financial strength had the advantage. Those who could afford high API fees could use more AI.

In the era of edge AI, the structure flips. If running costs approach zero, the decisive factor will be “where, who, and what they use AI for.”

Deploy a 290MB model in the browser to automate responses to inquiries on-site. Deliver AI to workers with their hands full using a wristwatch device. Bring offline models to areas without connectivity.

All of this is an area where SMEs can move faster than large corporations. Decision-making is quicker. The distance to the field is closer. They can “try it out with just one person.”

The change in the ‘location of AI’ is not just a technical trend. It signifies that “the rules of the game regarding who benefits from AI are changing.”

Should you continue paying monthly fees to the cloud, or operate at zero cost with devices at hand? It’s best to consider this choice sooner rather than later.