Hello, everyone, and welcome to the NEXT ACADEMY Podcast. Our theme this time is “AI that runs locally and the assumptions behind the division of labor that are beginning to break down.” Starting with Masa’s experience installing an environment for running a local LLM on his own machine, we asked him about the current state of Chinese open models, alternatives to dependence on data centers, and how each person’s way of working may change.
Starting to Run AI Locally
HAL: I hear you recently set up a system for running AI locally in your own environment. What prompted that?
Masa: I installed an environment for running local LLMs, and I’m experimenting with a setup that runs AI entirely on my own device. The background is that AI models from China are very active right now. They can run at considerably lower cost and locally as well. Those two things are beginning to come together, so I wanted to install one and see for myself.
HAL: The runtime environment itself and the model you run on it are separate things. Does that mean a setup that works entirely on your own device is becoming a realistic option?
Masa: Yes. Previously, the assumption was that you had no choice but to use a high-performance cloud service, but that assumption is breaking down. Once your own machine can deliver a practically useful level of performance, the meaning of having alternatives changes.
Where Chinese AI Stands Today
HAL: Let me outline the models coming out of China. Major players include DeepSeek, Alibaba’s Qwen, Zhipu AI’s GLM, and Moonshot AI’s Kimi. As one indicator, OpenRouter, a service that aggregates usage across models, reported that all five of the top endpoints in its monthly ranking as of July 2026 were Chinese. Do you feel they have come that far?
Masa: Even without looking at the figures, I can feel it when I use them. The cost difference is especially large. There are situations where comparing the same use case reveals an order-of-magnitude difference.
HAL: In terms of performance, one view places them around six to nine months behind the top US models. Depending on the application, costs range from a fraction to less than one-tenth. How do you see that gap?
Masa: If you focus only on the very frontier, I think there is a gap. But much of what we do in practical work doesn’t require the cutting edge. Seen that way, an order-of-magnitude difference in cost expands the range of what we can do itself.
The Value of Not Depending on Data Centers
HAL: I understand that cost isn’t the only thing you hope to gain from running models locally.
Masa: That’s right. What I value most is the environmental impact. Instead of building huge data centers somewhere and consuming resources such as water, we move the computing to users’ local devices and complete it there. I think that’s a wonderful direction for the natural world.
HAL: With large-scale generative AI operations, water consumption for cooling is an issue alongside electricity use. Moving computing onto your own device means restructuring that arrangement itself.
Masa: Exactly. Instead of concentrating everything in one place, we distribute it so that everyone handles a little on their own devices. As an idea, it makes sense to me. In fact, there are already efforts to make open models smaller and bring them to a level where they can run on a smartphone alone. If we get that far, it changes the picture.
The Assumption of a Division of Labor Is Breaking Down
HAL: Let’s turn to working styles. You mentioned feeling a fundamental change in how work is organized.
Masa: Until now, people have worked through a division of labor. I feel that its very foundations are breaking down significantly. One person can assign many roles to AI and have them run automatically. That has become something we can actually do.
HAL: So the scope of what one person can handle is expanding. Will that change come equally to everyone?
Masa: Honestly, I think there will be differences. The sooner people recognize it, the more wealth they can receive. Conversely, I expect the damage to those who are slow to adopt it could be immeasurable.
What Will Happen to Small and Medium-Sized Businesses?
HAL: How do you see the impact on companies that are slow to adopt it, particularly small and medium-sized businesses?
Masa: I think the damage to small and medium-sized businesses could be quite substantial. The more an organization is built around the division of labor, the longer it takes to reorganize. In the meantime, individuals and small teams that can move nimbly will pull ahead.
HAL: What about large companies with capital, or companies with strong resources? I’d like to return to that as part of the outlook for the future.
Masa: I hope we can explore that in more depth next time. But one thing I can say is that scale in itself is not the answer.
This time, we began with AI running locally and followed the discussion through cost structures, environmental impact, and the division of labor as an assumption underlying how we work. Choosing to complete things on our own devices is not simply a technical matter; it affects the way we organize our work itself. I think that was our biggest takeaway from this episode.
We look forward to seeing you next time.
