Could AI Be Turning from a Service Back into a Computer? The Surprisingly Big Shift Signaled by Kimi K3
Something interesting recently happened in the AI industry.
China’s Moonshot AI released Kimi K3 as an open-weight model.
For people who do not follow AI closely, this may sound like just another new model entering the market.
But when I saw the news, I wondered whether AI might be returning to the era of the traditional computer.
That may sound like an exaggeration, but I believe the development could be that significant.
The Age of Renting AI
For the past several years, we have grown accustomed to renting access to AI.
ChatGPT, Claude, and Gemini all work this way.
The AI does not live inside your computer. You are simply accessing a model running in a massive data center operated by a company such as OpenAI or Anthropic.
It is convenient, but the tradeoff is that you cannot examine or modify what is inside. How the model was trained and how it works remain largely hidden in a black box.
The “Brain” Is Now Available
Kimi K3 is different. It is a massive open-weight model with approximately 2.8 trillion parameters.
Of course, those parameters mean little when viewed directly by a person. Even to me, they would look like nothing more than enormous binary files.
The important point, however, is not simply that the contents have been released. It is that organizations can run the model themselves.
Universities, research institutes, and companies can build AI systems for their own use. They can keep confidential data within their own environments and carry out additional training or tuning for their particular needs.
This is more than the release of another model. To me, it represents the return of the idea that AI can be owned.
The Practical Barriers Are Still Enormous
That does not mean you can start running Kimi at home today.
At present, the model requires far more computing power than an individual could realistically provide. We are talking about servers equipped with dozens of GPUs.
Some people on social media are already celebrating the arrival of the local AI era, but for the average household, that era is still a long way off.
My own guess is that it may take about ten years before an AI of this scale can run comfortably on a standard desktop PC.
Of course, that is only a prediction. A major technological breakthrough could bring that future much closer.
Intelligence May Not Be the Only Thing That Matters
Let us look at the issue from a different angle.
Much of the recent discussion about AI has focused on the race for model performance: Which company’s AI is the smartest?
That question certainly matters. But another may become just as important: How do we run these models?
Today’s AI systems consume tremendous amounts of electricity. If doubling performance also means doubling power consumption, that approach will not remain viable indefinitely.
We may be approaching the limits of what improvements based on Moore’s Law alone can accomplish.
What we truly need may be a new kind of computer—one that can run AI far more efficiently while consuming much less power.
Japan May Still Have a Role to Play
People often say that Japan has fallen behind in AI.
It is probably true that Japan trails the United States and China in the development of very large language models. But the story does not end there.
Japan still has many world-class companies and research institutions working in areas such as semiconductor materials, manufacturing equipment, advanced packaging, and energy-efficient technology.
If the AI race shifts from “Who can build the smartest model?” to “Who can run these models most efficiently?”, Japan may still have a meaningful opportunity.
I am not suggesting that this guarantees a dramatic Japanese comeback. The real world is not that simple.
But there is still an arena in which Japan can compete.
AI Is Becoming a Computer Again
In the past, you bought a computer and ran software on it.
Then the cloud era arrived, and AI became a service that we rented.
From now on, open-weight AI models are being released one after another.
Perhaps we are standing at the entrance to a transition—from an era of using AI to an era of owning it.
If that happens, competition will no longer be limited to AI companies. CPUs, GPUs, memory, storage, electricity, and cooling will all become part of the same contest.
We may have started out discussing the latest AI model, only to find ourselves returning to a world that feels like an issue of a personal computer magazine from the 1980s.