The AI Race Isn't Just US vs China

If you only follow the headlines, the story of AI is a boxing match with two fighters. In one corner, the big American labs. In the other, China. Every few months one lands a punch, everyone updates the scoreboard, and the cycle resets.

It’s a clean story. It’s also incomplete.

There’s a third way of building AI that doesn’t get the same airtime, and the clearest example of it is a small lab in Tokyo called Sakana AI. Watching what they’re doing tells you something the two-horse-race framing hides, and it matters more to a small business than another round of “who scored higher on the benchmark this week.”

Who’s actually in the room

Sakana AI was started by people who would know. One of its founders, Llion Jones, was a co-author on the original Transformer paper, the 2017 research that basically every modern AI model is built on. He left Google to build something in Tokyo instead.

The bet they made is contrarian. While the famous labs race to build ever-bigger models that need ever-bigger data centers, Sakana’s whole thesis is that bigger isn’t the only way to get better. Their name is a hint: sakana is Japanese for fish, and a lot of their work borrows ideas from nature, like evolution, adaptation, and swarms, instead of just throwing more computers at the problem.

That’s not marketing fluff. The track record is real and a lot of it is published in peer-reviewed journals:

• They built a technique that breeds new AI models by combining existing open ones, with no expensive training run required, and one of their small models beat much larger ones at Japanese-language tasks.
• They shrank a large model down small enough to run on an iPhone, while keeping it useful.
• And as of this week, they shipped a system called Fugu that takes a genuinely different approach to building a top-tier AI (we’re writing a separate post on exactly how that one works).

The point isn’t that Sakana has “won” anything. They’re small, and a lot of their headline numbers are their own, not yet independently re-tested. The point is that there’s a serious, credible team proving the race has more than two lanes.

The phrase to watch: “sovereign AI”

Here’s the part that actually reaches down to your business.

Sakana talks openly about building sovereign AI: AI that a country (or a company, or a person) actually controls, instead of renting from someone overseas. Japan isn’t doing this as a science project. The government is funding domestic supercomputers and backing homegrown models on purpose, because depending entirely on a handful of foreign companies for something this important is a strategic risk.

A few concrete examples of that push:

• Japan stood up a national supercomputer (called ABCI) packed with thousands of high-end NVIDIA chips, funded through an economic-security program, explicitly so the country has its own AI horsepower.
• One of its big telecom companies, NTT, built an AI model from scratch that’s small enough to run on a single graphics card, so a business can keep it in-house instead of shipping sensitive data to someone else’s servers.

Strip out the geopolitics and you’re left with a simple idea: the people who depend on AI are starting to want to own it, not just rent it. Countries are realizing this at the national level. The exact same logic applies to a business on Main Street.

A nation worrying about depending on foreign AI and a small business worrying about depending on one vendor’s subscription are asking the same question: what happens to me if they change the deal?

Why a small business should care about any of this

You’re not buying a supercomputer. So why does Japan’s AI strategy matter to you?

Because it’s a preview of a choice you’re going to face, if you aren’t already.

Most businesses today get their AI the rented way: you send your text to a giant company’s servers, their model does the work, the answer comes back, and the meter runs every time. It’s convenient, and for plenty of tasks it’s fine. But it comes with the same three catches that make whole countries nervous:

Your data leaves the building. Every customer detail, internal number, and draft strategy you feed it travels to someone else’s computer. For a country that’s a security problem. For you it’s a privacy and trust problem, especially if you handle anything sensitive.

Someone else sets the rules. The price, the features, whether the model you built a workflow on even exists next year, none of that is your call. We’ve written before about why we tell clients to own your AI instead of renting it, and “sovereign AI” is just that idea wearing a flag.

You’re locked in. When AI is a service you rent, switching is painful by design. When it’s something you run, you can swap it out the day something better shows up.

The reason this is suddenly a real choice, and not just a nice idea, is that the ownable options finally got good. As we covered in open-source AI quietly caught up, the models you can download and run yourself have closed most of the gap on everyday business work. Sakana’s whole research program is more proof of the same trend: you no longer need to be one of two superpowers, or rent from them, to get capable AI.

What this actually changes for you

Nothing about this means you should rip out the tools you’re using. The honest, practical read is simpler:

The “US vs China” headline is a story about who builds the single biggest model. Your business doesn’t run on the single biggest model. It runs on reliable, affordable, private automation doing the same hundred tasks every day: answering questions, sorting messages, drafting replies, following a checklist without dropping a step.

For that work, the interesting frontier isn’t “bigger.” It’s “smaller, cheaper, and yours,” exactly the lane Sakana and Japan are betting on. A model small enough to run on hardware you already own. A system you control. AI that doesn’t leave your building and can’t have the rug pulled out from under it.

That’s the version of the AI race worth paying attention to, because it’s the one you can actually win on your own terms.

If you want AI agents built around your tasks and your data, on infrastructure you control instead of rented from whoever’s winning this month’s headline, that’s what we build. Tell us what your business does all day, and we’ll show you how much of it can run on something that’s actually yours.