Fable 5's Ban Will Probably Push AI Toward Open Source

Anthropic’s Fable 5 and Mythos 5 have apparently been cut off from foreign users under an instruction from the U.S. government.

The reason: national security.

Once those words appear, they are powerful.

“It is dangerous, so we are restricting it.” “There are national security concerns.” “We cannot disclose the details.”

When that three-piece set appears, everyone outside the room has very little room to argue.

Of course, there may be real risks. A model at the level of Fable 5 could plausibly be misused for cyberattacks or military purposes.

But when I saw this news, I also thought something else:

This will probably strengthen the momentum to make AI open source.

A Tool That Worked Yesterday Can Suddenly Become “Unavailable”

What this incident made visible is that AI is no longer just a convenient tool. It has become a strategic resource.

A model that worked yesterday can become inaccessible today.

It does not matter whether you were paying for it. It does not matter whether you had a contract. It does not matter whether your business workflow depended on it.

If the U.S. government says no, it becomes no.

This is not only an Anthropic problem. The same thing could happen with OpenAI or Google.

In fact, assuming it will never happen would be more unnatural.

People often say AI will become infrastructure like electricity or water. But if electricity or water suddenly said one morning, “From today, we will no longer supply your country,” that is not infrastructure. That is a collar.

So Should AI Become Open Source?

The next argument is obvious:

Then AI should be open source.

I understand the feeling.

Personally, I am quite sympathetic to that direction.

The major AI companies became powerful by absorbing text, code, images, papers, and other material from across the internet. Yet when their own models or outputs become the thing being used, the words suddenly change: intellectual property, national security, terms of service.

When I previously wrote about the DeepSeek distillation controversy, I felt the same asymmetry.

When you are the one absorbing, it is “learning.” When you are the one being absorbed, it is “infringement.”

When you are expanding access, it is “democratization.” When you are restricting access, it is “security.”

Language is convenient. It can change clothes depending on where you stand.

So it is natural that the demand for open-source AI will grow.

AI that cannot be stopped by one company’s convenience. AI that cannot be blocked at a national border. AI that cannot be choked by a terms-of-service change.

It is natural to want something like that.

But AI Does Not Train Itself on Good Intentions

The problem starts here.

Let’s make AI open source.

That is easy to say.

But who will train it?

To build a large model, you need something much less poetic:

• people
• hardware
• money

You need researchers. You need engineers. You need GPUs. You need electricity. You need data centers. You need cooling systems. You need training data. You need evaluation environments. You need safety testing.

And all of it costs money.

The phrase “open source” has a clean and refreshing sound.

But open-sourcing a large AI model is very different from volunteers building a useful GitHub library on weekends.

It is closer to saying, “Let’s build a power plant together.”

This is not a world where a README, some GitHub stars, and an enthusiastic Discord server are enough.

Could There Be a Grassroots Movement Like SETI?

One possibility is a grassroots distributed effort like SETI@home.

SETI@home was a distributed computing project where people around the world donated idle PC time to analyze radio telescope data and search for signs of extraterrestrial intelligence.

People around the world contribute spare GPU or PC resources.

Everyone brings a little compute.

“Your gaming PC becomes part of humanity’s shared AI.”

Written that way, it has a certain romance.

And I do think this direction will emerge. Distributed training, distributed inference, data contribution, evaluation work.

Systems that collect many small contributions will matter for AI too.

But training a large model is not that simple.

Unlike SETI-style workloads, you cannot always distribute small independent tasks and gather the results later.

Training requires communication. It requires synchronization. It requires quality control. It also requires protection against malicious participants.

If we try to build humanity’s shared AI, we first run into the problem that humanity is not very good at sharing.

There is something deeply ironic about that.

Then Maybe a Rich Person Can Fund It

Another possibility is that some billionaire funds it.

An AI version of the Linux Foundation. An AI version of Mozilla. An AI version of Wikipedia.

Or perhaps one of the world’s wealthy people suddenly awakens and says:

“I will build an open foundation model for humanity.”

Beautiful.

But also slightly worrying.

Will that person’s ideology not enter the system? Who chooses the foundation’s board? Which country’s laws govern it? Who decides the limits on dangerous use? How will politically inconvenient topics be handled?

In the end, are we only replacing “corporate AI” with “billionaire AI”?

The manufacturer of the collar changes, but the collar remains.

Then Japan’s Government Should Fund the Training

As a Japanese person, I want to say:

Maybe the Japanese government should seriously fund the training cost of open-source AI.

The government does not need to create its own AI company. If it tries to operate a government-made AI from scratch, it will probably spend most of its energy on meetings, procurement, and reports.

But funding compute resources and training costs to support open-source AI is different.

If AI has become this strategic, there is value in growing an open foundation model that can be used from Japan.

Continuing to depend on foreign models is risky. Access may suddenly be restricted. Prices may spike. Japanese language and Japanese business context may be deprioritized.

If so, national investment in compute resources, data preparation, researcher development, public data usage, and the training process itself makes sense.

In particular, I think the key is funding only the training.

If the government controls the model’s ownership, ideology, and operation, the discussion immediately becomes heavy.

But if public money covers the GPU bill on the condition that the resulting model is released as open source, that is a much cleaner structure.

It is closer to building roads or ports. Not infrastructure for one company, but a foundation that an entire industry can use.

As a public works project for the AI era, GPU bills may have more future value than another concrete building.

Of course, this also has a punchline.

If the Japanese government funds the training, first there will be a committee.

Then there will be an expert panel.

Then there will be something with a name like the Preparatory Working Group for the Promotion Council on the Development of Domestic Generative AI Infrastructure.

And then a report will be published.

As a Word document.

The fact that it is not written vertically in the ancient Japanese style may already count as progress.

Open-Source AI Is Not Idealism. It Is Defense.

Jokes aside, I think open-source AI will become very important.

Within that, I think one particularly promising approach is for the Japanese government to support open source by funding only the training.

The government would not own the model. It would provide the compute resources needed for training as public investment.

The result would be released as open source.

That structure may allow Japan to grow a foundation model it can use domestically while avoiding some of the heaviness of a fully government-led AI project.

This is not just idealism.

It is not only a beautiful story about free technology.

It is a very practical defense measure.

Do not depend on one company. Do not depend on one country. Prepare for sudden shutdowns. Make the technology inspectable. Adapt it to your country’s or company’s context.

Seen that way, open-source AI is less about “let’s all share knowledge together” and more about:

insurance against dying when access is cut off.

Open source in the early internet era had a kind of idealism.

But open source in the AI era is much more grounded.

Not freedom for its own sake, but avoiding dependence so deep that it becomes a choke point.

Not democratization, but preparation for supply interruption.

Not ideology, but business continuity planning.

That is not very romantic.

But reality is probably like that.

Closing

With the Fable 5 incident, AI has started to grow borders.

The intelligence that seemed to live beyond the cloud has suddenly started asking for a passport.

As a result, expectations for open-source AI will grow.

But opening AI requires more than open ideals.

It needs people. It needs hardware. It needs money.

Grassroots efforts will matter. Rich patrons may matter. Government funding may matter too.

In other words, to make AI truly open, someone must open a very closed wallet.

That is the most ironic part.

The future of open-source AI will probably begin like this:

“Let’s build intelligence shared by humanity.”

And the next line will surely be:

“First, let’s decide who pays for the GPUs.”