Computing power is becoming the new uranium, and the G7 and EU summit will for the first time place "AI sovereignty" on their official agenda this week.

Article author, source: 0x9999in1, ME News



TL;DR

• In June, the U.S. Department of Commerce imposed an overseas ban on Anthropic's flagship models Mythos 5 and Fable 5, prompting Anthropic to immediately take them offline globally, marking Europe's first direct encounter with the "kill switch."
• More than 40 security experts signed a joint statement opposing the ban, arguing that it targets the wrong product—the undisclosed Mythos Preview, not the commercial Fable 5, is the one with autonomous attack chain capabilities.
• Automated vulnerability analysis is no longer a proprietary technology; GPT-5.5, Opus, Sonnet, and Kimi 2.7 can all achieve the same level of performance—blocking it is equivalent to disarming yourself.
• The three-year "self-developed strongest AI golden window" from 2023 to 2026 has closed; nearly no one except Musk has made serious progress, and Europe is suffering from regulatory backlash.
• Mistral has raised €3 billion in emergency funding at a $20 billion valuation, taking on the full vision of European "AI autonomy"—but the window of opportunity is extremely narrow.
• Computing power is becoming the new uranium, and the G7 and EU summit will for the first time place "AI sovereignty" on their official agenda this week.

A letter, an order, a Europe awakened

It's actually quite simple.

The U.S. Department of Commerce ordered Anthropic to cease providing Mythos 5 and Fable 5 to non-U.S. users. Anthropic acted immediately, without delay or dispute, taking the models offline globally. The decisiveness was almost startling.

Immediately following, over 40 cybersecurity experts, led by former Facebook chief security officer Alex Stamos, signed an open letter opposing the ban. Their reasoning was straightforward: you're targeting the wrong issue.

The White House's basis was a vulnerability report submitted by Amazon's security team, which stated that Fable 5 can automatically generate proof-of-concept code for vulnerabilities and, under specific conditions, can evolve into a complete attack chain.

Sounds scary. But Stamos says, don't worry.

The standard Fable 5 cannot scan the entire Linux kernel or precisely locate all security vulnerabilities. The versions truly capable of "autonomous attack conversion" are Mythos 5 and Mythos Preview—these versions have never been publicly released and are only tested internally within targeted organizations.

In other words, the White House used a report on the "laboratory version" to ban the "commercial version."

What is this operation?

More critically, automatically generating vulnerability PoC code is not unique to Anthropic. OpenAI’s GPT-5.5, Anthropic’s own Opus and Sonnet, and even China’s Kimi 2.7 are roughly on par. Stamos wrote plainly in his letter: Chinese open-source models have now caught up closely with Fable 5 in analyzing system vulnerabilities.

Who does this ban protect?

The defender has lost its sharpest shield, while the attacker still holds the knife. This is not national security—this is national self-harm.

The "kill switch" was pressed for the first time.

For Europe, the impact of this matter far exceeds the technical debate itself.

Over the past few years, European policymakers have privately discussed a hypothetical: what if one day we can no longer use American large models? This question has always been treated as "theoretical anxiety," never taken seriously.

Until this week in June.

The responses from political leaders across multiple European countries were highly consistent—this was the first time the "kill switch" had been genuinely activated. Jordan Bardella, leader of France’s National Rally, bluntly stated that countries failing to rapidly develop their own models will increasingly become dependent on external powers. Former French Prime Minister Édouard Philippe went further, proposing the "Buy European Tech Act" to support domestic development through government procurement and incentives for low-carbon electricity.

Lawmakers from Finland, Bulgaria, and Germany have also begun coordinating to propose uniting with Canada, Australia, and Singapore to consolidate computing resources and advance an EU-level AI supercomputing facility.

Sounds lively. But to be honest, it's too late.

How late was it? Three years.

Anthropic included this statement in its 2023 funding documents: "The company that trains the strongest models in 2025 to 2026 will establish an advantage that competitors cannot catch up to."

Back then, it was dismissed as a startup’s empty promise. Looking at it today, the statement feels as cold and final as a verdict.

Three years of gold, now closed.

AI industry observer Andrew Curran wrote a piece this week saying it even more bluntly.

His core argument is that the period from 2023 to 2026 is the only window of opportunity for all parties to develop their own cutting-edge AI. During these three years, apart from Musk dedicating 26 months to relentlessly refining the Grok series and pushing it to the forefront, almost no other country has made a serious attempt.

What about Europe? Trapped in regulatory turmoil, it has missed valuable opportunities.

It’s not a lack of money or people—it’s that AI legislation came first, with layer upon layer of compliance reviews, turning what should have been a window for rapid progress into a hearing room.

China represents another path. Open-source models are advancing rapidly; products like Kimi 2.7 are now capable of matching Fable 5 in vulnerability analysis. This is why Stamos’s public letter repeatedly emphasizes that "adversaries are heavily stockpiling and exploiting vulnerabilities"—not alarmism, but a factual statement.

Curran also believes that traditional benchmarking has long become obsolete.

A model’s strength isn’t measured by its MMLU score or how well it performs on GSM8K. The real barrier lies in whether it can deeply infer user intent, autonomously reason through and iteratively solve complex tasks, and create an interaction that feels truly alive.

More importantly, leading models have begun to self-iterate, directly participating in the training of the next generation of models. What does this mean? Exponential self-acceleration.

The gap between first and second place is no longer a linear catch-up, but rather snowballing wider.

The golden three years are truly over.

Hash power is becoming the new uranium mine.

Curran presents a chilling prediction in the article.

By 2030, the computing power required for training large models will increase a thousandfold.

What does a thousandfold mean? Currently, training a state-of-the-art model requires roughly tens of thousands of H100-class GPUs. A thousandfold increase would mean computing power becomes concentrated at a scale only two countries can afford—the United States and China.

Top computing power will become the new "uranium."

Strategic materials. National security-controlled items. Licensed regulation. Politically sensitive.

Sounds exaggerated? Just look at Anthropic’s response time this time. One order from the Department of Commerce, and it was taken offline globally—with no room for negotiation. That’s what a “strategic asset” should look like.

What about national AI systems that rely on others' underlying technologies?

Curran’s assessment is harsh: it will effectively become a leasing shell for the China-U.S. model.

Healthcare, education, military, and government services—once their core AI capabilities are built on foreign models, the moment supply is cut off is the moment air raid sirens sound.

Not a metaphor, but literal.

What is Mistral betting on with its 300 million euros?

Europe doesn't have no cards.

Mistral is preparing a new funding round of €3 billion, with a valuation already reaching €20 billion. This is Europe’s only standout flagship in AI.

But frankly, this amount is too small.

Referring to the funding rhythms of OpenAI, Anthropic, and xAI over the past two years, single-round financings starting at $10 billion are no longer surprising. Mistral’s €3 billion is less a powerful strike saying "we’re going to win" and more of an entry ticket signaling "we’re still at the table."

But the significance of this matter has never been just about money.

This week’s G7 summit is being held in France, and both the European Parliament’s initiative on technological sovereignty and the EU leaders’ summit in Brussels have placed AI autonomy on their official agendas. This was a topic that, before the Anthropic supply disruption, couldn’t even secure a spot on the waiting list.

Fear drives action.

But there’s one unavoidable question: What does Europe really mean by "autonomy"—building a Mythos-level model from the ground up, or simply wanting a "backup key" to avoid being cut off in a crisis?

These two endeavors differ in scale by an order of magnitude. The former requires hundreds of billions of euros and five to ten years, while the latter merely requires preserving a seed like Mistral, along with sufficient local computing power.

My personal judgment is that Europe only has the possibility of pursuing the latter. The former is too late now, as it hasn’t been pursued over the past three years.

Open-source models, the next target

Curran's article also contained a judgment that I kept pondering.

He said that once open-source large models approach the cutting edge later this year, they are likely to face a coordinated global crackdown and ban by those in power due to their lack of control.

This sounds a bit conspiratorial, but upon closer thought, the logic holds up.

If a model like Mythos 5 can be instantly locked away by a single Commerce Department ban—what happens to open-source models with equal capabilities? No department can order them shut; they spread globally across Hugging Face in just hours.

For regulators, this is a nightmare.

Therefore, what’s most worth watching in the second half of the year isn’t just the release of GPT-6 or Claude 5, but whether international coordinated regulation for open-source large models will emerge. If this happens, the rules of the open-source ecosystem will be completely rewritten.

Will the Llama series continue to be open-sourced? Could the EU require stricter open-source policies for Mistral? Will Chinese open-source models face targeted restrictions in international dissemination?

These questions will be answered within this year.

The anger in the security community hides a larger issue.

Back to Stamos's open letter.

The fact that over forty top security experts signed this statement alone speaks volumes. The security community is notoriously cautious and would not openly oppose White House decisions unless pushed to the limit.

What exactly are they angry about?

On the surface, it’s a technical detail—the White House confused Fable 5 with Mythos Preview. But beneath that lies a deeper issue: policymakers no longer understand technical details.

This is the truly frightening part.

When a country's highest-level technology policy is built on a misunderstanding of the technology itself, it harms not only the banned companies, but the entire defense ecosystem.

Defenders need cutting-edge models more than attackers do. Attackers can wait, use open-source models, or buy resources on the black market. Defenders must act in real time, must comply with regulations, and must be the strongest. Taking away the strongest tools is like disarming the gatekeeper and telling thieves, "Help yourselves."

Stamos's phrase "disabling one's own abilities" in the letter is not figurative; it is a technical reality.

What did this actually change?

I tried to summarize this week in one sentence.

AI is no longer a product; it is geopolitics.

It was once a consumption tool, a productivity software, a valuation story. Starting this week, it is a nuclear weapon, oil, a strategic reserve, a national lifeline.

Every company that once treated "integrating GPT/Claude/Gemini APIs" as a technical decision must now reassess it as a political one. Every government that once viewed "AI sovereignty" as empty rhetoric must now include it in its national security white paper.

Every ordinary user may soon find that a tool they’ve taken for granted suddenly becomes unusable—not due to a crash, but because of a single signature.

Three years ago, we thought AI was the new internet.

Now I can see clearly—it’s a new uranium mine.

The internet is about connection; uranium mines are about control. The internet encourages diffusion; uranium mines guard against leakage. The winners in the internet are those with the most users; the winners in uranium mines are those with the strongest fists.

The game rules have changed. Some people just haven't caught on yet.

Those who woke up to it—such as the European officials who held an emergency meeting in Brussels this week—likely wore the same expression common to all latecomers.

Panicked, resentful, yet powerless.

After all, the door to three years of gold has closed.

When will the next opening be? Nobody knows.

Perhaps it will never open again.

Reference source

• Andrew Curran's analysis article on large model self-iteration and compute centralization, June 2026
• Anthropic Official Statement: Urgent Notice Regarding Global Access Restrictions on Mythos 5 and Fable 5, June 2026
• Full text of the open letter signed by over 40 cybersecurity experts, including Alex Stamos, June 2026
• U.S. Department of Commerce documents regarding export controls on Anthropic's flagship models
• Anthropic's 2023 funding materials section on projections of model advantages for 2025–2026
• Report on Mistral's €3 billion funding round and €20 billion valuation, June 2026
• Public statements by Jordan Bardella, leader of France’s National Rally, and former Prime Minister Édouard Philippe on European digital sovereignty
• Records of discussions on AI sovereignty at the G7 France summit and the EU Brussels summit agenda