The AI race just fractured — and the US did it to itself

Act one: the US government switched off its own best weapons

On June 12, 2026, the US government issued an export control order that forced Anthropic to suspend all access to Fable 5 and Mythos 5 for any foreign national, anywhere in the world. To comply, Anthropic had no choice but to disable both models entirely — for everyone. This was the second time in ten days the government had pulled a commercial AI model off the market for the whole planet.

What reportedly triggered it: during a classified red-team exercise, Mythos reportedly breached nearly every NSA system it was pointed at — within a few hours. The capability was real enough that officials treated the model itself as the threat. Anthropic complied with the legal order but pushed back on the reasoning publicly: "We disagree that the finding of a narrow potential jailbreak should be cause for recalling a commercial model deployed to hundreds of millions of people."

The argument over whether the ban was proportionate is real. But the mechanism matters more than the merits: the most capable AI on Earth was revoked with essentially no notice, by a party that was neither the buyer nor the vendor. If you had built on these models — or paid for them — you woke up to a product that no longer existed.

The credibility cost the US hasn't priced in

Council on Foreign Relations analyst Matthew Ferren put it plainly: "the U.S. is losing the AI credibility war to itself." After the ban, Canadian Prime Minister Mark Carney warned G7 leaders directly that "the export ban underscores the dangers of relying on U.S. models."

That sentence should land hard. The US has spent years arguing that democracies should adopt US-aligned AI rather than Chinese alternatives. The case for that rests entirely on reliability. If allies can't trust that access will be stable — if a government letter can switch the lights off overnight, globally, with no transparent process — the persuasion campaign collapses. You don't need China to undermine American AI influence. The policy process does it more efficiently.

Meanwhile, China's leading models are estimated to be three to eight months behind the US frontier. That window is closing regardless of what the US does with export controls. Restrictions that alienate allies while barely slowing adversaries are not a strategy — they're a self-inflicted wound dressed up as one.

Act two: Japan just proved you don't need a frontier model to match one

The same week the US was banning its own AI, a Tokyo-based lab called Sakana AI launched something philosophically different from anything in the current frontier race: Fugu Ultra.

Fugu is not a bigger model. It's a trained orchestration system that dynamically coordinates a pool of specialized AI agents — routing tasks, delegating to the right expert, running checks, and synthesizing results into a single response. Users see one OpenAI-compatible API endpoint. Internally, Fugu is assembling teams.

The benchmark numbers are striking:

• SWE Bench Pro (software engineering): Fugu Ultra scores 73.7, compared to Opus 4.8's 69.2.
• GPQA-D (graduate-level science reasoning): Fugu Ultra reaches 95.5, matching top performers.
• Humanity's Last Exam: 50.0 — competitive with frontier-class models.

Sakana notes pointedly that Fable and Mythos aren't in Fugu's agent pool — they're not publicly accessible. The implication is explicit: "Fugu would likely score even higher" if they were. The system is designed to route around exactly this kind of access problem. When a provider restricts a model, Fugu swaps it out for another.

The marketing copy lands differently now that the ban has happened. Sakana advertises Fugu as delivering frontier performance "without the risk of export controls." That line wasn't theoretical when they wrote it. It's documented recent history.

Fugu's bet is that orchestration beats scale. You don't need to train the world's largest model — you need to train the world's best conductor. Japan may have just proved that thesis correct.

What makes Japan's entry significant beyond the benchmarks is the philosophical shift it represents. Every major AI lab in the US, UK, and China has been racing to build a bigger, denser, more capable single model. Sakana is betting that how you coordinate existing intelligence matters more than how much raw capability you pour into a single system. If Fugu continues to improve as the pool of agents expands — including future open-weight releases from any country — the competitive moat shifts from training compute to orchestration intelligence. That's a race Japan has positioned itself to lead.

Act three: China's open-weight stack is here and it's serious

While Japan is winning on architecture and the US is losing on policy, China has been quietly stacking the leaderboards with open-weight models that organizations can download and self-host.

On BenchLM's Chinese model leaderboard, the current rankings look like this:

• GLM-5.2 (Z.AI) — score 91, open-weight, the top self-hostable model in the category
• Qwen3.7 Max (Alibaba) — score 90, proprietary
• DeepSeek V4 Pro (DeepSeek) — score 88, open-weight

Chinese labs have pulled into a position where their models "now compete directly with GPT and Claude on a growing share of practical benchmarks." More importantly, GLM-5.2 and DeepSeek are open-weight — which means any organization, anywhere in the world, can download, run, and fine-tune them without an API key, without a vendor relationship, and without fear of an export control letter.

Think about what the Fable/Mythos ban looks like from the perspective of a CTO in Germany, Singapore, or Brazil who had built on US models. They just lived through an overnight outage driven by American domestic policy. GLM-5.2 doesn't have that risk — not because China is more trustworthy, but because the model sits on their own servers. The open-weight advantage is suddenly a geopolitical argument.

What this actually means for the race

A year ago, the AI race looked like a two-horse sprint between the US and China, with the US comfortably ahead and the gap looking durable. That picture has fractured.

The US still has the most capable models — Mythos was capable enough to breach NSA systems in hours, which is both impressive and the reason it got banned. But capability is only one dimension of the race. Reliability, accessibility, and perceived independence from American foreign policy are also dimensions — and on all three, the US just took significant self-inflicted damage.

Japan's entry is a signal that the frontier race isn't purely about training compute anymore. A sufficiently clever orchestration layer over mid-tier models can match frontier benchmarks. As more open-weight models release from more countries, Fugu-style systems become more powerful — not less. The moat of "we trained the biggest model" erodes every time a capable open-weight checkpoint ships.

China's open-weight strategy is working precisely because it sidesteps the political risks that have now materialized in practice. You can't export-control a model that's already downloaded.

The most honest read of this week: the AI race is becoming multipolar faster than anyone expected, and US policy choices are accelerating that transition rather than slowing it.

One real bright spot: the datacenters are finally getting cleaner

Amid all of this, there's one piece of news that deserves attention and optimism — with an honest asterisk.

Nvidia announced a new datacenter cooling architecture that replaces traditional air-cooling with a closed-loop liquid system — a water and propylene glycol mixture that circulates without drawing fresh water from the environment. The system operates at up to 45°C (compared to the 30°C industry standard), which reduces the energy overhead of climate control. For a 50-megawatt facility, Nvidia estimates over $4 million in annual savings on cooling-related energy and water costs. The UN had projected that AI-related water consumption could equal the annual needs of 1.3 billion people by 2030 — this design directly attacks that problem.

It's genuinely good engineering. The asterisk is the one Nvidia's own team acknowledges: these systems are expensive upfront, and it's not yet clear whether or at what pace existing datacenters get retrofitted. Whether this actually bends the environmental curve depends on adoption velocity, not just the technology existing. Promising, real, and worth watching — but not a solved problem yet.

The takeaway

I build security and compliance software, so I read this week through that lens. What I see is a risk landscape that just got materially more complex — but also more interesting.

• Vendor concentration risk is no longer theoretical. If your architecture depends on a single frontier model from a single US vendor, last week showed what the tail risk looks like. It isn't abstract anymore.
• The open-weight case got stronger overnight. GLM-5.2 and DeepSeek aren't the best models in the world — but they're on your servers, and no government letter can take them away. For a large class of workloads, that trade-off now looks different than it did two weeks ago.
• Orchestration is the new moat. Fugu's architecture — a trained conductor over a swappable pool of agents — is a bet that the intelligence lives in the coordination, not the weights. If that bet pays off, it changes who wins the next five years of this race.
• The AI race is multipolar. Japan is competing on architecture. China is competing on openness and accessibility. The US is competing on raw capability while undermining its own network effects through policy instability. Watching which bet pays out is going to be one of the defining technology stories of this decade.

— Jasvant Singh Dosanjh. I build local-first, privacy-respecting security & compliance software at Dosanjh Labs. I write about AI, security, and the intersection of technology and governance.