You can block atoms, but you can't block bits.

In June 2026, the U.S. government banned foreign entities from accessing Anthropic's Fable 5 and Mythos 5 models. Anthropic urgently dispatched executives to Washington for negotiations. This is not the first time the U.S. has used export controls to restrict technological proliferation—nuclear bombs, high-end chips, EDA software have all followed this path. But this time, the target of restrictions is not a uranium ore, not a lithography machine, not a line of EDA code. It is a neural network weight file. One trillion floating-point numbers. Can run on an H100. Can be transferred via a torrent link.

1. The Governance Toolbox of the Atomic Era

Over the past 80 years, humanity has established three mature systems of technological embargoes. Each has been effective. The underlying logic behind each is the same: control physical bottlenecks.

First: Nuclear Non-Proliferation Treaty (NPT, 1968)

Enrichment of uranium requires centrifuges. Centrifuges need special steel, precision bearings, inverters. Each of these has a traceable physical supply chain. Physical bottleneck = controllable.

Second: Semiconductor export controls (Wassenaar Arrangement + BIS Entity List, 1996–present)

High-end chips require EUV lithography machines. Only ASML can produce EUV lithography machines. ASML’s extreme ultraviolet light source comes from Cymer (USA). Single point monopoly = blockable.

Third: EDA software bans (2022–present)

Designing 3nm chips requires Synopsys / Cadence toolchains. These are closed-source software with license servers that can be remotely shut down. License switches = stoppable.

The common premise of these three systems: what you want to control can only be physically produced by a few people.

Now, try applying this logic to AI models.

2. Heavyweight vs. Weightless

The essence of export controls is logistics management—you must physically move the items.

U-235 needs sea freight. Lithography machines require cargo planes. Even high-end GPUs need containers, customs declarations, end-user verification.

But a trained neural network model, like Fable 5, is about 1–2TB. With current fastest internet connections, this size of file can be transferred to any node on Earth in a few hours.

Export controls can block containers. But can they block a seed file?

This is not a technical detail. It’s a paradigm mismatch.

The technologies humans have previously controlled are "heavyweight": their physical form determines that they can be physically blocked. Nuclear weapons, chips, precision machinery—each has an insurmountable physical bottleneck. You don’t need to control everything; you only need to control that bottleneck.

AI models are "weightless." They don’t require uranium ore, lithography machines, or special steel. Their only physical dependency is computational inference power—and inference chips are shifting from "strategic resources" to "consumer electronics."

NVIDIA just announced RTX Spark at Computex Taipei—a consumer-grade chip capable of running a 120B parameter model locally, coming to laptops this fall. When inference power reaches "everyone has one," what exactly are you blocking?

3. Three Reasons Why It Can’t Be Blocked

It’s not that "blockades are ineffective." The fundamental assumptions behind blockades do not hold for AI. Three reasons.

First: Zero transmission cost

The physical cost of nuclear technology proliferation is building a centrifuge plant—billions of dollars, years of construction, clearly visible to commercial satellites. The physical cost of AI model proliferation is just transmitting a file. After Meta’s LLaMA was leaked to 4chan, copies appeared worldwide within 24 hours. It’s not "leaked"—it’s built-in leak channels.

Second: The open-source ecosystem is already ahead

OpenAI and Anthropic’s closed models are just the first tier. China’s Zhipu GLM-5.2 just open-sourced, with around 1 million tokens of context. Mistral, Falcon, DeepSeek, Qwen—frontier open-source models are narrowing the gap with closed-source ones on a monthly basis. Blocking closed models only accelerates the adoption of open alternatives.

Third: AI is not a "thing"—it’s a general capability

Nuclear weapons are a "thing"—having one or not is a qualitative difference, and once you have a bomb, you have all the capability. AI is not. The competition isn’t about "having a model," but about who can embed inference into real workflows faster, cheaper, and more reliably. Export controls can slow down access to a specific version, but they cannot stop others from training their own models in the same direction.

4. Not Just the U.S. Is Blocking, China Is Also Blocking

Most people only see the U.S. side—banning exports of Anthropic’s models, restricting high-end GPUs to China. But the blockade is two-way.

China has its own domestic alternatives accelerating:

• ByteDance is negotiating with TianShu Zhixin (Iluvatar CoreX) to purchase AI inference chips, expecting at least 50k units this year, becoming its third-largest domestic GPU supplier (after Huawei and Cambrian).

• Chinese AI chip manufacturers now hold about 41% of the domestic AI acceleration server market.

• Huang Renxun himself admits: NVIDIA’s market share in China has "effectively dropped to zero."

This situation reveals a deeper fact: export controls are not about "preventing others from obtaining technology," but about accelerating others to build their own tech stacks.

The more you block, the more motivated others are to develop their own ways to achieve the same.

5. The Real Consequence: Not Technology Being Blocked, but the Internet Being Divided

Blocking a model itself may not succeed. But the side effects of blockade are becoming reality.

Export controls force each side to build its own stack: its own chips, frameworks, models, cloud services, and application ecosystems. This is not "Iron Curtain 2.0"—it’s stack bifurcation.

Over the past 30 years, the value of the internet has been built on "a common protocol stack used worldwide." TCP/IP, HTTP, TLS—no matter where you are, the internet is the same. But now, from chips to models to applications, two parallel AI stacks are emerging:

• U.S. stack: NVIDIA GPU → CUDA → PyTorch → Anthropic / OpenAI → AWS / Azure → English-language application ecosystem

• Chinese stack: Huawei Ascend / Cambrian / TianShu Zhixin → MindSpore / PaddlePaddle → Zhipu / DeepSeek / Qwen → Alibaba Cloud / Huawei Cloud → Chinese-language application ecosystem

These stacks are not necessarily fully isolated, but frictions are increasing: interoperability decreases, standards diverge, talent flow is restricted, open-source communities fracture.

The worst-case scenario is not "one side wins." The worst case is both sides win—each building a complete, incompatible AI infrastructure.

When the entire world’s internet is one, AI is a tool for everyone. When it becomes two, AI becomes a weapon for both sides.

6. Final Words

The character "封" in oracle bone script originally meant "hands cultivating soil and planting trees"—not blocking, but establishing boundaries. In bronze inscriptions, it added the "寸" (hand), becoming "using earth to build boundaries." Later, "封锁" evolved into "sealing off."

In the atomic era, "封" was an effective verb. Because atoms are heavy, slow, and have single points.

In the bit era, "封" might just be a gesture. It won’t stop technological diffusion—but it will change the way it spreads: not through cooperation, but through confrontation; not through shared stacks, but through parallel stacks.

The true cost of export controls is not "the other side cannot get the model." The real cost is that the era of the internet as a unified infrastructure is ending, and no one is seriously discussing alternatives.

When bits cannot be blocked, what we need is no longer "how to block," but "how to coexist after blocking becomes impossible."