78 applications. That is the number. The US Commerce Department’s AI export plan—designed to regulate the flow of advanced AI models to adversarial nations—received exactly seventy-eight applications. Far below expectations. Far below the implied threat level.
I do not trust the silence. I audit the code. But here, the code is policy, and the policy is failing.
Let me establish context. In late 2023, the Bureau of Industry and Security announced a framework requiring licenses for the export of "advanced AI models"—covers model weights, training code, even API access to frontier systems. The stated goal: prevent China, Russia, and other rivals from acquiring capabilities that could threaten national security. The unstated goal: maintain US technological hegemony.
The expected volume of applications? Inferred from the administrative machinery—hundreds, perhaps thousands. The actual volume? 78. A number so low it screams one thing: the regulated parties do not believe the rules apply to them, or they have found alternative paths.
Based on my audit experience in 2017—when I spent three months manually verifying CryptoKitties smart contracts to find an integer overflow that could have collapsed the breeding logic—I learned that low participation in a security mechanism is not a sign of safety. It is a sign of avoidance. The same principle applies here.
Core insight: The low application count is not a bureaucratic hiccup—it is a market signal that decentralized, permissionless AI will fill the vacuum.
The logic is straightforward. US AI companies—OpenAI, Anthropic, Google—derive significant revenue from global customers. The license requirement introduces friction, delay, and uncertainty. For a startup, the cost of compliance outweighs the marginal gain from a few overseas clients. So they do not apply. They either ignore the rule (assuming low enforcement) or they restructure their offerings: release open-weight models, deploy via foreign subsidiaries, or partner with non-US cloud providers.
And here is where blockchain enters. The crypto ethos has always been about permissionless innovation. The same impulse that drove DeFi in 2020—against centralized intermediaries—now drives open-source AI. Projects like Bittensor, Render Network, and Akash Network are building decentralized compute and model marketplaces. They do not ask for permission. They do not file 78 applications. They distribute weight files via IPFS and let anyone run inference on a global GPU mesh.
This is not speculation. In 2020, during DeFi Summer, I built a Python framework to model oracle manipulation risks in Compound. The market ignored my warnings until the wETH oracle glitch—then they remembered. Now, the US government is building an oracle of its own: a central registry of who can export AI. But oracles lie. Data doesn't.
Truth is an oracle, not a price feed. The truth is that the US export plan will fail to contain AI diffusion, because the technology itself is inherently decentralized. Model weights are just numbers. They can be hashed, split, embedded in NFTs, or transmitted over Zero-Knowledge proofs. The cat is already out of the bag—or rather, the AI is already on the chain.
The contrarian angle: some will argue that 78 applications is a victory—it means most companies are complying by not exporting. I call that naivety. The absence of applications indicates either that the definition of "advanced AI model" is so narrow that it excludes most current products, or that companies are deliberately avoiding the system. If the first, the regulation is toothless. If the second, it creates a black market of AI models flowing through encrypted channels—worse than no regulation at all.
Fragility hides in the single point of failure. The US export regime is a single point of failure: a centralized gate that can be bypassed by any actor with a cryptographic key and a foreign server.
Let me bring this home with a personal example. In 2021, I founded a community around on-chain provenance for NFTs. I argued that the value lay not in the image but in the immutable history of creation. Today, the same principle applies to AI models. A model's provenance—its training data, weights, and architecture—can be anchored on a blockchain. Once on-chain, it cannot be subject to export control. The only way to stop it is to shut down the internet.
Proof precedes value; provenance is the only art. The US Commerce Department is trying to control the art, but the provenance is already distributed.
What does this mean for builders? First, double down on decentralized AI infrastructure. The bear market has weeded out the hype; now, survival depends on resilience. If your AI startup relies on centralized API access that can be cut off by a single government decision, you are structurally fragile. Second, monitor the regulatory response. The low application count will likely prompt a tightening—expanded definitions, stricter penalties. But every tightening creates an incentive for new evasion methods. It is a cat-and-mouse game that the government cannot win.
We do not buy pixels, we buy history. And the history of AI regulation will show that the 78 applications were the moment the US conceded that it cannot centrally control the most decentralized technology since the internet itself.
Code is law, but audits are conscience. I audit the silence. The silence from the Commerce Department speaks louder than any press release. They expected a flood of compliance; they got a trickle. The void will be filled by permissionless systems.
Takeaway: The future of AI will not be determined by export licenses. It will be determined by which models run on open, unstoppable networks. Build accordingly.