The 78-Application Anomaly: Why the US AI Export Program Is Already Failing

The data suggests a fundamental misalignment between policy intention and market behavior. The US Department of Commerce’s AI export licensing program has received only 78 applications. This is far below expectations. Far below the volume needed to exert meaningful control over the global flow of advanced AI models. Tracing the regulatory anomaly back to the incentive structure of AI firms reveals a deeper truth: compliance is a cost, not a commitment. Against the backdrop of a bull market in AI infrastructure, where every GPU hour is monetizable, the cost of applying for an export license—legal review, technical disclosure, uncertain approval timelines—outweighs the expected benefit for most mid-tier players. The big players? They have in-house legal teams. The small ones? They simply route through Dubai or Singapore. Context: The program, administered by the Bureau of Industry and Security (BIS), targets “advanced AI models” defined by training compute thresholds and model weight sensitivity. It requires licenses for exports to China, Russia, and other restricted destinations. The narrative from Washington is that this protects national security. But the raw number—78—tells a different story: businesses are voting with their inaction. Core analysis: Why 78? Let me break it down from a cost perspective. In 2023, I audited the cross-border API flows of a major AI startup. The due diligence for a single export license application costs roughly $50,000 in legal fees and engineering time. For a company with overseas revenue under $1M, that is a deal-breaker. They either ignore the rule or shift to model distillation—exporting only small, less-capable versions. The 78 applications likely come from a handful of hyperscalers (AWS, Azure, GCP) and frontier labs (OpenAI, Anthropic). That’s it. The rest of the ecosystem is either out of scope or out of compliance. Furthermore, the definition of “export” is ambiguous. Is a model served via API over the public internet an export? The BIS says yes, but enforcement is nearly impossible. Based on my discussions with compliance officers at three AI dev shops in 2024, I learned that most simply do not file. They assume the probability of being caught is low. And they are probably right—unless they are on a government watchlist. This creates a classic lemons market: only the most risk-averse apply, while the aggressive gain market share. Contrarian angle: The conventional take is that low application volume means the policy is weak and irrelevant. I argue the opposite: it signals a coming crackdown. The BIS will interpret this as industry non-cooperation, not policy failure. Expect tougher enforcement, broader definitions, and higher penalties by Q3 2025. We are heading toward a bifurcation where “compliant AI” becomes a premium product—like FedRAMP for the cloud—and the rest operate in a gray zone. That will accelerate the fragmentation of the AI stack: the US retains control over frontier models, but the rest of the world moves to open-weight alternatives from China (DeepSeek, Qwen) and Europe (Mistral). Tracing the geopolitical cost back to the BIS’s data mismatch shows that the agency likely used Chinese AI progress as a baseline, but underestimated how quickly enterprises would adapt. In 2022, when GPU export restrictions hit, Chinese firms pivoted to domestic chips and software workarounds within 12 months. The same will happen with model exports. The 78 applications are not a sign of policy success; they are a leading indicator of market exit. Takeaway: The real vulnerability forecast here is not about AI safety—it is about US economic influence. If the BIS fails to calibrate its licensing model to market realities, the US will not only lose export revenue but also the ability to shape global AI norms. The math doesn’t lie: 78 applications today means 78 lost opportunities for standard-setting tomorrow. The question is whether the regulatory architecture will adapt before the window closes.