Three Scenarios: How Chip Exports Reshape Economic Competition

Compute will become a decisive resource for cyber operations against highly strategic targets. As less sophisticated actors gain access to advanced cyber tools, vast compute reserves will allow nation-states to outpace many AI-enabled defenders.

Specifically, greater compute enables a rival nation-state to:

• Develop and train “cyber weapons-grade” models

  Purpose-built for offensive and defensive operations. Today’s frontier models with advanced cyber capabilities, such as Mythos or Opus 4.7, are trained as generalists. To gain an edge, nation-states could use compute to train or fine-tune specialized cyber models focused on vulnerability discovery, exploit generation, and attack automation.

• Deploy AI systems against hardened targets

  Cyber task performance scales directly with inference compute (compute used to run the model and generate tokens). UK AI Safety Institute evaluations found that increasing computational resources from 10 million to 100 million tokens improved autonomous attack performance by up to 59%. Many critical capabilities, such as discovering zero-day vulnerabilities, may be limited only by compute budgets.

A nation-state could use its large-scale compute to attack highly strategic and well-defended targets such as military networks, critical infrastructure, and the energy and financial sectors.

Compute enables adversaries to conduct distillation attacks and use that stolen intellectual property to undermine American AI companies’ competitiveness.

Distillation attacks are systematic campaigns to extract capabilities from American AI models. Adversaries can use such attacks to, in essence, steal the advanced capabilities of American models for a fraction of the original development cost.

American frontier AI companies have invested billions of dollars to develop frontier models. With distillation, adversary-linked competitors can release near-frontier models at a significantly lower cost or for free, threatening the commercial viability of American AI. By cutting into American companies’ revenue streams, adversary models could further lower American investment in research and infrastructure.

But distillation is not free: serving, training on, and refining distilled models at scale still requires substantial compute. Additional chip access would allow a rival to:

• Scale, increase, and improve distillation attacks

  Significant compute is required to train adversary models on illicitly acquired model outputs. Greater compute capacity can allow adversaries to run larger, faster, and more sophisticated distillation attacks.

• Refine distilled models to close the remaining gap

  Raw distillation outputs are a starting point, not a finished product. Post-training techniques such as compute-intensive reinforcement learning allow adversaries to further push a model toward a genuine competitor. Distillation attacks effectively subsidize some stages of the training pipeline, freeing up compute that can be redirected toward the other stages it does not substitute well for.

Exporting compute helps competitor countries build the most valuable parts of the AI industry, costing American firms global market share in those industries.

The AI tech stack is composed of many layers. Chips sit near the bottom of the stack and power the layers above them: data centers, AI models, and AI applications. These downstream layers are where much of the economic returns to AI accrue. AI is projected to add trillions to U.S. GDP through productivity gains. Nvidia CEO Jensen Huang himself claims, “This [applications] layer on top, ultimately, is where economic benefit will happen.”

Exporting compute lets strategic competitors:

• Train and serve AI that competes with its American counterparts

  Compute is the leading input to model capability and the binding constraint on Chinese model development. Chinese models only lag U.S. models by seven months. DeepSeek founder Liang Wenfeng acknowledged, “the problem we face has never been money, but the embargo on high-end chips.” Compute also powers the use of those models, meaning more customers reached. Many strong applications require strong models. American chips sold to strategic competitors power their data centers, models, and applications.

• Capture compute capacity that would otherwise serve American firms

  There is currently a shortage of advanced AI chips. U.S. AI companies would like to scale up their compute, as would many startups and universities. But limits in key production inputs have led to a compute shortage. Nvidia claims U.S. companies are losing revenue as a result of compute constraints. Chips exported are chips American firms cannot access.