What the headlines on DeepSeek are missing, according to RAND

The DeepSeek AI company has achieved significant technical progress while operating under U.S. export controls on advanced AI chips to China, demonstrating both efficiency gains and limitations in the current regulatory landscape.

Key developments: DeepSeek has managed to train advanced AI models using Nvidia H800 chips, which were specifically designed to comply with initial U.S. export controls.

• The company trained its V3 model using 2,000 H800 chips, showing impressive efficiency
• DeepSeek previously operated Asia’s first 10,000 Nvidia A100 cluster and reportedly maintains 50,000 “Hopper” chips
• The timing of their R1 model release coincided with President Trump’s inauguration, potentially for strategic messaging

Export control impact: The real effects of U.S. export restrictions on AI chips to China are still emerging, as the most stringent controls only began in October 2023.

• Chinese companies can still utilize pre-restriction data centers with tens of thousands of chips
• Future expansion and upgrading of computing infrastructure will be more challenging for Chinese firms
• Controls particularly affect deployment capabilities, company growth, and synthetic training capabilities

Technical context: DeepSeek’s efficiency improvements reflect broader trends in machine learning development, while also highlighting specific challenges.

• Machine learning algorithms have historically become more efficient over time
• DeepSeek’s founder acknowledges needing 4x more computing power to achieve results comparable to U.S. competitors
• The company’s efficiency gains likely stem from their previous access to substantial computing resources

Competitive landscape: The comparison between U.S. and Chinese AI capabilities remains complex and partially obscured.

• Leading U.S. companies often keep their most advanced capabilities private
• DeepSeek’s more open approach to sharing model weights and methods contributes to their visibility
• The compute gap between U.S. and Chinese companies continues to widen under export controls

Deployment considerations: The impact of export controls extends beyond initial model training to affect broader AI ecosystem development.

• AI companies typically spend 60-80% of their compute resources on deployment
• Restricted compute access increases costs and limits widespread AI service deployment
• Deployment limitations affect synthetic data generation and capability feedback loops

Strategic implications: The effectiveness of export controls must be viewed within a broader context of technological competition and policy objectives.

• While AI capabilities will inevitably diffuse, controls can maintain technological advantages
• Export restrictions need to be paired with policies strengthening societal resilience and defense
• The controls buy valuable time but cannot completely prevent capability proliferation

Future outlook: DeepSeek’s achievements demonstrate both the potential and limitations of AI development under export controls, suggesting a need for comprehensive policy approaches that go beyond simple restrictions while maintaining technological advantages in democratic nations.