Dylan Patel

Founder and Chief Analyst at SemiAnalysis, the leading independent research firm on semiconductor and AI hardware intelligence. Specialises in GPU compute economics, chip architecture, AI training cluster intelligence, and export control policy. Known for primary-source estimates of AI lab GPU counts, training costs, and hardware sourcing — including the widely-cited ~50K total GPU estimate for DeepSeek/High-Flyer.

Background

SemiAnalysis provides detailed technical and business intelligence on the semiconductor and AI hardware supply chain. Dylan Patel’s work sits at the intersection of chip engineering and geopolitics: he tracks which labs have which hardware, what it costs, and how export control policies are shaping AI capability distribution globally.

Known for: below-the-surface GPU cluster intelligence; rigorous cost analysis of AI training and inference; sceptical takes on AI policy that cut against both triumphalism and alarmism.

Appearances in this wiki

Key positions

• DeepSeek’s training efficiency comes from two compounding innovations: MoE (671B total / 37B active) and Multi-head Latent Attention, plus below-CUDA custom implementation
• Export controls aim to limit inference-scale deployment, not prevent training — a meaningful but constrained objective
• Restrictions paradoxically accelerate China’s domestic semiconductor independence; may guarantee long-term Chinese self-sufficiency
• AGI capabilities may already exist but deployment cost ($5–$20/query) is the barrier, not capability
• Physical compute constraints (fabs, manufacturing, cluster construction) limit sudden AI deployment even when algorithmic capabilities exist