Presentation: From Fab To Token - The State Of The Market

Details

This 51-minute QCon AI presentation from Jordan Nanos, a Member of Technical Staff at SemiAnalysis, offers a hardware-first lens on AI system design. SemiAnalysis is a semiconductor and AI research firm with 280,000 newsletter subscribers and 85+ staff, known for detailed financial modeling of chip supply chains and GPU performance across cloud providers. The talk centers on identifying current choke points in AI infrastructure: where semiconductor constraints, networking bottlenecks, and data center expansion limits create hard ceilings on AI scaling. Nanos draws on SemiAnalysis''s ClusterMAX project, which benchmarks GPU performance across cloud providers and tests new hardware. The core argument is that software architects who understand their hardware constraints make systematically better decisions — particularly around inference cost optimization, context window tradeoffs, and model serving architecture. For practitioners, the value lies in the framing around tokenomics: the cost-per-token chain from chip fabrication through inference, and how supply chain events upstream translate into pricing changes downstream. Full access to the supporting data and financial models requires a SemiAnalysis subscription.