rob

Inference will become the largest market in the world within a decade. Whoever produces the most tokens at lowest cost per watt becomes the most valuable company. We are at 1/1000th of global population using paid AI; scaling to billions of concurrent agents requires orders-of-magnitude more infrastructure from wafer to watt.

Etched builds chip, board, cold plate, interconnect, rack, and production line in-house. This eliminates vendor dependencies, enables 24/7 parallel development (40 days from silicon to inference vs 10 months for peers), and ensures producability at gigawatt/month scale. 'Best vendor is no vendor' — vertical integration compounds speed and yield.

Data center power scales non-linearly: 100MW to 1GW is easier than 1GW to 10GW. Etched's low-voltage + cluster-memory approach maximizes tokens/megawatt. Nations will compete on inference energy efficiency; majority of energy may flow to data centers. 'Agents per gigawatt' replaces GDP per capita as productivity metric.

Inference-time compute scaling (e.g., 6-month tasks compressed to days) plus massive agent concurrency (millions per task) demands cluster-scale memory for low latency and high FLOPs for throughput. Etched's architecture targets this: giant distributed brains (MoE models across thousands of chips) serving billions of concurrent agents. 2027 predicted as first year agents exceed humans in knowledge work.

Leading-node capacity is scarce but not zero-sum if designs target different process nodes and memory types. Etched chose 4nm and different HPM than Nvidia's Rubin (3nm) to avoid direct wafer competition. Vertical integration of non-scarce components (boards, cold plates, interconnects) adds net new production capacity beyond foundry limits.