Etched - Building AI Hardware to Make Inference Faster and Cheaper - [Invest Like the Best, EP.480]

Specialized AI inference hardware offers a path toward superior performance and cost-efficiency compared to general-purpose GPUs. Etched founders Gavin Uberti and Rob Wachen detail their strategy of extreme vertical integration, designing everything from chip architecture to rack systems and power delivery. Current hardware often relies on outdated constraints, necessitating custom solutions that optimize for low-voltage inference and cluster-scale memory. By prioritizing velocity and parallelization over traditional industry heuristics, the team successfully taped out a functional chip on their first attempt. The shift from training-focused infrastructure to inference-centric systems is critical, as the proliferation of AI agents requires massive, concurrent, and low-latency compute power. This approach enables the production of tokens at a scale and efficiency level that will eventually support a global economy driven by artificial intelligence.