Gavin Uberti

Challenging the General-Purpose GPU Paradigm

Gavin Uberti and his team dropped out of Harvard with a bold, controversial conviction: general-purpose GPUs are fundamentally inefficient for the future of artificial intelligence. While traditional semiconductor giants build chips designed to run any arbitrary workload, Etched took the ultimate gamble by hardcoding the transformer architecture directly onto microchips.

Their flagship ASIC chip, Sohu, eliminates unnecessary hardware abstraction layers. By optimizing exclusively for transformers—the backbone of modern LLMs, video generators, and multimodal models—Etched unlocks orders-of-magnitude faster throughput and dramatically lower power consumption.

The Silicon Bet on Transformers

Building custom silicon is widely considered one of hardware engineering's highest-risk ventures. However, Sohu's architecture offers a decisive advantage for real-time AI inference, allowing massive models to run at blazing speeds previously thought impossible on standard hardware clusters.

As AI deployments shift heavily from model training to large-scale operational inference, Etched provides giant tech enterprise workloads, frontier labs, and cloud providers with a crucial alternative to compute bottlenecks and sky-high chip costs.

Building the Next Semiconductor Giant

Backed by Sequoia Capital, Primary Venture Partners, Peter Thiel, and leading AI luminaries, Gavin Uberti is leading Etched through a defining moment in the hardware race. By betting that transformers will remain the dominant AI paradigm for years to come, Etched is redefining performance boundaries across the entire global semiconductor ecosystem.