Etched partners with a16z to advance AI inference hardware, raising questions about Nvidia’s GPU dominance

A startup most people have never heard of is quietly assembling the ingredients to challenge Nvidia’s stranglehold on AI compute. Etched, founded in 2022, has partnered with Andreessen Horowitz to build purpose-built inference hardware from the ground up, and a16z partners are now publicly declaring that AI inference is “the most important workload of our era.” Etched has raised approximately $800 million in total funding and locked in over $1 billion in signed customer contracts for its rack-scale inference systems. What Etched is actually building The core product is Sohu, an application-specific integrated circuit (ASIC) designed exclusively to run transformer models. Instead of building a general-purpose chip that can do lots of things pretty well, Etched built a chip that does exactly one thing, run the architecture behind every major large language model, and does it extraordinarily well. The Sohu chip is fabricated on TSMC’s advanced N4P process node. Etched has already completed what’s known as an A0 tape-out, meaning the first version of the physical chip design has been sent to TSMC for manufacturing. Etched says Sohu delivers more than 10x performance gains over leading ...