The dual-round structure is itself a signal of the leverage leading AI startups now hold: selling a stake at one valuation and returning almost immediately for more capital at a far higher price has become a defining feature of the current investment cycle, according to the Wall Street Journal. Neither transaction had closed as of July 17, and terms remain subject to change, but the reported $20 billion figure would place a four-year-old company among the world's most highly valued private semiconductor startups — before its first chip has been commercially validated at scale.
Etched's technical bet is specific: rather than a general-purpose GPU that handles a vast array of computational tasks, its Sohu chip bakes transformer architecture directly into the silicon, eliminating the overhead that comes with Nvidia's CUDA software stack. The company exited stealth in June with $800 million raised to date and more than $1 billion in signed customer contracts, claiming 20x throughput over Nvidia H100 GPUs on transformer inference workloads using fixed-function attention circuits.
Founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen, Etched joins a crowded field of inference-focused challengers including Cerebras Systems, Groq, SambaNova, and UK-based Fractile — all pursuing the same thesis: as inference workloads scale with adoption, demand is growing for chips optimized for that specific job rather than Nvidia's do-everything approach. The skeptical read, as one venture investor put it, is that “the most expensive thing in the AI hardware market right now is the word 'claims.'”
