Analysis
Etched disclosed on June 30, 2026 that it has crossed $1 billion in booked orders for Sohu, its transformer-only inference chip, and revealed that it quietly closed a $500 million round in December 2025 at a $5 billion post-money valuation. Co-founders Gavin Uberti (CEO) and Robert Wachen, both Harvard dropouts and Thiel Fellows, have been building toward this moment since 2022, betting that a chip designed exclusively to run transformer-architecture models — rather than a general-purpose GPU — could beat Nvidia decisively on cost and speed for inference workloads.
The product packaging matters as much as the chip itself. Etched sells Sohu not as a standalone part but bundled into 'frontier inference clusters' — racks and software co-designed around the chip — which lets the company capture more value per deal and simplifies adoption for AI labs that don't want to build their own inference infrastructure from silicon up.
“It also puts pressure on Groq, which has yet to disclose comparable booked-order figures publicly.”
The competitive landscape for specialized inference silicon has gotten crowded fast: Groq (secondary market value near $6B) and Cerebras (which IPO'd earlier in 2026 raising $5.55B at a ~$38B market cap) both pursue large-die, non-GPU architectures, while Nvidia's own Blackwell and upcoming Rubin generations keep raising the performance bar general-purpose silicon has to clear. Etched's bet is narrower and riskier than either — full specialization to transformers only means it loses relevance instantly if the industry moves to a fundamentally different architecture, but it wins decisively on cost-per-token if transformers remain dominant.
The numbers in context: $1B in bookings against a $5B valuation is a 5x multiple, reasonable for infrastructure hardware compared to the much higher revenue multiples software-layer AI companies are commanding this year. It also puts pressure on Groq, which has yet to disclose comparable booked-order figures publicly.
What to watch: whether Etched converts its December raise and order book into a public listing given Cerebras' successful IPO comp, how quickly Sohu ships into production versus remaining in pilot deployments, and whether Nvidia responds to specialized-inference competition with its own narrower, cheaper inference-only silicon line.
