Sohu — the transformer-only chip betting it can beat NVIDIA at inference
HQ: Cupertino, CA, United States | Founded: 2022 | Employees: ~100 (2026, est.) | Stage: Late Stage VC ($10B+ valuation, 2026) | Website: https://www.etched.com
Etched was founded in 2022 by Gavin Uberti (CEO) and Chris Zhu (CTO), two Harvard students who dropped out to build AI hardware — following early work on AI accelerators at OctoML (and the 'Prod' project that predated Etched). Their thesis is simple and radical: the transformer architecture — invented in 2017 and now powering virtually every frontier model (GPT, Claude, Llama, Gemini) — is not a passing phase, so why waste transistors on generality? Sohu, Etched's flagship chip, is hardwired for transformers: attention and feed-forward layers are implemented in silicon rather than executed in software, which the company says delivers up to 20x faster inference than an NVIDIA H100 at a fraction of the power.\n\nThe bet requires extraordinary conviction — if the industry shifts to a fundamentally different architecture, a transformer-only chip becomes worthless — but Etched argues the industry's investment in transformers (trillions of dollars of models and infrastructure) makes the architecture effectively permanent. The company emerged from stealth in 2024, secured over $1 billion in signed customer contracts, and shipped five systems by 2026, with production scaling through TSMC and its supply chain.\n\nCapital followed the conviction: roughly $800M to $1B raised, including a $500M round at a $5B valuation (January 2026), a $300M Series C at a $10.3B valuation (July 2026), and reports of a further raise at a ~$20B+ valuation (August 2026), with Sequoia among the backers. With Sohu's performance claims, a book of $1B+ in contracts, and the AI-inference market exploding, Etched is the highest-profile bet that specialized silicon — not GPUs — will win the inference layer.
Etched is the purest 'bet-the-company on one architecture' story in AI hardware: if transformers remain the dominant AI architecture, a chip hardwired for them (Sohu) can crush general-purpose GPUs at inference — 20x claims, $1B+ in booked contracts, and valuations rocketing from $5B (Jan 2026) to $10.3B (July 2026) toward a reported ~$20B+. The founders (two Harvard dropouts who previously built an AI accelerator at OctoML/Prod) have conviction, capital (~$800M-1B), and customer commitments. The bull case is the inevitability of specialized inference silicon; the bear case is architecture risk (if AI shifts beyond pure transformers, the chip is stranded), NVIDIA's ecosystem moat, and the brutal capital and manufacturing reality of competing with the world's most valuable hardware company.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki