Etched closed a $300 million Series C at a $10.3 billion valuation, led by Sequoia with Andreessen Horowitz, Jane Street, and SK hynix joining, per Seeking Alpha. Its Sohu chip is an ASIC that hardcodes the transformer architecture directly into silicon, which the company says delivers roughly a 20x inference speedup over Nvidia's H100. Etched says it holds over $1 billion in signed customer contracts and begins rack shipments this summer, per Yahoo Finance. Total funding is now about $800 million.
The read is the bet the valuation encodes. Etched has no product at revenue scale yet, so $10.3 billion prices a wager that a chip which can only run transformers beats a chip that can run anything, at least for inference. That is a genuine architectural argument. If the transformer is the durable shape of AI models, hardcoding it removes the flexibility overhead a general-purpose GPU carries, and the efficiency gain follows. The $1 billion in signed contracts is what separates this from a pure thesis round, because customers committed before a single rack shipped.
The risk on the cap table is that same lock-in cutting the other way. A transformer-only chip is a bet the transformer stays dominant. If the frontier moves to a different architecture, a general GPU adapts and Sohu does not. SK hynix in the round is the tell that matters most. The memory maker that supplies the high-bandwidth memory for these systems is hedging into the challenger, not only the incumbent, which is a key supplier voting that inference will fragment away from Nvidia.
Bottom Line
A $10.3 billion mark prices a bet that transformers are permanent enough to hardcode. The $1 billion in contracts says customers will try it. Watch whether Sohu's real-world inference economics hold once racks ship this summer.