A single piece of news lit up both the AI and chip worlds yesterday: Silicon Valley startup Etched has closed $800 million in cumulative funding, pushed its valuation to $5 billion, and landed a $1 billion chip order.
The company has been quiet. Founded in 2022 by three Harvard dropouts, its investor list is anything but: Nobel laureate Geoffrey Hinton, Stanford vision pioneer Fei-Fei Li, OpenAI co-founder Andrej Karpathy, venture godfather Peter Thiel, quant giant Jane Street and a TSMC-linked fund all came in.
What they are betting on sounds extreme: a chip that runs only the Transformer architecture.
Sohu, Etched’s first chip, uses TSMC’s 4-nanometre process and has already taped out. The company claims a server of eight Sohu units can replace 160 Nvidia H100s on Llama 70B inference.
The bet against general-purpose
Nvidia’s GPU is, at its core, a general-purpose chip. It runs Transformers, CNNs, RNNs, even Bitcoin mining. That “does everything” design is the root of Nvidia’s success.
But generality wastes efficiency. Running a Transformer, huge chunks of the GPU’s transistors do other things. It is like driving an all-terrain vehicle to deliver takeout: it gets there, but the fuel cost is absurd.
Etched’s move, in one line: etch the Transformer architecture into the silicon. Sohu is an ASIC. Unlike a GPU, its circuits are fixed, built to do one thing: run Transformer forward inference.
That extreme trade-off buys extreme efficiency. Etched claims Sohu delivers 20 times the throughput of an H100 on Llama 70B, and 140 times the performance per dollar.
Why backers dare to believe
Chips are one of the few industries where “PPT fundraising” can yield no product for a decade. That Etched got Hinton, Karpathy and Li to put money in comes down to team.
All three founders are Thiel Fellowship recipients. CEO Gavin Uberti traded at Jane Street and knows low-latency compute. CTO Chris Zhu interned on Google’s TPU team and saw a dedicated chip go from design to production. The company has over 400 engineers, mostly from Nvidia, Google’s TPU team, Broadcom and TSMC.
Nvidia’s moat was never just chip performance, but the CUDA ecosystem. Etched’s clever move: it does not ask developers to rewrite code. Sohu runs Transformer models directly; a PyTorch-trained model runs on Sohu with no changes. That is rare in ASIC history.
The $1 billion pre-order is the market’s verdict. Signing that scale before volume production means a customer almost certainly ran real workloads on early prototypes, and they were good enough.
Editor’s note: This is an adapted translation of the original LeiPhone report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.leiphone.com/latest/index/id/4753.