Etched
Transformer-only AI chips and rack-scale inference clusters built to outrun GPUs on LLM serving.
About
Etched is an AI hardware company that took the radical bet of hardcoding the transformer computation graph directly into silicon. Its Sohu chip is a transformer-specific ASIC fabricated on TSMC's 4nm process with 144GB of HBM3 per chip, and the company claims an 8-chip server can serve over 500,000 tokens per second on Llama 70B, roughly the throughput of 160 NVIDIA H100s. Those numbers are the company's own and have not been independently verified. Etched now sells frontier inference clusters: co-designed chips, racks, and software aimed at labs serving models at scale.
The company, founded in 2022 by Harvard dropouts Gavin Uberti, Robert Wachen, and Chris Zhu, has raised around $1.9 billion at a $21 billion valuation (Series D, August 2026), led by Jane Street, which is also its first customer, with Sequoia, a16z, and Thiel among backers. Its own site says it has booked $1 billion in contract orders, shipped working A0 silicon from TSMC N4P, and is validating its first rack-scale product with customers ahead of shipments planned for summer 2026. The honest caveat: a transformer-only chip cannot run mixture-of-experts, multimodal, image, or training workloads, so it is a narrow but deep bet on transformer inference staying dominant.
What it can do
- Sohu transformer-only ASIC on TSMC 4nm with 144GB HBM3 per chip
- Low Voltage Inference (LVI): math blocks at under half typical voltage for FLOPs density
- Cluster Scale Memory (CSM): low-latency shared memory pool across the scale-up domain
- Frontier inference clusters: co-designed chips, racks, and software
- Open-source software stack pitched as compatible with existing model code
- Vendor-claimed 500k+ tokens/sec on Llama 70B from an 8-chip server
Setup
- Etched sells complete inference clusters, not bare chips; start with a sales conversation at etched.com.
- Scope your transformer inference workload (prefill vs decode, context lengths, throughput targets).
- Evaluate the cluster against your current GPU fleet on cost per token and power efficiency.
- Plan rack, power, and networking for the co-designed cluster hardware.
Prerequisites
A large-scale transformer inference workload; enterprise engagement with Etched.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: $1B in booked orders and $800M+ raised disclosed on the live etched.com site; TechCrunch-tracked funding rounds at $5B, $10.3B, and $21B valuations; ~400 engineers hired from NVIDIA, Google TPU, Broadcom. Strengths: real silicon from TSMC, named first customer (Jane Street), strong capitalization. Docked for: (1) performance claims are company-stated with no independent benchmarks; (2) transformer-only architecture is a narrow bet that excludes MoE and multimodal workloads. License: proprietary.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 9.0/10
- · $1B in booked contract orders and $800M+ raised disclosed on live etched.com; $21B valuation at Aug 2026 Series D (Jane Street lead, also first customer)
- · Sohu transformer-only ASIC on TSMC 4nm, A0 silicon validated; first racks planned summer 2026
- · Team of 400+ engineers hired from NVIDIA, Google TPU, Broadcom, SK Hynix, TSMC
- · Performance claims (500k+ tokens/sec on Llama 70B per 8-chip server) are company-stated, no independent benchmarks
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.