Breaking

Etched Discloses $800M Raise and $1B in Sohu Sales Contracts

June 30, 2026 at 16:00 EDT

Infra & Chips
Foundation Models
Funding & M&A

Etched, the AI chip startup building the Transformer-only ASIC called Sohu, said on June 30 it has raised a cumulative $800 million and signed more than $1 billion in system sales contracts. Positioning itself as an inference-focused challenger to NVIDIA's GPUs, the company plans to ship its first rack-scale product in the summer of 2026.

June 30, 2026 · Etched

A "Transformer-Only" Chip Exits Stealth to Challenge NVIDIA

AI chip startup Etched emerges with $800M raised and over $1B in customer contracts, betting that its Sohu ASIC — with the attention mechanism burned directly into silicon — can outrun general-purpose GPUs on inference.

$800M

raised across four undisclosed rounds

$1B+

signed in customer contracts

$5B

post-money valuation (late 2025)

400+

engineers from NVIDIA, TPU, TSMC

Claimed throughput — Llama 70B inference

Tokens per second per server (FP8, 2048 in / 128 out). Self-reported by Etched; no independent benchmarks published.

~500,000

8× Sohu

~45,000

NVIDIA B200
10×+ slower

~23,000

NVIDIA H100
20×+ slower

One 8× Sohu server is positioned as equivalent to 160 H100 GPUs.

The bull case

Logical consistency of transformer specialization
Deep NVIDIA-alumni engineering talent
A0 silicon on TSMC N4P already operational

The skeptics

No independent third-party benchmarks yet
High migration cost — incompatible with CUDA
Locked to transformers only; no flexibility
Past dedicated-ASIC failures (e.g. Graphcore)

What's next

First Sohu shipments are slated for summer 2026. With NVIDIA still holding 86–92% of the inference market, the focus now shifts to volume production and publishing real benchmarks to back its claims.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.