Breaking

AI Chip Startup Etched Reaches $5B Valuation, Secures $1B in Orders

June 30, 2026 at 14:17 EDT

Infra & Chips
Funding & M&A
Foundation Models

Etched, a startup building AI chips specialized for the Transformer architecture, said on June 30, 2026 that it has already secured roughly $1 billion in customer contracts for inference systems and reached a post-money valuation of $5 billion. The company plans to ship the first racks of systems powered by its inference-only chip, Sohu, in summer 2026, and is moving through production and customer validation after TSMC manufactured its A0 silicon in the first half of the year.

Transformer ASIC · Etched / Sohu

A chip that runs only Transformers — built to outrun the GPU

Etched's first silicon for Sohu has returned from TSMC's 4nm line. The startup is now valued near $5 billion with over $1 billion in customer demand — and claims throughput that dwarfs today's flagship GPUs. First rack ships summer 2026.

~$5B

Valuation after latest raise

$1B+

Customer contract demand

500K+

Tokens/sec, Llama 70B (8-chip)

~400

Engineers (NVIDIA, TPU, TSMC alumni)

Claimed throughput vs GPU servers

One 8-chip Sohu server, relative to a standard server — company figures, unverified

1×

H100 server

~2×

B200 server

20×+

Sohu
8-chip

vs H100

One Sohu server is claimed to match roughly 160 H100 GPUs on Llama 70B.

Why bet everything on one architecture

Transformer-only ASIC on TSMC N4P, HBM3E memory
Low Voltage Inference (LVI) lifts FLOPs density
Cluster Scale Memory (CSM) for low-latency chip sharing
Single-purpose focus targets throughput, cost & power

Open questions

All benchmarks are company-provided & unverified
Still pre-shipment at A0 silicon validation stage
No CNNs or non-Transformer models — concentration risk
Pricing and detailed benchmarks undisclosed

2026

First rack ships summer 2026

Customers are validating the first rack-scale system — chip, server and software bundled. The summer shipment will be the real test of these claims.

Continue reading

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