First Jane Street Sohu Deployment Drives $21B Etched Valuation

AI hardware manufacturer Etched secured a $700 million Series D funding round led by quantitative trading firm Jane Street. The investment, detailed in the August 18 announcement, pushes the company’s valuation to $21 billion. This marks a $10.7 billion increase from its previous valuation set just 26 days earlier on July 23.

The round shifts Jane Street from an early investor to the startup’s first major production customer. Additional participants in the Series D include Kleiner Perkins, Sequoia, Andreessen Horowitz, Tiger Global, Bain Capital Ventures, Blackstone, Neo, Primary, Stripes, Positive Sum, Diffusion, Argo, and Peter Thiel. The company’s total funding to date now stands at $1.9 billion.

First Production Deployment

The valuation surge directly follows Jane Street’s installation and testing of the first production-ready inference rack. Etched shipped the hardware to the trading firm in July 2026. Jane Street confirmed the rack is now fully operational within its own data center and is supporting high-precision, demanding workloads.

Etched focuses exclusively on the inference stage of model deployment. If you evaluate what AI inference is at scale, the primary bottlenecks are memory bandwidth and compute utilization. The company’s flagship chip, Sohu, is manufactured using TSMC’s 4-nanometer (N4P) process.

Unlike general-purpose GPUs, Sohu hardwires the Transformer architecture directly into the silicon. This architectural decision creates a strict tradeoff. The system abandons future architectural flexibility for extreme power efficiency and processing speed. The hardware optimizes both the prefill and decode stages of generation using two newly designed cluster components.

Etched claims this specialization allows Sohu to run certain Transformer models up to 20x faster than an NVIDIA H100 GPU. Independent benchmarks are not yet available to verify this performance claim. Competitors have questioned the long-term viability of baking a single architecture into silicon as model designs evolve. However, the immediate enterprise demand for pure inference throughput continues to drive capital toward specialized hardware, echoing market moves like the $300M SN50 chip order validating SambaNova’s ASIC-native cloud.

Order Backlog and Talent Acquisition

The company reports an order backlog exceeding $1 billion in signed contracts from public cloud providers and private AI organizations. To execute on these manufacturing orders, Etched has built a 400-person workforce heavily weighted with former NVIDIA talent.

Alumni from the incumbent chipmaker comprise roughly 15% of the staff. This includes VP of Platform Brian Loiler, a 22-year NVIDIA veteran who previously contributed to the development of the HGX and DGX hardware systems.

If you manage large-scale Transformer deployments, the arrival of production ASICs changes your infrastructure math. Dedicated inference hardware forces a decision between maintaining flexible GPU clusters for diverse model experiments or locking into specialized silicon to drastically reduce the cost and latency per generated token.