A startup that has spent years in stealth mode just walked into the AI hardware arena and essentially said, “we’re here now.” Etched revealed its Sohu chip and a rack-scale inference system on June 30, purpose-built for transformer-based large language model inference, claiming state-of-the-art results in throughput, latency, and power efficiency.

The company isn’t just shipping slides and promises. Etched says it has secured over $1 billion in signed customer contracts, with first rack shipments scheduled for this summer and plans to scale production toward gigawatt capacity by 2027.

What Etched actually built

The Sohu chip is an application-specific integrated circuit, or ASIC. Instead of being a general-purpose processor that can do many things reasonably well (like Nvidia’s GPUs), it’s designed to do one thing exceptionally well. That one thing is transformer inference, the process of running trained AI models to generate outputs.

The chip was fabricated using TSMC’s N4P process node and achieved first-pass silicon success. That’s a notable detail for anyone who follows chip development. First-pass success means the design worked correctly on its initial manufacturing run, something that saves enormous amounts of time and money in the notoriously unforgiving world of semiconductor development.

Etched says preliminary customer tests have validated performance claims on several prominent AI models, including Llama, DeepSeek, Qwen, and Mamba. The system is designed as a complete rack-scale solution rather than just a standalone chip, which means Etched is selling integrated infrastructure, not just silicon.

The money and the people behind it

Etched has raised $800 million in total funding. The most recent round brought in $500 million at a valuation of $5 billion.

The investor list includes Jane Street, Peter Thiel, and Geoffrey Hinton, widely regarded as one of the godfathers of deep learning.

The company operates with a team of over 400 engineers, many recruited from Nvidia and TSMC.

Why this matters for the AI hardware market

The AI industry is entering a phase where inference costs matter as much as, or more than, training costs. Training a frontier model is a one-time (well, periodic) expense. Running that model billions of times per day for users, applications, and autonomous agents is the ongoing cost that actually determines profitability.

The $1 billion in signed contracts before shipping a single production rack suggests that potential customers have seen enough in testing to commit real capital. The trajectory Etched is describing—from summer 2026 shipments to gigawatt-scale production in 2027—is aggressive. Gigawatt-scale data center capacity is the kind of language typically reserved for hyperscalers like Microsoft and Google.

Nvidia’s moat isn’t just hardware. It’s the CUDA software ecosystem that makes its chips easier to program and deploy. Etched will need to prove that the performance advantages of its ASIC approach are large enough to justify customers adopting new toolchains and workflows.