Etched doubles valuation to $21B in three weeks as Jane Street leads $700M round

Etched has raised $700 million at a $21 billion valuation, led by Jane Street. The round closed on August 18, three weeks after the company's $300 million Series C at $10.3 billion, led by Sequoia Capital on July 23. That is a doubling of valuation in less than a month.

Kleiner Perkins, Sequoia, Andreessen Horowitz, Tiger Global, SK Hynix, and Diffusion Capital also participated. Etched has now raised $1.9 billion to date.

The San Jose-based company, founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen, builds specialized AI inference hardware. Its chip, Sohu, is an application-specific integrated circuit designed to run only transformer models. Etched claims the chip delivers 20x the throughput of Nvidia's H100 GPUs on inference workloads by hard-coding transformer attention mechanisms directly into silicon.

Jane Street is both lead investor and Etched's first customer. The quantitative trading firm received its first server rack last month and is now deploying Sohu in its own workloads. That makes Jane Street a live proof point, not just a financial backer.

"Inference is becoming one of the most important infrastructure markets in AI, and the winners will be measured by tokens per dollar and per watt," said Kleiner Perkins Managing Partner Mamoon Hamid.

Etched now employs more than 400 people, drawing engineers from Nvidia, Broadcom, Google's TPU (Tensor Processing Unit) team, and SK Hynix. The company reported first-pass silicon success on TSMC's N4P process and operates a 2-megawatt data center at its San Jose headquarters, with a new 80,000-square-foot, 10-megawatt facility in nearby Milpitas.

The company said it has secured more than $1 billion in customer contracts across public and private AI companies and cloud providers, though it has not named any customer beyond Jane Street.

Why the valuation is moving this fast

The jump from $10.3 billion to $21 billion in three weeks reflects a specific market dynamic. The inference market is projected to reach $1.3 trillion by 2032, according to Bloomberg Intelligence, doubling the size of the AI training market. Iron Mountain and Structure Research expect inference compute to overtake training capacity this year and account for 80% of total AI compute load by 2030.

Etched is positioned at the center of that shift. The company splits inference into two phases: a prefill phase that reads and interprets prompts using low-voltage transistors that run cooler at higher clock speeds, and a decode phase that generates output tokens one at a time using a shared memory design that lets accelerators in one rack draw data from adjacent memory pools without copying.

The company says its systems are already running DeepSeek, Qwen, Mamba, and Llama models. That range matters. A chip that only works for one model family is a research project. A chip that runs the most widely deployed open model architectures is a product.

The Nvidia question

The broader context is the rush for Nvidia alternatives. Etched's valuation explosion sits alongside a separate deal where Groq raised at $3.5 billion, roughly half its prior mark, after signing a supply deal with Nvidia. The contrast is telling. Companies that partner with Nvidia see their valuations compress. Companies that compete on a different architectural bet see theirs expand.

Etched's bet is that specialization beats generalization at inference scale. Nvidia's GPUs handle both training and inference. Sohu handles only inference, and only for transformers. If transformer architecture remains dominant, that specialization is a permanent advantage. If the architecture shifts, the chip becomes obsolete. That is the binary risk every buyer and investor is pricing.

"This round reflects a growing industry conviction that the challenge demands a new entrant willing to rebuild the stack from first principles," said Etched Chief Executive Officer Gavin Uberti.

Source