Inference Chip Startup Etched Exits Stealth With $800M Raised and $1B in Signed Contracts

Quick Facts

• Etched raised $800M total, including a $500M round led by Stripes that closed in December at a $5 billion valuation.
• The company’s first chip, Sohu, claims a single 8-chip server outperforms 160 Nvidia H100 GPUs running Meta’s Llama 70B model.
• Etched has more than $1 billion in signed customer contracts and plans to begin shipping rack-scale inference systems this summer.

Etched Inc. came out of stealth on June 30, announcing a working chip, $800 million in total funding, and more than $1 billion in signed customer contracts. The San Jose-based company builds purpose-built inference chips and complete rack-scale systems for running transformer AI models.

The bulk of the funding came from a $500 million round led by Stripes that closed in December, valuing the company at $5 billion. Jane Street separately led a previously unannounced round and has invested more than $100 million in total, according to Bloomberg. The company closed a $5.4 million seed round in March 2023 and a $120 million Series A in June 2024.

The Chip

Etched’s first-generation product is Sohu, a custom ASIC built exclusively to run transformer model inference. The company says a single 8-chip Sohu server processes around 500,000 tokens per second running Meta’s Llama 70B model, outperforming 160 Nvidia H100 GPUs while using less power and space.

Sohu achieved first-pass silicon success on TSMC’s N4P process. Current systems run models including DeepSeek, Qwen, Mamba, and Llama. Etched developed two proprietary technologies: LVI, which reduces thermal throttling, and Cluster Scale Memory, which combines HBM with a shared low-latency memory architecture to cut inference latency.

Vertical Integration

Etched is not selling chips alone. The company designed its entire server rack, including circuit boards, cooling plates, and networking connections. Co-founder Rob Wachen says no other chip startup has done this.

The company operates a Taiwan factory and has built a data center, test house, and prototyping lab at its San Jose headquarters. It is targeting gigawatt-scale production by 2027. Its team of more than 400 people includes veterans from Nvidia, Broadcom, Google’s TPU team, and SK Hynix.

Leadership

Etched was founded in 2022 by Gavin Uberti, Chris Zhu, and Robert Wachen, three Harvard University dropouts who became among the first recipients of the Thiel Fellowship. Uberti and Wachen previously worked at Nvidia and on Google’s TPU team.

Key hires include CTO Mark Ross, former CTO of Cypress Semiconductor; VP of platform Brian Loiler, who spent 22 years at Nvidia building HGX and DGX systems; VP of ASIC and architecture Saptadeep Pal, an architect on Nvidia’s H100, A100, and V100 teams; and VP of software David Munday, who built Google’s TPU software team.

Market Position

Etched’s strategy centers on a single-purpose architecture optimized for transformer models, the same architecture powering ChatGPT, Anthropic’s Claude, and Google’s Gemini. Analysts expect the inference chip market to exceed $100 billion annually by 2030.

The founders nearly ran out of money in 2023. Major investors passed when pitched a 30-page memo arguing AI would need specialized inference chips. Etched has now gone from a $34 million seed valuation to a $5 billion post-money valuation in under three years.

Gavin Uberti, co-founder and CEO, said: “We recognized early on that frontier AI would become one of the most economically significant technologies ever created, but that the infrastructure needed to serve those models in a sustainable and economically viable way simply did not exist.”

Read more: Inference chip startup Etched launches with $800M in funding