AI Chip Startup Etched Nears $20 Billion Valuation in New Funding Talks

American startup Etched, maker of the Sohu chip built specifically for running language models, is negotiating two funding rounds at once, including one at a valuation reaching $20 billion. That's four times the company's valuation from just seven months ago, when it closed a $5 billion round in December 2025.

From the start, Etched has bet exclusively on one segment of the AI chip market: inference, meaning running already-trained models rather than training them. That's where the company sees an opening to challenge Nvidia's dominance, which is centered mainly on training and general-purpose chips.

What Sohu Is

The company's flagship product is the Sohu chip, built on a 4-nanometer process and designed from the ground up to run only the transformer architecture that underpins most of today's large language models, including Llama and the GPT family. By giving up the general-purpose programmability that Nvidia's GPUs offer, Etched's designers were able to concentrate transistors on operations central to transformers, such as attention, projections and feed-forward layers.

The company says this specialization lets Sohu run up to 20 times faster than Nvidia's H100 when serving the Llama 70B model, while cutting cost and energy use per query. It unveiled its first rack-scale system built around the chip on June 30, 2026, and says deliveries to customers will begin this summer.

A Billion Dollars in Orders Before Launch

Even though the chip is still going through validation, the company says on its website that it has more than $1 billion in signed contracts from customers looking to deploy Sohu to run their AI systems. That level of pre-launch demand for a product that hasn't yet reached mass production is rare even in the red-hot AI infrastructure sector.

$1 billion in customer demand for a chip still in the validation stage is the clearest signal yet that corporate buyers aren't waiting for inference silicon to mature before committing to it.

Rounds, Back to Back

The pattern of a startup selling shares at one valuation and almost immediately raising fresh capital at a much higher one has become, according to the Wall Street Journal, a hallmark of the current AI investment cycle. It reflects the negotiating leverage that leading AI companies hold over investors competing for a spot in the next round before valuations climb even higher.

Among Etched's existing investors, besides Sequoia Capital and Jane Street, are funds including Stripes, Ribbit Capital and Primary Venture Partners, as well as billionaire Peter Thiel. Its angel investor list includes well-known names from the AI industry, among them Andrej Karpathy and Geoffrey Hinton.

Competition for the Inference Niche

Etched isn't the only startup trying to take a bite out of Nvidia's share of the inference market. Cerebras Systems, Groq, Fractile and SambaNova are active in the same segment, and several of them have also closed large funding rounds in recent months. Nvidia, however, remains by far the dominant supplier of AI infrastructure, both for training and for production deployments.

For tech companies and cloud providers, growing competition in the inference segment could mean cheaper, more energy-efficient options for running AI models at scale. With compute bills climbing steadily, that could eventually have a real impact on the cost of AI-powered services, including those offered in European and Polish markets.