Etched Is Fielding Bids at $40 Billion, Twice Its August Price

In December 2025, Etched was worth $5 billion. In July 2026 it was worth $10.3 billion. In August it was worth $21 billion. Now, according to TechCrunch, the inference-chip startup is reviewing bids that would value it at $40 billion or more, which would be the third straight time its price has roughly doubled.

TechCrunch reported on October 5, citing people familiar with the company, that the incoming offers range from $40 billion from top-tier investors to $50 billion from lesser-known backers. The talks are early and the terms of any deal may change. Etched declined to comment.

Three rounds in a year, four if this one closes

The pace is the story. TechCrunch reported on July 23 that Etched had closed a $300 million Series C at $10.3 billion, led by Sequoia, with Andreessen Horowitz, SK Hynix, Jane Street and Diffusion Capital participating. The company said it was the highest valuation ever for a Sequoia-led Series C. Its previous mark had been $5 billion, set when it raised $500 million in December 2025.

Then, on August 18, Etched raised $700 million at $21 billion. The Next Web reported that Jane Street led that round, and noted that the valuation had doubled in 26 days. If a new round lands at $40 billion, the company will have gone from $5 billion to eight times that in under a year.

Back-to-back rounds at separate valuations have become a pattern among the most contested AI startups, as TechCrunch noted. The logic for investors is simple: if the price will be higher in six weeks, the cheapest entry point is the one in front of you. TechCrunch also pointed out what the cash buys. If Etched raises as much as it did in August, it could have as much as 3.5 years of runway, a cushion that matters for a company building complete hardware systems, one of the most capital-hungry corners of AI.

The customer is also the lead investor

Jane Street's dual role explains a lot of the momentum. The quantitative trading firm was already a backer before it led the August round, and it is also Etched's first paying customer. The Next Web reported that Etched's first customer delivery was a rack shipped to Jane Street, and TechCrunch's latest report says the firm received an early system.

That arrangement is unusual for a chip company and useful for its fundraising. Hardware startups usually ask investors to believe performance claims made in a lab. Etched can point to a sophisticated buyer that tested the machine, bought it, and then wrote the largest check in the company's history. It is not independent validation, and The Next Web was careful to note that the August announcement came without published benchmark numbers. But it is a stronger signal than a slide deck.

The order book supports the same read. Etched said in June that it had booked $1 billion in customer orders, according to TechCrunch, after TSMC manufactured its first batch of chips. The company's website now describes more than $1 billion in demand and customer contracts and says its first rack-scale product is being validated with customers.

The company quietly changed its pitch

Flux covered Etched in July as the startup that burned the transformer into silicon and bet everything on the architecture never changing. That is no longer how the company describes itself.

In its July interview with TechCrunch, co-founder and COO Robert Wachen said Etched's systems can run any AI model, including Mixture of Experts models such as DeepSeek and Qwen and non-transformer designs such as Mamba. The Next Web put it directly: the company was founded on etching a single model architecture into silicon, and it has moved away from that premise.

What replaced it is a split view of inference. "Inference is built in two stages," Wachen told TechCrunch, "prefill and decode." For prefill, the stage where a model reads the prompt, Etched built a chip that runs at a much lower voltage, which it calls low-voltage inference. For decode, the stage where tokens are generated one at a time, it designed a memory and interconnect technology it calls cluster-scale memory, which lets many chips share one fast memory pool. The company's website claims the low-voltage approach can run trillion-parameter sparse models at more than 80% of peak compute without thermal throttling. Those are Etched's figures, not third-party measurements.

The shift matters for the valuation. A transformer-only ASIC is a binary bet on one architecture. A system that targets the two physical bottlenecks of inference, compute during prefill and memory bandwidth during decode, is a bet on the workload itself. Investors pricing the company at $40 billion are buying the second story.

What $40 billion has to cover

Etched is not a fabless chip designer selling parts. It sells full systems, which means it is also in the business of racks, factories and data centers. TechCrunch reported in July that the company runs a 2-megawatt data center at its San Jose office and had opened an 80,000-square-foot, 10-megawatt facility in Milpitas. It also has a facility in Taiwan to coordinate production near TSMC. Headcount is around 400, and The Wall Street Journal reported that roughly 15% of those employees previously worked at Nvidia.

That footprint is why the money keeps coming in large tranches. The Next Web reported that the August capital was earmarked for factories, supply chains and fleet software on the path to gigawatt scale. Building toward that scale against Nvidia, which sets the price and the software standard for nearly every AI data center, takes more than one good chip.

The risk sits in the gap between orders and deployments. Etched has $1 billion in signed demand and one disclosed delivered rack. Its performance claims are still largely self-reported. A $40 billion price assumes the company converts that order book into fleets of working systems on schedule, and that customers beyond its lead investor keep buying once independent numbers exist.

For now, the market is paying for the possibility. The founders, Gavin Uberti, Chris Zhu and Wachen, left Harvard in 2022 to build a chip company. Four years later, investors are reportedly competing to pay twice what they paid seven weeks ago, before the company has published independent benchmark results.