Etched, a San Jose semiconductor startup that builds chips wired to run only transformer-based AI models, is in talks for a funding round that would value it at $20 billion, the Wall Street Journal reported July 18 — four times the $5 billion valuation it set with a $500 million round last December.
The new round is being led by Jane Street, an existing investor that has put more than $100 million into the company to date, according to the report. Neither that deal nor a separate, concurrent round led by Sequoia Capital at a lower $10 billion valuation had closed as of the report, and terms on both could still change, the Journal said, citing people familiar with the matter. Running two parallel negotiations — locking in a lower price with one backer while shopping a higher one to another — has become a common tactic among fast-growing AI chip startups this funding cycle.
Betting against general-purpose GPUs
Etched was founded in 2022 by Harvard dropouts and Thiel Fellows Gavin Uberti, Chris Zhu and Robert Wachen on a narrow bet: that hardwiring a chip to run only transformer math, instead of the general-purpose flexibility of an Nvidia GPU, would make AI inference — running an already-trained model to answer a query — cheaper and faster. Its Sohu chip, manufactured by TSMC, is the product of that bet.
By its own account, Etched has raised $800 million to date from backers including Jane Street, Peter Thiel, Stripes and Ribbit Capital, with angel checks from AI researchers Andrej Karpathy, Geoffrey Hinton and Fei-Fei Li. The company says it has booked more than $1 billion in customer contracts, opened a factory in Taiwan, and expects its first racks to ship this summer.
A crowded race for Nvidia’s inference business
Etched is one of several companies chasing Nvidia’s dominance in AI hardware with custom silicon, alongside efforts such as OpenAI’s inference chip built with Broadcom. Nvidia still supplies the large majority of AI accelerators in use today, but investors have grown more willing to fund alternatives as inference — rather than training — becomes the larger, more persistent cost for AI companies serving users at scale.