Etched raised $300 million at a $10.3 billion valuation on Thursday, July twenty third, a round that roughly doubles the value of a chip startup much of the industry wrote off two years ago. The money scales Sohu, an inference chip that does one thing on purpose, it hardwires the transformer, the architecture behind nearly every large language model, straight into silicon. Sequoia led, with Andreessen Horowitz, Jane Street, Diffusion, and memory maker SK hynix joining, and the total raised now passes one billion dollars. For anyone who sizes inference budgets, the bet is simple to state and hard to make. Here is what Etched built and where the risk sits.
The short answer
On Thursday, July twenty third, Etched raised $300 million at a $10.3 billion valuation. Sequoia led, with Andreessen Horowitz, Jane Street, Diffusion, and SK hynix joining, and total funding now passes one billion dollars. The money scales Sohu, an inference chip that hardwires the transformer into silicon rather than running general purpose math like a GPU. First rack shipments are due in summer 2026. The bet is that the transformer stays dominant long enough for fixed hardware to pay back, a wager that looked reckless two years ago and now carries one of Sequoia's richest Series C valuations.
Two years ago the idea sounded reckless. Build a chip that runs one thing, the transformer, and nothing else, and sell it against the most flexible processors in the world. Plenty of people in the chip industry said it could not work. On Thursday, July twenty third, Etched answered with a $300 million Series C at a $10.3 billion valuation, roughly double its last mark, and total funding now above one billion dollars. Sequoia led, joined by Andreessen Horowitz, Jane Street, Diffusion, and memory maker SK hynix. For anyone who plans inference capacity, the raise is a signal about where the hardware conversation is heading.
A chip that bets on one shape
Sohu is an ASIC, an application specific integrated circuit, which is a fancy way of saying a chip built for exactly one job. A GPU is the opposite. It runs almost any math you throw at it, which is why it powers everything from games to model training, and that flexibility is precisely what costs it efficiency on any single workload. Etched took the other road. Sohu hardwires the transformer, the attention step, the linear projections, softmax, and layer normalization, directly into fixed silicon.
The wager underneath is a claim about the field. Nearly every large language model in production today is a transformer. If that stays true, a chip that does only transformers can serve those models with far more tokens per watt than a general purpose GPU, because none of the silicon is spent on flexibility nobody is using. That is the whole pitch, and it is why the round drew the names it did.
The money and who is behind it
The headline is the valuation, $10.3 billion, up sharply from the prior round, and reported as the highest for a Series C that Sequoia has led. Andreessen Horowitz and Jane Street returned the confidence with capital, Diffusion joined, and the detail worth pausing on is SK hynix. Inference chips are gated by memory bandwidth as much as by compute, so a major memory supplier taking a stake reads as more than a financial bet. It hints at a supply relationship for the fast memory these chips depend on.
Where the risk sits
Here is the part worth reading twice, because the strength and the risk are the same fact. Sohu is fast because it is fixed. Fixed hardware cannot adapt. If a new architecture ever displaces the transformer the way the transformer displaced what came before it, a chip that runs only transformers loses its advantage the day that happens. A GPU absorbs that shock by running whatever arrives next. Etched has no such hedge, and that is by design.
So the real question is not whether Sohu is fast. It is whether the transformer stays the dominant architecture long enough for specialized silicon to earn back its cost. So far the transformer has held for years, through wave after wave of larger models, and investors are betting that streak continues. That is a defensible bet in 2026, but it is a bet, and honest coverage should name it as one rather than treat the valuation as proof.
What it means for people who buy inference
You are not going to rack a Sohu system next week, and for homelabs it is not on the table at all. First shipments are set for summer 2026, aimed at operators serving transformer inference at high volume, where cost per token and tokens per watt decide whether a product makes money. The signal for everyone else is the direction. Specialized inference silicon is now well funded enough to pressure general purpose GPUs on price and efficiency, and more competition at the chip layer usually flows downstream into cheaper inference. Track it as a market signal, and keep your serving stack loose enough to move if the economics shift.
Sources and further reading
- TechCrunch: AI chip startup Etched defies skeptics, hits $10.3B valuation from big-name investors
- SiliconANGLE: AI chip startup Etched more than doubles valuation to $10.3B in new $300M round
- Tech Startups: Etched raises $300M at $10.3B valuation to take on Nvidia
- MLQ: Etched raises $300M Series C at $10.3B valuation to scale GPU free inference chips
Frequently asked questions
What did Etched announce on July twenty third?
Etched closed a 300 million dollar Series C that values the company at 10.3 billion dollars, roughly double its prior valuation. Sequoia led the round, with Andreessen Horowitz, Jane Street, Diffusion, and SK hynix joining. The company said the total it has raised now passes one billion dollars. The money funds production and customer deployments of Sohu, its inference chip, with first rack shipments scheduled for summer 2026. Reports noted it as the highest valuation recorded for a Series C that Sequoia has led, a strong vote of confidence in a specialized hardware bet.
What is the Sohu chip and how is it different?
Sohu is an ASIC, an application specific chip, that hardwires the transformer directly into silicon. A general purpose GPU runs almost any kind of math, which is flexible but leaves performance on the table for any single workload. Sohu does the opposite. It bakes the transformer computation, the attention step, the linear projections, softmax, and layer normalization, into fixed hardware. Because nearly every large language model today is a transformer, Etched argues a chip built only for that shape can serve those models far more efficiently per watt than a flexible GPU, at the cost of running nothing else.
Who invested and what is the valuation?
The 300 million dollar Series C was led by Sequoia Capital, with Andreessen Horowitz, Jane Street, Diffusion, and memory maker SK hynix participating. The round set the valuation at 10.3 billion dollars, up sharply from the prior round, and brought Etched's total funding above one billion dollars. SK hynix taking part is notable because inference chips live or die on memory bandwidth, so a major memory supplier at the table signals more than a purely financial stake.
What is the main risk in the Etched bet?
The risk is the same as the strength. Sohu is fast because it is fixed, but fixed hardware cannot adapt if the field moves. If a new model architecture displaces the transformer the way the transformer displaced what came before, a chip that only runs transformers loses its edge overnight. GPUs hedge that risk by running whatever comes next. Etched is betting that the transformer stays dominant long enough for specialized silicon to pay back. So far the transformer has held, which is why investors keep writing larger checks.
When will Sohu ship and who is it for?
Etched said first rack shipments are scheduled for summer 2026, with the new funding going toward expanding production and customer deployments. The audience is operators running large scale transformer inference, the teams serving chatbots, coding assistants, and agents at high volume, where cost per token and tokens per watt decide the economics. For smaller teams and homelabs the chip is not a near term option, but the direction, specialized inference silicon competing with general purpose GPUs, is worth tracking because it shapes where inference pricing goes next.