A Silicon Valley chip startup called Etched has announced cumulative funding of 800 million dollars, a valuation of 5 billion dollars, and a single chip order worth 1 billion dollars. For a company that has kept a low profile since it was founded in 2022 by three Harvard dropouts, the investor list is the headline.
Nobel laureate Geoffrey Hinton is in. So is Stanford computer vision pioneer Fei-Fei Li, OpenAI co-founder Andrej Karpathy, venture capital’s Peter Thiel, quantitative trading house Jane Street, and a fund affiliated with TSMC.
What they are backing looks extreme on paper. A chip that runs Transformers and nothing else.
The bet: burn the architecture into silicon
Sohu is Etched’s first chip, built on TSMC’s 4-nanometre process, and it has already taped out successfully. The company claims that on Llama 70B inference, a server holding eight Sohu chips replaces 160 NVIDIA H100 units.
That claim points at a deeper question. Is the general purpose era of AI silicon coming to an end?
To understand what Etched is doing, start with the difference between general and specialised. NVIDIA’s GPU is fundamentally a general purpose chip. It runs Transformers, it runs convolutional networks, it runs recurrent networks, it can even mine Bitcoin. Being able to do everything is the root of NVIDIA’s success, because whatever the next AI wave turns out to be, the GPU will be useful for it.
The price of general purpose is wasted efficiency. When a GPU runs a Transformer, a large share of its transistors are doing something other than that job. It is like delivering takeaway food in an all-terrain vehicle. It arrives, but the fuel bill is absurd.
Etched’s answer, in one line, is to carve the Transformer architecture into the hardware. Sohu is an application specific integrated circuit. Unlike a GPU it was not designed for programmability. The circuits are fixed and they do one thing, which is forward inference on Transformer models.
That means Sohu cannot run a convolutional network, cannot run a recurrent network, and cannot even run a Transformer operation that does not match Etched’s hardware implementation. The extreme trade-off buys extreme efficiency. Etched claims Sohu delivers 20 times the throughput of an H100 on Llama 70B inference, and 140 times the performance per dollar of a GPU.
Why the trade-off is defensible now
Sohu’s breakthrough is not in the process node and not in raw compute scale. It is in the nerve required to make an architectural bet.
Chip design has an iron law. The more general the chip, the larger the market and the lower the efficiency. The more specialised the chip, the higher the efficiency and the smaller the market. An ASIC sits at the extreme of efficiency and therefore at the extreme of risk, because if the algorithm architecture changes, a specialised chip becomes expensive scrap silicon.
Etched can make that bet only because it judges the Transformer to be stable enough to deserve dedicated hardware. That judgement is not unreasonable. The architecture has dominated mainstream AI for more than five years, from language models to text-to-video, and the foundation is unlikely to be overturned quickly.
The more immediate driver is cost. Large operators are absorbing enormous inference bills, with OpenAI alone spending billions of dollars a year on that line item. That pressure makes the market intensely hungry for anything that cuts cost per token.
The 1 billion dollar pre-order is the strongest available validation. Signing a contract of that size before volume production means the customer has almost certainly run real workloads on early prototypes and liked the result.
Why serious investors backed the team
Semiconductors are one of the few industries where a company can raise on a deck and fail to ship a product for a decade. Getting Hinton, Karpathy and Fei-Fei Li to write cheques says something about the founding team.
All three founders were Thiel Fellows. The programme, created by Peter Thiel, funds students who drop out to build companies, offering 100,000 dollars on the condition that they leave university and work full time. Past alumni founded Ethereum and Linear Technology among others.
Chief executive Gavin Uberti worked in high frequency trading at Jane Street, which gives him an unusually deep feel for low latency compute. Chief technology officer Chris Zhu interned on Google’s TPU team and watched a specialised chip travel from design to deployment.
The company now employs more than 400 engineers drawn mainly from NVIDIA, the Google TPU team, Broadcom and TSMC, covering essentially every critical link in the AI silicon chain. Four hundred chip veterans, a successful tape-out and a 1 billion dollar order is about as validated as a chip startup gets.
The CUDA problem, sidestepped
NVIDIA’s moat was never only chip performance. It is the CUDA ecosystem. Every challenger eventually faces the same question. Your chip may be faster, but will developers rewrite their code?
Etched’s clever move is that it does not ask them to. Sohu runs Transformer models directly, and a model trained in PyTorch runs on Sohu without code changes.
That is rare in ASIC history. Specialised chips usually mean a new programming model, a new software stack and a new learning curve. Sohu avoids that because the Transformer ecosystem is already unified. Model formats are standard, inference frameworks are mature, and a dedicated chip only needs to meet that common layer.
For NVIDIA that is an unfamiliar kind of challenge. Its core advantage is general purpose plus ecosystem. If customers discover that a specialised chip is faster, cheaper and requires no code changes, the premium attached to generality shrinks fast.
Three routes, one shakeout window
The global AI silicon chain is splitting into three routes.
The first is the general purpose GPU route, run by NVIDIA and AMD, selling shovels to everyone regardless of what they are digging for. Mature ecosystem and broad applicability on one side, low efficiency and high cost on the other.
The second is the cloud in-house route, covering Google TPU, AWS Trainium and Meta’s own silicon. Deeply tied to the owner’s business with controllable cost, but a closed ecosystem and limited technology spillover.
The third is the third party specialised route, which includes Etched, Cerebras and Groq. Optimised for a specific computation and sold to everyone. Extreme efficiency on one side, concentrated risk on the other.
Most people in the industry expect all three to coexist for a long time while market share is redistributed. The next two to three years look like the window in which the shape of that redistribution is decided.
Editor’s note: This is an adapted translation of the original Leiphone report. It has been trimmed and restructured for readability for an international business audience.