Etched
Etched builds Sohu, an application-specific integrated circuit designed to do one thing: transformer inference. It targets inference, not training.
Highlights
- Sohu — transformer-specific ASIC chip optimized for inference (not training)
- Founded 2022 by Gavin Uberti and Chris Zhu in Cupertino, California
- $120M Series A led by Primary Venture Partners
- Individual investors include Peter Thiel and Stanley Druckenmiller
- Targets transformer inference at lower cost per token than NVIDIA H100/B200
- Architecture bet: transformer dominance is durable enough to justify fixed-function silicon
- Compete with NVIDIA inference for hyperscale customers and AI-native inference clouds
External link — opens etched.com in a new tab. Etched is a third-party product; we are not affiliated with it.
Link checked 17 November 2026: this site responded and still names the product.
About Etched
What it is
Etched builds Sohu, an application-specific integrated circuit designed to do one thing: transformer inference. Where a GPU is programmable and therefore general, Sohu fixes the architecture in hardware, which is what makes the speed and cost claims possible. Founded in 2022 in Cupertino by Gavin Uberti and Chris Zhu, with a $120M Series A and Peter Thiel among the individual investors. It targets inference, not training.
Why it's different
The bet is specific and falsifiable: that transformers have stabilised enough that fixed-function silicon is worth building for them. If that holds, the economics are dramatic, because a chip that cannot do anything else can do that one thing far more efficiently. If the dominant architecture shifts — and diffusion language models and state-space models are both live attempts to shift it — the silicon cannot follow. This is a hardware startup competing with Nvidia, so the realistic assessment is that it is an interesting bet rather than an available product for most readers.
How people use it
This is infrastructure, not something you use directly. It matters to organisations whose costs are dominated by inference at very large volume, where a step change in tokens per pound is worth re-architecting for. For everyone else it is worth watching as a signal about where inference pricing goes, since credible specialised silicon is the main thing that would move it.
Written by the n3os team. We are not affiliated with Etched.
More Model Platforms & APIs tools
- Actualyze AIPaid
Actualyze AI is a single OpenAI-compatible layer between an organisation's applications and the model providers that governs and secures every request.
- AurikoFreemium
Auriko is an LLM gateway: one OpenAI-compatible endpoint in front of more than a dozen model providers, routing requests by price, latency and reliability.
- HirundoPaid
Hirundo removes the effect of particular training data, such as bad labels or copyrighted material, from an already-trained model without full retraining.
- Hydra HostPaid
Hydra Host rents dedicated, non-virtualised NVIDIA GPU servers for AI, HPC and data workloads, across a network of more than forty data centres.
- Inception LabsPaid
Inception Labs is a Stanford spinout building Mercury, a family of diffusion-based language models that refine the whole output in parallel.
- NebiusPaid
Nebius operates NVIDIA GPU clusters for training and inference, with managed Slurm and Kubernetes and an inference API on top. It is listed on Nasdaq.
Every Model Platforms & APIs tool in the directory →
Listed by how many tags they share with Etched, then by name. Nobody pays to appear here.
This listing was written from public information, without Etched’s involvement. If you own it and something here is wrong — or you would rather not be listed at all — email us and, once we have confirmed it is yours, we will correct or remove it.
Get the ones worth knowing about
We write one of these for every tool worth the trouble. Get the new ones, plus what we have found genuinely useful lately.
Your address goes to Buttondown, who send the emails on our behalf. One click unsubscribes, and the list is never sold or shared.