Etched

Roadmap & Position in AI Silicon

Transformer-only ASIC and inference supercomputer stack positioned against Nvidia B200.

Company Overview

Etched is a semiconductor company that makes Sohu, a transformer-only ASIC, plus the full rack-to-cluster inference supercomputer around it. Customers are frontier-model inference buyers: hyperscalers, LLM API providers, and enterprise AI teams running Llama-class and GPT-class workloads.

What They're Building

The company's public product roadmap & what they're committed to building.

Sohu Chip

A transformer-only ASIC on TSMC 4nm with 144GB HBM3E, targeting over 10x B200 performance for transformer inference.

Sohu Developer Cloud

A hosted environment for prospective customers to benchmark workloads on Sohu before taking delivery.

Supercomputing Stack

Rack and multi-rack cluster software covering orchestration, RDMA networking, telemetry, and fleet reliability for customer deployments.

Latest Intelligence

Zeitgeist tracks private signals to determine where the company is heading strategically.

Competitors

Groq:

Also inference-only ASIC, but their LPU is a more general language-model accelerator rather than a transformer-graph-in-silicon bet.

Cerebras:

Wafer-scale engine optimized for training and inference breadth, competing on scale rather than transformer specialization.

NVIDIA:

The incumbent Etched benchmarks against explicitly; flexible, software-mature, and the default choice Etched has to displace.

Etched

's Moat:

Architectural specialization combined with vertical integration: hardwired transformer math in silicon, plus owning the pod, rack, and cluster software layer that hyperscalers would otherwise capture, which is what makes the cost-per-token math defensible against NVIDIA flexibility.

How They're Leveraging AI

AI Use Overview:

Etched builds a transformer-specific ASIC architecture (Sohu, on TSMC 4nm with 144GB HBM3E, targeting over 10x B200 performance for transformer inference) and an inference software stack optimized for LLM serving and cost-per-token reduction rather than general-purpose compute.

More Similar Companies

Aurorin CAD

Lets mechanical engineers design parts in seconds instead of hours with AI-native CAD.

SolidWorks has not been meaningfully challenged in 30 years. Aurorin built a parametric kernel from scratch with an AI chat interface, and a 3x SpaceX intern with GPU engineering experience is exactly the right person to do it.

BaseFrame

Helps hardware engineers find components in seconds with AI search inside Slack and Teams.

Hardware engineers waste hours cross-referencing datasheets across distributor sites. BaseFrame embeds a parts copilot directly in Slack and Teams, which is where the buying conversations already happen.

Chasi

Deploys 24/7 AI agents that automate sales, rental, and service for equipment dealers.

Industrial equipment dealers run sales, rental, and service across phone, email, and web chat with no automation. Chasi deploys 24/7 AI agents across all channels, and the founder led AI deployments at Cummins and Harley-Davidson, so he already has the Rolodex.

DAIVIN!

Builds tankless dive gear that extracts breathable oxygen from water via electrolysis.

Tankless dive gear using electrolysis to extract oxygen from water. If the physics works at depth, it replaces scuba tanks entirely. The founder is a Finnish electrical engineer with the country's highest certification and military diving experience. Moonshot hardware with multi-domain potential (sea, land, space).