Etched

Product & Competitive Intelligence

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

Technology

|

AI Silicon

|

Series B

|

Valuation:

~$5B

Company Overview

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

Latest Intel

Zeitgeist tracks private signals to determine where the company is heading and what it means competitively.

May 8, 2026
|
Tech Stack and Architecture
|
Confidence:
High

New Intel: Etched is building kernel drivers with SR-IOV, VFIO, and device attestation support. Those are the features you need to sell into multi-tenant clouds, so the intended buyer is a hyperscaler or neocloud, not a research lab picking up a dev kit.

May 8, 2026
|
Go-to-Market Motion
|
Confidence:
High

New Intel: Etched is building customer RMA workflows, 8D/CAPA reporting, and Taiwan ODM ramp capability right now. You do not build that apparatus until silicon is about to reach paying customers, so first shipments are a near-term event.

May 8, 2026
|
Product Direction
|
Confidence:
High

New Intel: Etched is building a water-cooled rack with its own RDMA fabric, collectives library, and Nsight-equivalent profiler. Their competitor is a DGX pod, which means the sales motion will look more like Supermicro than like selling cards to developers.

View All The Latest Signals

What They're Building

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

Sohu Chip:

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

Sohu Developer Cloud:

Hosted environment for prospective customers to benchmark their 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.

Competitive Landscape & 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.

Direct 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.

Add To Your Portfolio

Founding Team

Gavin Uberti

CEO and Co-Founder (ex-OctoML, Harvard)

Chris Zhu

CTO and Co-Founder (ex-Optiver, Harvard)

Mark Ross

Technical Leadership (ex-Cypress Semiconductor CTO)

Ajat Hukkoo

Senior Technical Leader (ex-Broadcom, ex-Cypress)

Founder Force Multiplier

Uberti brings ML compiler and systems experience from OctoML, Zhu brings latency-obsessed HPC engineering from Optiver, and they recruited silicon veterans from Cypress and Broadcom to handle what two Harvard dropouts cannot. That pairing of software-first architectural bets with hardened ASIC execution is the right shape for a transformer-only chip company also building its own deployment stack.

Funding History

2023 | Seed, $5.4M led by Primary
2024 | Series A, $120M led by Primary
2026 | Series B, $500M led by Stripes

Back To All Companies >