What Etched.ai does

Etched.ai (DBA: Etched) is an AI semiconductor startup developing Sohu, claimed to be the world's first transformer-specific ASIC, designed exclusively to accelerate transformer model inference. The company's core thesis is that general-purpose GPUs like the Nvidia H100 are suboptimal for transformer workloads, and that a purpose-built ASIC can deliver an order-of-magnitude improvement in tokens per dollar and per watt for dense models, sparse mixture-of-experts architectures, and diffusion models. Etched is headquartered in San Jose, California, was founded in 2024 by Harvard dropouts Gavin Uberti (CEO), Chris Zhu (Co-Founder/Thiel Fellow), and Robert Wachen (President), and has assembled a leadership bench drawn from NVIDIA, Google DeepMind, Broadcom, Intel, and Cypress.

The company sells or plans to sell Sohu as a rack-scale AI inference system through a direct B2B enterprise sales motion targeting AI companies, hyperscalers, and cloud providers running transformer-based LLMs and generative AI workloads. Manufacturing is handled in partnership with TSMC, with HBM memory sourced from SK Hynix, Samsung, and Micron; chip design completion, manufacturing, and shipping are targeted for 2027. The company also partnered with Decart to launch Oasis, a real-time AI video generation demonstration intended to showcase Sohu's transformer inference throughput.

Etched is a venture-backed private company that has raised approximately $625M across a Series A ($120M, June 2024, led by Positive Sum and Primary Venture Partners) and a Series B ($500M, January 2026, led by Stripes) at a $5B post-money valuation, with additional participation from Peter Thiel, Stanley Druckenmiller, Ribbit Capital, Jane Street, Marvell, Spark Capital, and others. The company does not currently disclose revenue, consistent with its pre-shipment status.

Name

Etched.ai

Legal name

Etched

Website

https://etched.com

Company type

Private

Founded year

2024

Operating status

Operating

Headcount range

251–500 employees

Short description

Etched.ai develops Sohu, a transformer-specific AI ASIC chip designed to accelerate inference for large language models, sparse MoEs, and diffusion models. The company targets hyperscalers and AI enterprises via direct B2B sales, with manufacturing partnered through TSMC.

Ownership category

akta.pro rank

Product category

AI Inference Semiconductors

NAICS

Semiconductor and Related Device Manufacturing (334413), Semiconductor and Other Electronic Component Manufacturing (3344)

SIC

Electronic Computers (3571)

akta.pro primary industry

Compute & Acceleration ICs (CPU/GPU/AI/FPGA/ASIC) (HDAHABAC)

akta.pro secondary industries

AI Accelerators (GPUs/TPUs/NPUs/ASICs) (HDAAAAAA), AI Hardware IP & Chiplet/Packaging for Accelerators (EDA/IP, advanced packaging) (HDAAAAAL)

Keywords

• AI inference chips
• Transformer ASIC hardware
• AI accelerators
• Specialty semiconductors
• Rack-scale AI systems

Headquarters

HQ city

San Jose

HQ country

United States

HQ region

North America

Offices1 record

Markets served

GTM type

B2B

Offering type

Hardware or Manufacturing

Cost components

Technology or R&D, Personnel, Supply Chain, Operations, Marketing or Sales

Revenue model

1. AI Chip Hardware Sales: Sale of specialized transformer ASIC chips (Sohu) to AI companies, hyperscalers, and enterprises requiring high-performance inference compute. Manufacturing handled through TSMC partnership with memory supplied by SK Hynix, Samsung, and Micron.

Go-to-market motion1 record

Distribution channels1 record

Marketing channels1 record

Core offering

Etched.ai designs and sells the Sohu chip, the world's first transformer-specialized ASIC, purpose-built to accelerate AI inference for transformer-based models such as large language models, sparse Mixture-of-Experts, and diffusion models. The company also delivers a rack-scale AI inference system that combines the Sohu ASIC with high-bandwidth memory and partners with TSMC for chip fabrication.

Product overview

Etched.ai is a semiconductor startup developing specialized ASIC chips optimized exclusively for transformer model inference. The company's primary product is the Sohu chip - a purpose-built transformer ASIC that competes with Nvidia's general-purpose GPUs in the AI accelerator market. Sohu is designed to provide superior performance and energy efficiency for transformer-based AI workloads including large language models (LLMs), sparse mixture-of-experts architectures, and diffusion models. The company partners with TSMC for chip manufacturing and has raised significant funding ($120M Series A in June 2024, $500M Series B in early 2026) to support chip design completion with manufacturing and shipping targeted for 2027.

Differentiator

Problem solved

Functional benefit

Products and services

• Sohu Sohu is a transformer-specialized ASIC chip built by Etched.ai to run transformer-based AI models (LLMs, sparse MoEs, diffusion) with significantly higher tokens per dollar and per watt compared to general-purpose GPUs. It is sold to AI companies, hyperscalers, and cloud providers requiring efficient inference compute.
• Oasis Oasis is the first experiential real-time open-world AI model that generates interactive video entirely through AI, enabling users to move, jump, pick up items, and interact with a physics-based environment. It is powered by transformer-based AI inference that Etched's Sohu chip is designed to accelerate.

Quantifiable outcome

• 10-20x faster than Nvidia H100 GPUs for transformer inference tasks
• +2 more outcomes

Segments1 record

Ideal customer profiles1 record

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability4 records

Feature2 records

Partnerships

Four partnerships are on record, tiered core.

• TSMC (Taiwan Semiconductor Manufacturing Co.)coreStrategic or Co-development Partner · 24 February 2026Manufacturing partnership with TSMC for chip production. TSMC will manufacture Etched's Sohu transformer-specific ASIC chips, providing advanced semiconductor fabrication capabilities essential for the company's hardware strategy.
• SK HynixcoreStrategic or Co-development Partner · 24 February 2026Memory provider partnership for supplying high-bandwidth memory (HBM) used in AI chips, working alongside Samsung and Micron for memory supply chain.
• SamsungcoreStrategic or Co-development Partner · 24 February 2026Memory provider partnership for supplying high-bandwidth memory (HBM) used in AI chips alongside SK Hynix and Micron.
• MicroncoreStrategic or Co-development Partner · 24 February 2026Memory provider partnership for supplying high-bandwidth memory (HBM) used in AI chips alongside SK Hynix and Samsung.

Scale indicators3 records

Recent moves6 records

Expansion highlights5 records

Broad incumbents

• Google (TPU): Google's TPU is a vertically integrated AI accelerator optimized for transformer workloads and the closest analog to Etched's transformer-specialized ASIC. It is a major internal alternative for hyperscaler-class AI customers and sets the architectural benchmark for the category.
• Nvidia: Nvidia is the dominant incumbent in AI accelerators with its H100/H200/Blackwell GPUs and CUDA software stack. It is Etched's primary direct competitor in AI inference compute, and the standard against which Etched's 10-20x performance claims are benchmarked.

Direct peers

• Cerebras Systems: Cerebras builds wafer-scale AI accelerators (WSE) targeting both training and inference workloads for hyperscalers and enterprises. It competes head-to-head with Etched in selling dedicated AI silicon to large AI customers, though Cerebras's chip is general-purpose rather than transformer-specific.
• Groq: Groq develops LPU (Language Processing Unit) inference accelerators optimized for low-latency LLM serving. It directly targets the same transformer inference use case as Etched's Sohu and competes for similar hyperscaler and AI-company design wins.
• SambaNova Systems: SambaNova builds RDU (Reconfigurable Dataflow Unit) AI accelerators aimed at enterprise and government AI training and inference deployments. It is a comparable fabless AI chip startup targeting transformer-class workloads with similar enterprise go-to-market motion.
• Tenstorrent: Tenstorrent designs RISC-V based AI accelerators for training and inference, led by Jim Keller. Like Etched, it is a well-funded fabless AI chip startup competing for design wins in hyperscaler and enterprise AI infrastructure.
• Graphcore: Graphcore (now part of SoftBank) developed IPU (Intelligence Processing Unit) accelerators for AI workloads. It is a comparable specialized AI silicon company targeting transformer and ML inference, though it has shifted focus following acquisition.

Emerging players

• Lightmatter: Lightmatter builds photonic AI accelerators and interconnect products for data center AI workloads. It targets adjacent compute-infrastructure opportunities for AI inference and is a comparable well-funded AI hardware startup.
• Mythic AI: Mythic develops analog compute-in-memory AI inference chips for edge and data center deployments. It overlaps with Etched in the AI inference accelerator category, though its analog approach targets a different power/performance envelope.
• Hailo: Hailo builds edge AI inference processors for automotive and industrial use cases. While primarily edge-focused rather than data center, it represents an adjacent fabless AI accelerator company targeting transformer and CNN workloads with similar hardware specialization philosophy.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat4 records

Key risks6 records

Key highlights6 records

Customer concentration

• LinkedInlinkedin.com/company/etched-ai

Revenue estimate

Valuation estimate

Number of profiles

Profiles8 records

Funding overview

Funding rounds6 records

Investors33 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Etched.ai