Etched

A specialized AI hardware play that trades general-purpose flexibility for extreme Transformer inference performance—essential for massive scale, but useless for non-Transformer architectures.

Excellent for high-throughput Transformer inference at scale, weaker for experimental model structures or non-Transformer workloads.

Gpu Hardware Inference Llm Ai Acceleration Ai Development Machine Learning Models

Analysis based on product data, pricing structure, traffic signals, and public user sentiment.

Visit Website

Share this tool

Facebook X

LinkedIn Reddit

WhatsApp Telegram

Who Should Use Etched?

Typical users

Infrastructure engineers at LLM labs, hyperscalers, and enterprises running production-grade Transformer models at massive scale.

Maturity fit

advanced

Choose this if…

Your primary workload is inference for Transformer-based models like Llama or GPT
You need to maximize tokens-per-second while minimizing power and rack space
Latency is the primary bottleneck in your production AI application

Skip this if…

You need to run CNNs, RNNs, or other non-Transformer architectures
Your workflow relies heavily on the mature NVIDIA CUDA ecosystem for custom kernel development
You are in an R&D phase where model architectures change frequently

About Etched

Etched is a hardware company building Sohu, the first ASIC (Application-Specific Integrated Circuit) designed exclusively for Transformer models. By burning the Transformer architecture into silicon, they aim to bypass the overhead of general-purpose GPUs.

What it actually does

Etched provides specialized AI chips and a supporting software stack designed to run LLMs with significantly higher throughput than standard GPUs. Their Sohu chip is hard-coded to handle the specific mathematical operations of Transformers, allowing for massive parallelization of inference tasks.

What makes it different

Unlike NVIDIA GPUs which use programmable cores to handle any type of computation, Etched uses an ASIC approach. This 'bet on Transformers' means the hardware cannot be repurposed for other tasks, but it achieves performance gains that general-purpose silicon cannot match.

Transformer-only hardware acceleration FP8 precision support High-bandwidth memory (HBM) integration Massive throughput for long-context windows Multi-chip scaling for large model deployments Hardware-level beam search acceleration

Key Features

Sohu ASIC

Hard-coded Transformer architecture that eliminates the need for general-purpose compute overhead.

8x H100 Throughput

Claims to process over 500,000 tokens per second on Llama 70B models.

Open-source Software Stack

A compiler and kernel library designed to integrate with PyTorch and other frameworks.

Single-node Density

Designed to replace multiple GPU nodes with a single server, reducing data center footprint.

Low-latency Decoding

Hardware-level optimizations for faster token generation in real-time chat applications.

Scalable Interconnects

Built-in support for clustering chips to handle models with hundreds of billions of parameters.

Pricing

Popular

Enterprise Reservation

Custom one-time/deployment

Sohu ASIC hardware
Full software stack access
Direct engineering support
Custom deployment integration

Pricing checked 5 months ago

Pricing guidance

Best plan for most users: The Enterprise Reservation is currently the only path to access, intended for high-scale AI providers.

Free plan enough? No — this is high-end physical hardware with no free tier.

Upgrade when:

When GPU cloud costs for inference exceed millions per year
When latency requirements cannot be met by H100/B200 clusters
When data center power limits prevent further GPU scaling

Watch out for:

Hardware is physically incapable of running non-Transformer code
Requires specific server chassis and cooling configurations
Minimum order quantities likely apply

Premium infrastructure positioning aimed at the top 1% of AI compute spenders.

Pros & Cons

Strengths

Extreme Inference Throughput
By specializing the silicon, Sohu can handle token volumes that would typically require dozens of NVIDIA H100s, significantly lowering the cost per token.
Reduced TCO for LLM Providers
Lower power consumption and higher density per rack mean significantly lower operational costs for companies running models at scale.
Latency Leadership
The lack of general-purpose scheduling overhead allows for near-instantaneous inference, which is critical for voice and real-time agents.

Weaknesses

Zero Architectural Flexibility
If the industry shifts away from the Transformer architecture to something new (like SSMs or Mamba), this hardware becomes functionally obsolete.

Affects: Long-term infrastructure planners
Immature Software Ecosystem
Compared to NVIDIA's decade-old CUDA ecosystem, Etched's software stack is new and will likely have more friction for custom implementations.

Affects: Machine learning engineers
Limited Availability
As a hardware startup, initial supply is likely to be constrained and prioritized for major cloud partners and labs.

Affects: Mid-market enterprises

Real User Sentiment

The technical community is highly optimistic about the performance claims but cautious about the 'ASIC gamble' regarding architectural lock-in.

Users tend to like

The bold engineering decision to specialize on Transformers
Potential to break the NVIDIA monopoly on AI compute
Massive promised improvements in tokens-per-dollar

Users commonly complain about

Concerns about the lack of flexibility for future model types
Skepticism regarding real-world vs. theoretical performance
High barrier to entry for smaller developers

Recurring tradeoffs

Performance vs. Flexibility: You get world-class speed but lose the ability to run diverse model types.

Happiest users

Infrastructure leads at major LLM labs who need to slash inference costs for millions of users.

Often frustrated

Researchers experimenting with novel, non-Transformer architectures who find the hardware unusable for their work.

Use Cases

LLM Inference at Scale

Running models like Llama 3 for millions of concurrent users at lower cost.

Real-time AI Agents

Powering voice and chat bots that require sub-100ms response times.

Long-context Analysis

Processing massive documents or codebases that require high memory bandwidth.

Cost-optimized Cloud Hosting

Cloud providers looking to offer cheaper 'Transformer-only' compute instances.

Enterprise Private LLMs

Deploying high-performance models on-premise with a smaller hardware footprint.

Frequently Asked Questions

Can Etched Sohu run models like CNNs or Mamba?

No. Sohu is an ASIC hard-coded specifically for the Transformer architecture. It cannot run convolutional neural networks (CNNs) or newer architectures like Mamba or State Space Models (SSMs). It is designed to do one thing—Transformer inference—extremely well.

How does Etched compare to NVIDIA H100 in price?

While specific unit pricing is not public, Etched claims that a single Sohu server can replace several NVIDIA H100 nodes. This leads to a significantly lower Total Cost of Ownership (TCO) for inference-heavy workloads due to reduced power, cooling, and physical space requirements.

Is there a developer kit available for Etched?

Etched primarily targets large-scale enterprise and cloud partners. Individual developer kits are not widely available yet, though they offer a software stack for developers to begin preparing models for their architecture.

Does Etched support PyTorch?

Yes, Etched provides a specialized compiler and software stack that allows models written in PyTorch to be compiled and run on Sohu hardware, provided they follow the Transformer architecture.

What happens if the Transformer architecture is replaced?

This is the primary risk of using Etched. If a new architecture becomes the industry standard, Sohu hardware cannot be 'reprogrammed' to support it. Users would need to switch back to general-purpose GPUs or wait for a new ASIC design.

Is Etched hardware available for purchase now?

Etched is currently in the pre-order and reservation phase. They are working with early-access partners and hyperscalers to deploy the first generation of Sohu chips.

Why trust this page?

This evaluation combines product positioning, pricing analysis, traffic and market signals, and public user sentiment into a single decision-support page. Content is generated editorially — not copied from the vendor's website.

Funding & Company

Founded

2022

Stage

Series b

Total Raised

$625.4M

Latest Round

Series B (Jan 2026)

Notable Investors

Stripes Primary Venture Partners Positive Sum Peter Thiel Ribbit Capital

Etched has raised over $625 million in three rounds since its founding in 2022, culminating in a $500 million Series B in January 2026. This aggressive funding from notable investors like Stripes and Peter Thiel is aimed directly at challenging Nvidia's dominance by producing specialized ASIC hardware for AI.

Full funding report high confidence

Market Signals & Traffic

Estimated visits, global rank, geography, traffic sources, monthly visit trends, and organic search keywords (Similarweb)—on a dedicated page built for depth and search.

Estimated visits: 49,347
Global rank: #684,585
Snapshot: Apr 2026
Traffic trend: Falling