# Etched

> Source: https://aiwiki.ai/wiki/etched
> Updated: 2026-06-24
> Categories: AI Companies, AI Hardware
> From AI Wiki (https://aiwiki.ai), a free encyclopedia of artificial intelligence. Quote with attribution.

**Etched** is an American semiconductor startup building [Sohu](/wiki/sohu), the first [application-specific integrated circuit](/wiki/asic) (ASIC) designed to run only [transformer](/wiki/transformer) models, the architecture behind most large language models. Founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen, Etched hardwires the transformer directly into silicon, trading the flexibility of a general-purpose [GPU](/wiki/gpu) for raw [AI inference](/wiki/ai_inference) speed: the company claims a single eight-chip Sohu server runs Meta's Llama 70B at over 500,000 tokens per second, equivalent to replacing roughly 160 [Nvidia](/wiki/nvidia) H100 GPUs.[1][2] The bet is explicit. "In 2022, we made a bet that transformers would take over the world," co-founder and chief executive Gavin Uberti said when the chip was unveiled in June 2024.[1] By early 2026 the company had raised roughly $1 billion across several rounds and reached a reported valuation of about $5 billion, while Sohu remained pre-production and not yet commercially available.[2][3][4]

## Overview

Etched designs a single product, the Sohu chip, an inference [ai_chips](/wiki/ai_chips) ASIC specialized for the transformer architecture that underlies most large language models, including those in the [GPT](/wiki/gpt) and [Llama](/wiki/llama) families. The company's central argument is that a chip dedicated to one architecture can dedicate far more of its die area to transformer-specific compute than a general-purpose GPU, and that the resulting throughput and efficiency advantages justify the loss of flexibility. Etched describes Sohu in marketing as the basis for "the world's first transformer ASIC."[1][5]

The strategy is deliberately narrow. Because Sohu cannot run other model types, Etched's commercial prospects are tied to the continued dominance of the transformer. The founders have framed this publicly as a calculated risk, arguing that the AI industry has converged on transformers to a degree that makes a specialized chip economically attractive. "We've hit a point in the evolution of AI where specialized chips that can perform better than general-purpose GPUs are inevitable, and the technical decision-makers of the world know this," Uberti said.[1][2]

## Who founded Etched and when?

Etched was founded in 2022. Gavin Uberti and Chris Zhu both left Harvard University to start the company; Uberti is a [Thiel Fellowship](/wiki/thiel_fellowship) alumnus. The third co-founder, Robert Wachen, is also a Harvard dropout. The early team recruited semiconductor veterans, most notably Mark Ross, a former chief technology officer of [Cypress Semiconductor](/wiki/cypress_semiconductor), who serves as Etched's CTO, along with engineers drawn from companies such as [Broadcom](/wiki/broadcom).[1][4][6]

The company is based in Cupertino, California. At the time of its Series A in mid-2024 it employed roughly 35 people.[1]

## How does the Sohu chip work?

Sohu is an inference accelerator manufactured on [TSMC](/wiki/tsmc)'s 4-nanometer process. Each chip carries 144 GB of HBM3E high-bandwidth memory, roughly 1.8 times the memory bandwidth of an Nvidia H100 SXM5, which Etched cites as a major source of its performance advantage on memory-bound inference workloads.[2][4][7]

Etched markets Sohu around aggressive, company-supplied benchmarks. Its headline claim is that a single server containing eight Sohu chips can deliver about 500,000 tokens per second running Meta's Llama 70B model, which the company says is equivalent to replacing roughly 160 Nvidia H100 GPUs. Etched further asserts that Sohu is "an order of magnitude faster and cheaper than even Nvidia's next generation of Blackwell GB200 GPUs when running text, image and video transformers," a margin the company quantifies as about 10 times faster than Blackwell and 20 times faster than the H100 (Hopper) generation.[1][2][7] These figures are forward-looking marketing claims rather than independently verified results; as of mid-2026 they had not been confirmed by third-party testing.[4][7]

| Detail | Etched claim |
| --- | --- |
| Architecture | Transformer-only ASIC (inference) |
| Process node | TSMC 4 nm |
| Memory | 144 GB HBM3E per chip (~1.8x H100 SXM5 bandwidth) |
| Headline throughput | ~500,000 tokens/sec on Llama 70B per 8-chip server |
| GPU comparison | One Sohu server stated to replace ~160 Nvidia H100s |
| Claimed speedup | ~10x Nvidia Blackwell GB200; ~20x H100 (Hopper) |

The same specialization that drives Sohu's claimed advantages also constrains it. Because the chip hardwires the transformer, it is reported to be unable to run [mixture-of-experts](/wiki/mixture_of_experts) routing, vision encoders, [diffusion models](/wiki/diffusion_model), [state space models](/wiki/state_space_model), CNNs, or LSTMs in the way a GPU can, which would exclude some architectures used in newer production systems.[7]

## Funding

Etched has raised capital in three principal stages. It closed a seed round of $5.4 million, reported by EE Times in 2023 at a valuation of about $34 million.[6] In June 2024 it announced a $120 million Series A co-led by Primary Venture Partners and Positive Sum Ventures, bringing total funding raised to about $125.4 million at the time. Named backers in that round included [Peter Thiel](/wiki/peter_thiel), GitHub chief executive Thomas Dohmke, Cruise and Bot Company co-founder Kyle Vogt, and Quora co-founder Charlie Cheever.[1]

In January 2026 multiple outlets reported that Etched had raised approximately $500 million in a round led by the investment firm Stripes, with participation from Peter Thiel, Positive Sum, and Ribbit Capital. The round valued the company at about $5 billion and brought its total funding close to $1 billion.[3][4]

| Round | Amount | Date | Lead(s) |
| --- | --- | --- | --- |
| Seed | $5.4M | 2023 | (~$34M valuation) |
| Series A | $120M | June 2024 | Primary Venture Partners, Positive Sum Ventures |
| Later round | ~$500M | January 2026 | Stripes (~$5B valuation) |

## Why bet on transformers, and what are the risks?

Etched's business is an explicit wager that the transformer will remain the standard AI architecture for years to come. A transformer-only ASIC offers no fallback if the field shifts toward a fundamentally different design; unlike a GPU, the silicon cannot be repurposed for another architecture. The founders have argued that the convergence of frontier models onto the transformer makes this risk acceptable, and that the payoff is capabilities GPUs cannot reach: Uberti has said that "video generation, audio-to-audio modalities, robotics, and other future AI use cases will only be possible with a faster chip like Sohu."[1] Commentators have nonetheless noted that betting an entire company on architectural stability is unusual in a fast-moving field.[2]

A second risk is execution and timing. Designing, taping out, and bringing a leading-edge accelerator to volume production is capital-intensive and slow, and Etched competes against Nvidia's entrenched ecosystem as well as other inference-focused challengers such as [Groq](/wiki/groq), [Cerebras](/wiki/cerebras), and [SambaNova](/wiki/sambanova). Because much of what Etched has published consists of projected benchmarks rather than shipping silicon, the company's claims remained unproven in independent testing as of mid-2026.[4][7]

## Is Sohu available yet?

As of 2026, Sohu had not become widely commercially available. Etched partnered with TSMC's Emerging Businesses Group to manufacture the chip, and the company has demonstrated Sohu to investors and in controlled benchmarks. Reporting from April 2026 indicated that the chip was not publicly available for purchase or rental, placing it in a pre-production or early customer-engagement stage rather than mass production.[4][7]

At its 2024 Series A, Etched said unnamed customers had reserved tens of millions of dollars' worth of hardware and that it planned to offer early access through a developer cloud preview.[1] The $500 million raised in January 2026 was framed by the company as fuel for the continued push to bring Sohu to market against Nvidia.[2][3][4]

## References

1. Wiggers, Kyle. "Etched is building an AI chip that only runs transformer models." TechCrunch, June 25, 2024. https://techcrunch.com/2024/06/25/etched-is-building-an-ai-chip-that-only-runs-transformer-models/
2. Bass, Dina. "AI Chip Startup Etched Raises $500 Million to Take on Nvidia." Bloomberg, January 13, 2026. https://www.bloomberg.com/news/articles/2026-01-13/ai-chip-startup-etched-raises-500-million-to-take-on-nvidia
3. "Etched Raises $500 Million to Challenge Nvidia." Yahoo Finance / Bloomberg, January 2026. https://finance.yahoo.com/news/etched-raises-500-million-challenge-115812959.html
4. "Harvard dropouts' Etched raises $500M at $5B valuation to challenge Nvidia." Tech Funding News, January 2026. https://techfundingnews.com/nvidia-rival-ai-chip-maker-etched-founded-by-harvard-dropouts-lands-500m-at-5b-valuation/
5. "AI chip startup Etched.ai raises $120 million to take on Nvidia." Data Center Dynamics, June 2024. https://www.datacenterdynamics.com/en/news/ai-chip-startup-raises-120-million-to-take-on-nvidia/
6. "Harvard Dropouts Raise $5 Million for LLM Accelerator." EE Times, 2023. https://www.eetimes.com/harvard-dropouts-raise-5-million-for-llm-accelerator/
7. "Etched AI Sohu vs NVIDIA: Transformer ASIC vs General-Purpose GPU for LLM Inference (2026)." Spheron Blog, 2026. https://www.spheron.network/blog/etched-ai-sohu-vs-nvidia-transformer-asic-inference/