Etched is a three-year-old startup making an unusual bet: instead of a general-purpose AI chip like Nvidia’s GPUs, it designs hardware that can run only one kind of model — the transformer, the architecture behind nearly every major chatbot today. That narrow focus, the company argues, lets its chip run transformer models dramatically faster and cheaper than a general-purpose GPU, and it has reportedly drawn investor interest at a valuation of up to $20 billion.
What Etched actually builds
Founded in 2022 by three Harvard dropouts — Gavin Uberti, Chris Zhu, and Robert Wachen — Etched spent two years largely in stealth before unveiling its chip, called Sohu, in June 2024. Sohu is a specialized chip known as an ASIC, or application-specific integrated circuit. Rather than the programmable circuits a general-purpose AI chip uses to handle many different kinds of math, Sohu’s circuits are wired at the hardware level to do only the calculations a transformer needs — mainly attention and matrix multiplication — and nothing else.
That tradeoff is deliberate. A GPU has to stay flexible enough to train new model architectures, render graphics, and run scientific simulations. Etched’s bet is that this flexibility is no longer worth paying for at inference time — the stage where an already-trained model is put to work answering prompts — because nearly every large AI system in production today, from chatbots to coding assistants, is built on the same transformer architecture. By stripping out that generality, Etched says a server of eight Sohu chips can push over 500,000 tokens per second on a 70-billion-parameter Llama model, versus roughly 23,000 tokens per second on eight of Nvidia’s H100 GPUs. The chip is built on a TSMC 4-nanometer process with 144GB of high-bandwidth memory per chip.
Those figures are Etched’s own benchmarks, not independently audited ones, and as of mid-2026 the company still has no large-scale, third-party-verified deployment to point to — its website says its first hardware racks are only now shipping to customers. The design also carries a permanent, hardware-level limit: because the attention circuits are fixed in silicon, Sohu cannot run models built on a different architecture — convolutional networks, diffusion image models, or state-space models like Mamba — even if something eventually replaces the transformer as AI’s dominant design.
Why investors are betting billions
Etched raised a $120 million Series A in June 2024, led by Primary Venture Partners, on the strength of the Sohu announcement alone, before a single chip had shipped. By January 2026 it had raised a $500 million round led by Stripes and Peter Thiel, pushing its valuation to roughly $5 billion and total funding past $600 million, with backers citing around $1 billion in customer demand. As of July 2026, Etched is reportedly in talks for two back-to-back funding rounds: one valuing the company at around $10 billion, led by Sequoia Capital, and a separate one at roughly $20 billion, led by Jane Street — a fourfold jump in a matter of months, though neither deal had closed at the time of writing and terms could still change.
The rush of capital reflects a broader shift in AI spending. As companies move from training frontier models to running them at scale for billions of daily queries, inference — not training — is absorbing a growing share of AI compute budgets. That shift has opened room for challengers to Nvidia’s dominance, from hyperscalers’ own custom silicon to startups like Etched betting that specialization, not flexibility, wins the next round of AI infrastructure spending.
Not a product you can buy directly
Etched doesn’t sell individual chips to developers; it sells server racks and data-center access to AI labs and cloud providers, positioning Sohu as an alternative to renting Nvidia GPU capacity. There’s no consumer signup or public price list — the company directs prospective customers to request access directly rather than publishing pricing.
In the news
The valuation talks that put Etched back in the spotlight this month are detailed in our news coverage. Read the full story →
Sources: Wall Street Journal reporting via Investing.com and PYMNTS (July 2026); Jon Peddie Research; Etched company statements.