The archive · Hardware & Devices · Technical decision · 2022–2026
Etched bets everything on a transformer-only chip; $1B orders before first racks ship
Etched's Sohu ASIC runs only transformer inference, claiming 20x an H100; after a near-death 2023, it books $1B in orders and a $10.3B valuation by July 2026.
Etched
What the business is
Etched designs Sohu, an application-specific integrated circuit that only runs transformer-model inference, and sells frontier inference clusters — racks combining its chips with custom hardware and software — to AI companies.
Starting capital：Disclosed rounds: $5.4M seed at a $34M valuation (Mar 2023); $120M Series A led by Primary Venture Partners (Jun 2024); $500M at a $5B post-money valuation led by Stripes (Dec 2025); $300M Series C led by Sequoia at a $10.3B valuation (Jul 2026).
How it started
Gavin Uberti, Robert Wachen and Chris Zhu dropped out of Harvard and founded Etched in 2022 after writing a 30-page memo arguing that AI would eventually need specialized silicon, not general-purpose GPUs. In 2023 every major investor they pitched passed, and the company ran month-to-month, close to running out of cash, before closing a $5.4M seed in March 2023 at a $34M valuation.
What happened
In June 2024 Etched launched Sohu publicly and raised a $120M Series A led by Primary Venture Partners with Peter Thiel and Amjad Masad participating, claiming a Sohu server runs Llama 70B at 500,000 tokens per second and replaces 160 H100s. TSMC fabricated the chip on its 4nm process, and the A0 tape-out succeeded in early 2026. In December 2025 Etched quietly closed a $500M round at a $5B post-money valuation led by Stripes, with Jane Street, Hudson River Trading, Two Sigma and Ribbit Capital participating.
How it ended up
On 2026-06-30 Etched came out of stealth: TSMC had manufactured the chip, the company had booked over $1B in signed contracts for frontier inference clusters before shipping production racks, and it disclosed about $800M raised. On 2026-07-23 Sequoia led a $300M Series C at a $10.3B valuation — which Forkast reported as the highest valuation ever for a Sequoia-led Series C — with SK Hynix investing. First production racks were scheduled for summer 2026; the chip had still not shipped at scale or been independently benchmarked.
Background
Etched's bet is the purest form of hardware specialization: build a chip that can only run transformer models, strip away every other capability, and win on speed and cost. Founders Gavin Uberti, Robert Wachen and Chris Zhu wrote a 30-page memo in 2022 arguing that AI would need dedicated silicon, watched every major investor pass in 2023, and survived month-to-month before a $5.4M seed at a $34M valuation.
The 2024 AI boom changed everything. In June 2024 Etched launched Sohu — an ASIC fabricated by TSMC on 4nm that it claimed runs Llama 70B at 500,000 tokens per second, with one 8-chip server replacing 160 H100s — and raised a $120M Series A led by Primary Venture Partners. 'We're making the biggest bet in AI,' CEO Uberti said: 'If transformers go away, we'll die.'
By December 2025 investors were betting the bet was right: Stripes led a $500M round at a $5B post-money valuation, joined by Jane Street, Hudson River Trading, Two Sigma and Ribbit Capital. In early 2026 the A0 tape-out succeeded at TSMC, and on June 30 Etched came out of stealth with over $1B in signed customer contracts for full frontier inference clusters — chips, racks and software — before shipping production hardware.
On July 23, 2026 Sequoia led a $300M Series C at a $10.3B valuation, with SK Hynix joining, bringing disclosed funding to roughly $1.1B. The company employs over 400 people and runs 2 MW and 10 MW data centers in San Jose and Milpitas, but as of September 2026 no Sohu system has shipped at scale or passed an independent benchmark — the $10.3B still rests on the claim that transformer attention stays the dominant architecture.
What has to be true
- Fixed-function silicon only wins if the workload is stable, so Etched bet on transformer attention outlasting every rival architecture.
- The 2023 funding failure shows the bet was not obvious: specialized inference silicon became fundable only after the AI boom made inference the bottleneck.
- Vertical integration (chips plus custom racks, software and data centers) lets Etched sell a system, not just a component, to customers without Nvidia-scale engineering teams.
- $1B of pre-orders before first shipments converts architectural conviction into a demand signal investors can price.
What can be applied
Specialization only pays off when the market arrives: Etched's 2023 memo was refused by every major investor, and the AI boom funded a chip that still has not shipped.
Aftermath
As of 2026-09-02, Etched is a roughly $10.3B company (per its July 2026 Sequoia-led Series C) with about $1.1B raised, over 400 employees, and more than $1B in signed contracts for Sohu-based frontier inference clusters. TSMC manufactured the A0 chip on its N4P process in early 2026; first production racks were scheduled for summer 2026, and Etched operates a 2 MW data center in San Jose plus a 10 MW facility in Milpitas. The chip has not yet shipped at scale or been independently benchmarked, and the entire thesis remains exposed to a shift away from transformer architectures.
Sources
- Transformer model chipmaker Etched.ai raises $120M to challenge Nvidia's market dominance
- Nvidia competitor Etched hits $5B valuation, $1B in sales for AI chip
- Etched's $10.3 billion valuation is a bet on silicon specialization, not chip performance
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card