Etched Doubles to a $21B Valuation in One Month — and Its Lead Investor Is Running the Chips in Production
The inference-chip startup raised $700M led by Jane Street, which tested Etched's hardware, bought it, and then funded the company. Etched's valuation has quadrupled since December.
Etched raised $700 million at a $21 billion valuation on August 18 — roughly double what the inference-chip startup was worth a month ago, and more than four times its December 2025 valuation of $5 billion. The round was led by Jane Street, and that’s the detail that matters: the quant trading firm tested Etched’s hardware, bought it, installed a rack in its own datacenter, and then led the funding round. “We tested the chip and are pleased with the early results,” the firm said.
That’s not momentum investing. That’s a customer writing a check after production validation.
The fastest re-rating in AI hardware
The velocity here is remarkable even by 2026 standards. Etched closed a $300 million Series C at $10.3 billion on July 23, led by Sequoia. Twenty-six days later it closed $700 million at $21 billion. Total funding now stands at $1.9 billion, with a cap table that reads like a venture index: Kleiner Perkins, Sequoia, Andreessen Horowitz, Peter Thiel, Tiger Global, Bain Capital Ventures, and Blackstone, among others.
In June, the company said it had successfully manufactured its chips with TSMC and booked $1 billion in orders. Two months later it completed its first customer delivery. Hardware startups usually die in exactly this gap — between tape-out and revenue — and Etched appears to have crossed it.
Why an inference-only chip is winning deals
Etched builds specialized silicon for transformer inference rather than general-purpose GPUs. Co-founder Robert Wachen says the speed comes from optimizing both halves of the workload: the prefill phase, where the model ingests the prompt, and decode, where it generates tokens. The company designed a dedicated prefill chip that runs at low voltage to process tokens faster, plus cluster-scale memory that lets multiple chips share memory at low latency.
For latency-obsessed customers — and nobody is more latency-obsessed than a high-frequency trading firm — that architecture is the pitch. Jane Street validating it in production is the strongest possible reference customer.
The Nvidia subtext
Etched is openly poaching Nvidia engineers and positioning itself against the company that owns AI compute. It won’t threaten Nvidia’s training business, and it doesn’t need to: inference is where the market is compounding, as usage grows faster than model training. And the moat is visibly narrower on this side of the workload — Cerebras already powers OpenAI’s Ultrafast API tier, and inference-first silicon keeps landing real customers.
A $21 billion valuation on one delivered customer is still aggressive. But the difference between Etched and the paper unicorns is now concrete: someone’s money runs on these chips.