AI ALPHA

Etched Hits $21 Billion, Ships First Chip to Jane Street

Custom AI inference silicon on a circuit board, representing Etched's chips now deployed at trading firm Jane Street
Etched shipped its first production rack to Jane Street in July and the quant firm then led its $700 million round. Source: Wall Street Journal
TLDR

Jane Street led the round and took the first rack off the line

Etched, the inference-chip startup that spent two years as the most doubted name in AI silicon, said it has raised $700 million at a $21 billion valuation. The lead investor is Jane Street, joined by Kleiner Perkins, Sequoia, Andreessen Horowitz, Tiger Global, Bain Capital Ventures, and Blackstone, among others. The number that matters is not the raise. It is the speed of the repricing behind it.

Bar chart showing Etched's valuation rising from 5 billion dollars in December 2025 to 10.3 billion dollars in July 2026 to 21 billion dollars in August 2026
Etched's valuation quadrupled in eight months and doubled in the last one. Source: Santage analysis of Etched funding announcement, August 18, 2026.

Eight months ago Etched was worth $5 billion. In July it closed a $300 million round at $10.3 billion. This new round doubles that again. Repricings that steep usually follow a product that is selling, not a promise, and this one did. In the same announcement, Etched said it had shipped its first rack to Jane Street in July and that the system is already running production workloads inside the firm's data center.

The round in brief
New raise$700 million
Valuation$21 billion, up from $10.3 billion in July
Valuation in December 2025$5 billion
Customer contracts signedMore than $1 billion
First customer deliveryJane Street, July 2026
Time from seed to working siliconUnder three years
Source: Etched funding announcement, August 18, 2026.

Why a trading firm is buying inference silicon before the hyperscalers

Jane Street is not a lab and not a cloud. It is a proprietary trading firm whose edge is measured in the precision and latency of the models it runs against markets. That makes it a revealing first customer. When a company that keeps its methods secret and answers to no one but its own profit and loss chooses an unproven chip over Nvidia's shipping product, it is making a statement about the economics, not the branding.

Etched built its name on Sohu, a chip specialized for the transformer, the architecture underneath almost every modern AI model. Rather than a general accelerator, it now sells complete inference clusters built from two custom pieces. A Low Voltage Inference design targets the compute-heavy prefill stage, where a model reads a prompt, while a Cluster Scale Memory system pools memory across an entire rack for the memory-bound decode stage, where a model writes each token. Splitting inference into those two problems and optimizing each separately is the wager, and it is meant to buy speed and cost a general-purpose GPU cannot.

Etched componentStage it targetsWhy it matters
Low Voltage InferencePrefill, reading the promptHigher compute density at lower power
Cluster Scale MemoryDecode, writing each tokenShared memory across a full rack, lower latency
Etched splits inference into two hardware problems and builds a custom part for each. Source: Etched.

The firm framed the delivery as a test that passed, not a favor to a portfolio company.

We tested the chip and are pleased with the early results. Etched's unique approach to inference delivers the precision we will need to support our most demanding workloads.
Jane Street, in Etched's funding announcement

What changes when the lead investor is also the first customer

For most of Etched's life the criticism was structural. Founded by two college dropouts, Gavin Uberti and Chris Zhu, it had raised heavily on silicon tied to a single architecture, and skeptics argued that betting a chip on the transformer was reckless in a field that reinvents itself yearly. The company has now inverted that story. It says its clusters run any frontier model, and it has a paying deployment to point to rather than a benchmark slide.

The financing structure is the real signal. Jane Street did not write a check and wait for a roadmap. It bought the hardware, ran it on real workloads, and then led the round. Customer and lead investor became the same entity, which compresses the usual gap between someone funding a chip and someone depending on it to zero. Chief executive Gavin Uberti said the company has felt the urgency to get its hardware into customers' hands and run real workloads since day one, and the raise rewards exactly that sequencing.

For the market, this lands at a specific moment. Inference, not training, is now the cost that scales with every user and every agent, and it is the line item hyperscalers most want to pull away from a single supplier. A credible second source for inference silicon is worth a premium to anyone whose margins bend to the price of a token. That is why a chip startup can quadruple in eight months while still shipping its first units.

The safest way to fund hardware turned out to be selling it first.

The detail most coverage skipped

Etched reached first-pass working silicon in under three years from seed. First-pass success, meaning a chip that works without a costly re-spin, is rare even at established firms and rarer still at a startup, and it is the quiet technical fact underneath the valuation. The $21 billion is not paying for a deck. It is paying for a working part that a demanding customer already trusts in production.

The lesson of this round is not that a chip company got expensive. It is that the surest way to raise money for hardware turned out to be selling it first. Etched did not raise to go prove its silicon. It proved its silicon, delivered it to the buyer, and let the buyer lead the round.

Santage is committed to independent, transparent journalism. This article is produced in accordance with Santage's Editorial Standards and aims to provide accurate and timely information. Readers are encouraged to verify information independently.