Etched’s Valuation More Than Doubles to $21 Billion After Jane Street Tests Its AI Inference Hardware

Etched announced $700 million in new funding on August 18, 2026, at a $21 billion valuation. Jane Street led the round after testing Etched’s hardware—and became the startup’s first customer after receiving its first inference-cluster rack. 156

The financing marks an unusually rapid rise in Etched’s private-market valuation: from $5 billion in December to $10.3 billion after a $300 million Series C on July 23, and then to $21 billion less than a month later. 3917

Why the round matters

The investment is a bet on specialized hardware for AI inference: the process of running an already-trained model after a user submits a prompt. Etched is not presenting its product as a conventional accelerator card. It sells complete rack-scale systems—what it calls frontier inference clusters—that combine chips, memory, networking, cooling, software and other infrastructure into an integrated serving platform. 524

That puts Etched in a part of the AI infrastructure market that is increasingly important as models move from training into large-scale, real-time use. Nvidia describes its broader rack-scale systems as “AI factories”; Etched’s systems pursue a similar full-system approach but are optimized specifically for inference workloads. 89

How Etched’s inference architecture works

AI inference has two distinct phases, and Etched designed a separate technology around each one.

1. Low Voltage Inference targets prefill

Prefill is the compute-intensive stage that reads a prompt and its context, then converts that information into the model’s working state. It is responsible for processing the input before the model begins generating an answer. 20

Etched’s approach, called Low Voltage Inference (LVI), runs the chip’s mathematical operations at less than half the voltage used by many conventional AI chips, according to the company. The intended benefit is lower heat generation, allowing Etched to fit more transistors into a given power and thermal envelope and increase compute density. 1827

The company says this design is intended for high-throughput workloads, including large sparse mixture-of-experts models. Those are company performance and compatibility claims; the supplied evidence does not independently verify every advertised benchmark or workload. 27

2. Cluster Scale Memory targets decode

Decode is the generation stage: the model produces an answer one token at a time. Unlike prefill, decode is heavily constrained by memory access because the system repeatedly consults the accumulated context, including the model’s key-value cache. 2025

Etched’s answer is Cluster Scale Memory (CSM), a shared-memory design that allows multiple chips across a rack to access a common high-speed, low-latency memory pool. The goal is to reduce the duplication and data movement that can occur when each accelerator relies primarily on its own local memory. In theory, that can improve throughput and lower the cost of serving large models. 72228

Together, LVI and CSM reflect Etched’s central product thesis: inference should be designed around the different bottlenecks in prompt processing and token generation, rather than treating both phases as the same workload.

Can Etched run any frontier model?

The available evidence points to a system intended to support more than one narrowly fixed model. Etched says its architecture can run large sparse mixture-of-experts models, and other reporting describes the company as positioning its clusters for frontier-model workloads. 1327

That is different from proving that the racks can run every frontier model without software changes, optimization or other adaptation. The supplied sources do not establish that broader, unqualified claim. The safer conclusion is that Etched is building general-purpose inference infrastructure around a specialized architecture, not a chip locked to one specific model.

Jane Street’s role as investor and customer

Jane Street’s involvement provides a customer-validation signal in addition to the financing itself. Etched said the trading firm tested its chip, took delivery of the first rack the following month, and began deploying the system in its data center. 56

Jane Street said:

“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. We’re excited to now have our own rack running in our data center.” 6

For Etched, the significance is that a demanding latency-sensitive customer moved from evaluation to deployment before leading the next funding round. It is still early validation rather than proof that the company has displaced incumbent AI hardware at scale.

Who participated in the $700 million round?

The announced participants included Jane Street, Kleiner Perkins, Sequoia, Andreessen Horowitz, Tiger Global, Bain Capital Ventures, Neo, Primary, Stripes, Positive Sum and Blackstone. 45

Etched’s own announcement also names Peter Thiel among its investors. The supplied material does not clearly establish whether that reference means he participated in this specific $700 million round or is being listed among the company’s cumulative backers. 1628

The evidence provided does not substantiate Diffusion or Argo as participants in this particular financing. They should not be included in the round’s investor list on the basis of the available sources.

The takeaway

Etched’s announcement combines three developments: a $700 million financing, a valuation that rose from $5 billion in December to $21 billion in August, and the first customer deployment of its rack-scale inference hardware. 159

The investment case rests on the idea that AI serving will need infrastructure designed specifically for inference economics. Etched’s LVI architecture targets the compute-heavy prefill phase, while CSM targets the memory-intensive decode phase. Jane Street’s deployment gives that thesis an early commercial test—but the company will still need to demonstrate reliable performance, broad model compatibility and production-scale delivery before the $21 billion valuation can be fully justified.