Etched's Valuation Hits $21 Billion a Month After Jane Street Investment

The pace of this valuation increase is unusual even by AI industry standards. Etched was valued at $5 billion last December, then climbed to $10.3 billion in July of this year following a $300 million Series C round. This latest round doubled the valuation again, adding roughly $11 billion in the process.

Date	Valuation	Note
December 2025	$5 billion
July 2026	$10.3 billion	Series C, $300 million raised
August 2026	$21 billion	Led by Jane Street, $700 million raised

Background: what Etched actually sells

Etched doesn't sell individual chips — it sells complete systems, which the company calls "frontier inference clusters." Selling finished systems is similar to how rival NVIDIA describes its own systems as "AI factories." But where NVIDIA started out selling gaming graphics cards before expanding into picks-and-shovels for the broader AI chip market, Etched was built from the ground up as a startup focused solely on inference-specific hardware.

Chips split for two stages: prefill and decode

Inference refers to the computation that happens between when a user enters a prompt and when the system produces an answer. According to Etched co-founder and COO Robert Wachen, this process breaks down into two stages: prefill and decode. Prefill is the compute-intensive stage where the system parses the prompt and context, while decode is the memory-intensive stage where the system generates the actual answer, token by token, that the user sees.

Etched designed a new low-voltage chip specifically to handle the prefill stage. By lowering voltage, the company packed in more transistors while avoiding the heat problems common to advanced AI chips, which it says lets it process more tokens, faster. For the decode stage, the company built a new form of memory along with a dedicated interconnect to link it together. The company calls this "cluster-scale memory" — an architecture that connects multiple chips so they can share a memory pool at very high speed with minimal latency. Together, the company claims, this combination boosts speed while cutting costs.

Jane Street is already running racks

In a blog post announcing the investment, Jane Street said it "tested the chips and was pleased with the early results." The firm added that Etched's approach to inference delivers the precision its demanding computational workloads require, and said it is now running its own racks in its data centers.

The name no longer means "a model etched into a chip"

The name Etched came from an early-stage idea: etching a specific model directly into a chip. The concept was that each chip would be custom-designed for a single frontier model — and that was in fact the company's original plan. But that's no longer the case. Etched's systems have evolved to run any frontier model rather than being locked to one, and the company still contends with the confusion its name creates as a result.

Alongside Jane Street, this round included participation from Kleiner Perkins, Sequoia Capital, Andreessen Horowitz, Peter Thiel, Tiger Global, Bain Capital Ventures, NEA, Stripes, Primary, PositiveSum, Diffusion, Argo, and Blackstone.

Editor's take

A valuation doubling in a single month isn't driven by investor sentiment alone — it's the result of an actual use case. Jane Street is a quant trading firm where millisecond-level latency directly affects profit and loss. The fact that such a firm bought the chips and is now running racks in its own data centers is a far heavier form of validation than any benchmark number. Companies like Groq and Cerebras have made similar bets in the inference-chip market for years, but it's rare for a real production customer to publicly say it's satisfied.

There's a recurring pattern among inference-chip startups worth watching. They tend to start by drawing attention through extreme optimization — chips built for one specific model — and then grow their valuations as they broaden into general-purpose hardware. Etched has followed the same path. Its name still evokes "a chip etched for a model," but the actual product has shifted into a general-purpose system capable of running any frontier model. It's no coincidence that this shift coincided with the valuation jump — hardware that can run multiple models tends to have an easier time capturing a larger customer base than hardware locked to a single model.

One thing worth noting for domestic companies is the split design between prefill and decode. For teams concerned about inference costs, distinguishing between workloads bottlenecked by long, complex prompts (prefill) and those bottlenecked by long answer generation (decode) can make a real difference in infrastructure costs. Few domestic companies are likely to adopt Etched's hardware anytime soon, but the underlying idea of optimizing inference by splitting it into stages is worth keeping in mind even when simply choosing cloud GPU instances.

In the coming weeks, more real-world feedback from other major customers beyond Jane Street will likely emerge. Given how quickly this valuation has climbed, the key question for any future round will be whether it's backed by actual revenue and deployment scale.