Etched raises $700M at $21B valuation - doubling in a month

👁 115 reads

AI chip startup Etched announced on Tuesday, 18 August, that it has raised $700 million at a $21 billion valuation in a round led by Jane Street, the quantitative trading firm, which tested and bought the startup’s hardware before committing capital. The speed of the step-up is extraordinary even by the standards of the current AI funding market: Etched was valued at $5 billion in December, raised a $300 million Series C at a $10.3 billion valuation in July, and has now seen investors add nearly $11 billion to that figure in the space of a single month.

What Etched actually builds

Etched sells AI hardware as complete systems it calls “frontier inference clusters” - a framing that mirrors Nvidia’s terminology of “AI factories” for its own full-system products. The distinction from traditional chip vendors is that Etched’s offer covers the entire stack, from silicon to interconnect, rather than discrete components a customer must assemble.

The technical core of the product centres on a deliberate split of the inference workload into its two constituent stages. Co-founder and COO Robert Wachen explained the architecture to TechCrunch:

“Inference is built in two stages - prefill and decode.”

The prefill phase is mathematically and computationally intensive: the system must parse a user’s prompt and process all relevant context before generating any output. The decode phase is memory-intensive: it produces the output tokens - the actual answer a user reads. Most AI hardware treats these two stages with a single generalised design. Etched has built a separate chip for each.

For the prefill stage, Etched designed a chip that operates at low voltage. Lower voltage means less heat per transistor, which in turn allows the company to pack more transistors into the same physical space without hitting the thermal limits that constrain competing high-end AI chips. More transistors in a denser arrangement means more tokens processed per second during the prompt-understanding phase.

For the decode stage, Etched built what it calls cluster-scale memory - a new type of memory and interconnect that lets many chips share a single large memory pool.

“It allows many chips to connect together and use a shared memory pool at a very, very fast, low latency.”

The practical claim from the company is a combination of higher throughput speeds and lower operating costs relative to existing inference infrastructure.

Jane Street’s endorsement carries weight

The identity of the lead investor matters here as much as the dollar figure. Jane Street is not a venture fund placing bets on promising roadmaps - it is one of the world’s most technically rigorous trading firms, with quantitative workloads that place extreme demands on low-latency computation. Its decision to lead the round came after the firm ran Etched’s hardware internally and was satisfied with the results.

In the blog post accompanying the announcement, Jane Street stated:

“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 datacenter.”

That last sentence - confirming a rack is already live in Jane Street’s data centre - is significant. It moves Etched from the category of well-funded hardware promise to deployed production infrastructure at one of the most demanding computational environments in finance. For AI chip companies, where the gap between announced silicon and real customer workloads has historically been wide, a live production deployment at this scale is a meaningful proof point.

Other investors in the company include Kleiner Perkins, Sequoia Capital, Andreessen Horowitz, Peter Thiel, Tiger Global, Bain Capital Ventures, Neo, Stripes, Primary, Positive Sum, Diffusion, Argo, and Blackstone.

A perception problem being shed

Etched still carries baggage from its original positioning. In its early days the company described its approach as etching a specific model into its chips - implying each chip was permanently optimised for one particular frontier model and would become obsolete as that model was superseded. That is no longer how the product works. Etched’s current systems can run any frontier model, and Wachen has been explicit in correcting the record with investors and potential customers.

The shift matters commercially. A chip locked to a single model is a hard sell to an enterprise customer planning infrastructure investments over a multi-year horizon. A chip that runs any frontier model at higher speed and lower cost than the incumbent - and that has already been validated by a buyer as demanding as Jane Street - is a different proposition entirely. At $21 billion, the market is pricing in the second story, not the first.

// read moreGeneral Intuition in Talks to Raise at $6B ValuationMagazineGroq Raises $350M at $3.5B Valuation for Neocloud BusinessMagazineOpenAI-backed Thrive Holdings raises $2B at $12B valuationMagazine

Kinetyra