Etched Raises $700M at $21B Valuation After Jane Street Deploys Its Hardware
In this article
Etched on August 18 announced a $700 million raise at a $21 billion valuation, led by quant trading firm Jane Street — which didn't just write a check, but tested the hardware and deployed a rack in its own datacenter before committing. That sequence matters: this is not a speculative bet on a roadmap but a production deployment at one of the most computationally demanding institutions in finance. For engineers tracking alternatives to NVIDIA's inference stack, Etched's trajectory from a $5 billion valuation in December to $10.3 billion in July to $21 billion today — roughly doubling in a single month — marks a concrete inflection in the inference hardware market.
Two-Stage Inference Architecture
The valuation acceleration reflects two engineering decisions made at the silicon level, both targeting the two-stage structure of transformer inference.
Inference runs in two distinct phases. The prefill phase — parsing the prompt and building context — is compute-bound. The decode phase — generating each output token the user sees — is memory-bandwidth-bound. Etched built separate hardware primitives for each stage rather than forcing a single design to serve both.
For prefill, Etched designed a chip that operates at low voltage, enabling higher transistor density without the thermal dissipation problems that constrain competing high-end AI chips at standard voltages. Higher transistor density at lower thermal output translates directly to faster token processing during the compute-intensive context-understanding stage.
For decode, Etched engineered what it calls cluster-scale memory: a new memory type paired with an interconnect that allows multiple chips to share a single memory pool at low latency. Co-founder and COO Robert Wachen described the result as enabling chips to "connect together and use a shared memory pool at a very, very fast, low latency." The architectural claim is that pooling memory at cluster scale removes the per-chip memory ceiling that constrains throughput during decode — the stage where most wall-clock latency accumulates in production deployments.
Correcting the Early Narrative
Etched still carries reputational baggage from its founding thesis, which proposed etching a specific frontier model directly into silicon — an ASIC approach that would have made each chip model-specific and commercially fragile given how rapidly frontier models turn over. That is no longer the architecture. Etched's current systems can run any frontier model, which substantially changes the risk profile for customers making multi-year hardware procurement decisions. A cluster locked to one model vintage is a liability; a cluster optimised for the inference compute pattern shared across transformer-class models is a durable asset.
Valuation Trajectory and Investor Base
| Date | Event | Valuation | Amount Raised |
|---|---|---|---|
| December 2025 | Prior round | $5 billion | Not disclosed |
| July 2026 | Series C | $10.3 billion | $300 million |
| August 18, 2026 | Round led by Jane Street | $21 billion | $700 million |
Other investors include Kleiner Perkins, Sequoia Capital, Andreessen Horowitz, Blackstone, Tiger Global, Bain Capital Ventures, Neo, Stripes, Primary, Positive Sum, Diffusion, Argo, and Peter Thiel. Jane Street's public statement confirmed it now has "our own rack running in our datacenter," making this a go-live deployment rather than a pilot.
What the Jane Street Deployment Signals
Financial institutions running quantitative workloads represent some of the most demanding inference environments outside frontier AI labs. As explored in software extraction dynamics at the AI frontier, the firms best positioned in this cycle are those that control the inference layer rather than simply procuring it. The fact that Etched's first announced production customer is a high-frequency trading operation — rather than a hyperscaler or AI lab — suggests traction in verticals where latency and cost-per-token economics are existential rather than aspirational.
Etched's month-over-month valuation doubling is grounded in an actual shipment rather than benchmark projections. Whether the cluster-scale memory architecture holds up at broader deployment scale remains the open engineering question, but the commercial momentum is no longer speculative.