Etched Raises $700 Million at $21 Billion Valuation, Completes First Customer Delivery
Etched announced on August 18, 2026 that it raised $700 million at a $21 billion valuation in a round led by Jane Street, capping a climb from a $5 billion valuation in December 2025 through a $10.3 billion Series C in July 2026. The same day, the company completed its first customer delivery to Jane Street, which tested and purchased Etched's AI hardware and said it was pleased with early results. Jane Street will run an Etched rack in its datacenter. TechCrunch GlobeNewswire
The round's lead investor doubling as a paying customer is a notable signal. Jane Street, primarily a quantitative trading firm, is not a typical hyperscaler or cloud provider. Its willingness to deploy Etched hardware in production rather than merely participate in a financing gives the valuation a concrete anchor that pure financial rounds lack.
Etched's investor base extends well beyond Jane Street. The cap table includes Kleiner Perkins, Sequoia Capital, Andreessen Horowitz, Peter Thiel, Tiger Global, Bain Capital Ventures, Neo, Stripes, Primary, Positive Sum, Diffusion, Argo, and Blackstone. TechCrunch The July Series C at $10.3 billion was led by Sequoia Capital alongside Andreessen Horowitz, Jane Street, Diffusion, Argo, and SK Hynix. Etched Etched also raised $800 million across four unannounced financings, including a strategic investment from VentureTech Alliance. Etched
The company has roughly 400 employees, and around 15% previously worked at Nvidia. WSJ Etched's co-founder and COO is Robert Wachen.
On the technical side, Etched designed a low-voltage prefill chip that packs in more transistors without the typical heat problems, allowing faster processing of more tokens — the individual units of text a model generates. The company also created a new type of memory and interconnect it calls cluster-scale memory, which lets many chips connect and share a memory pool at low latency for the decode process. Etched delivers its AI technology as full systems it calls "frontier inference clusters." Critically, these systems can run any frontier model and are no longer custom-built to run a single model, a meaningful evolution from the company's earlier thesis of single-model chips. TechCrunch
The prefill-decode split is worth pausing on. AI inference — the process of a trained model actually answering a query — has two phases: prefill, where the model processes the prompt, and decode, where it generates output tokens. Etched's approach addresses both with distinct hardware strategies: a low-voltage prefill chip for throughput and a cluster-scale memory pool for low-latency decode. This is architecturally different from running both phases on the same GPU, as Nvidia's H100 and B200 do. Whether the performance gains hold at scale across diverse model architectures remains the open question, but the first production deployment at Jane Street provides initial validation.
The shift away from single-model chips is strategically significant. A chip hardwired for one model architecture risks obsolescence the moment training paradigms shift — like building a factory that can only manufacture one product. By building systems that run any frontier model, Etched broadens its addressable market and hedges against the rapid architectural churn in the LLM space. The trade-off is that a more flexible system generally cannot match the raw efficiency of a fixed-function chip for a specific workload. Etched appears to be betting that flexibility and system-level optimization — memory pooling, interconnect, low-voltage design — can close that gap.
The valuation trajectory itself warrants attention. Going from $5 billion to $21 billion in roughly eight months implies that investors are pricing in either substantial customer traction, breakthrough hardware performance in benchmarks, or a combination of both. The Jane Street delivery is the first public evidence of a paying customer in production. For a hardware startup, first delivery is a genuine milestone: it means silicon that was designed, manufactured, packaged, integrated into systems, and deployed under real workloads. Many AI chip startups stall at the prototype or sample stage.
Around 15% of Etched's headcount coming from Nvidia is a double-edged indicator. It brings deep expertise in GPU-class system design, thermal management, and datacenter integration. It also means Etched is competing for the same talent pool as its primary adversary, and the institutional knowledge those engineers carry is both an asset and a potential constraint — muscle memory from one architecture does not always transfer cleanly to a fundamentally different approach.
The broader context here is an inference market that is expanding rapidly as model deployment shifts from training-heavy spending toward serving. If Etched's frontier inference clusters deliver on their performance claims across multiple model families, the company has a credible path to capturing share in a market currently dominated by Nvidia. The $21 billion valuation prices that possibility aggressively. The next data point that matters is not another financing round but a second customer delivery and public performance data under production workloads.
