Etched raised $700 million at a $21 billion valuation on Tuesday, led by Jane Street after the quant fund tested the startup's AI inference hardware, bought a rack, and installed it in its own datacenter. The round doubles a valuation that was $10.3 billion just a month earlier and quadruples the $5 billion mark Etched carried in December.
The step-up is aggressive even against 2026's inflated AI-hardware comps. Etched closed a $300 million Series C at $10.3 billion in July. The new $700 million check adds roughly $11 billion of paper value in four weeks and signals that a customer-turned-investor validation — Jane Street running production workloads on the chip — is worth more to the market than another round of pitch decks.
Etched sells full systems it calls 'frontier inference clusters,' the same product category Nvidia markets as AI factories. That framing matters: Etched is not pitching a component that slots into someone else's rack. It is pitching a rack that replaces Nvidia's, aimed at the inference side of the workload where cost per token has become the dominant economic metric for AI deployments.
Key facts
- 01Etched raised $700M at a $21B valuation, led by Jane Street after the quant fund tested and bought its hardware.
- 02The valuation is up roughly $11B in a month, from $10.3B in July and $5B in December.
- 03Etched sells full 'frontier inference clusters,' competing directly with Nvidia's AI factory systems.
- 04The company built a low-voltage prefill chip and a new cluster-scale memory interconnect for the decode phase.
- 05Backers include Kleiner Perkins, Sequoia, Andreessen Horowitz, Peter Thiel, Tiger Global, and Blackstone.
Jane Street, announcing its lead investment, said it has already put the hardware to work.
“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.”— Jane Street, Lead investor
Co-founder and COO Robert Wachen said the enthusiasm comes from two custom components Etched designed for the two distinct phases of inference. Prefill, where the model ingests and understands a prompt, is compute-heavy. Decode, where the model generates the response tokens the user sees, is memory-heavy. Most general-purpose AI chips compromise on one to serve the other.
Etched's prefill chip runs at low voltage, which lets the company pack more transistors onto the die without hitting the thermal walls that constrain high-end GPUs. More transistors at a manageable temperature translates to more tokens processed per second on the ingestion side.
For the decode phase, Etched built a new memory type and an interconnect it calls cluster-scale memory, which pools memory across many chips at low latency.
“It allows many chips to connect together and use a shared memory pool at a very, very fast, low latency.”— Robert Wachen, Etched Co-founder and COO
The company is still working to shake an early-days perception problem. When Etched first pitched investors, the plan was to etch a specific model's architecture directly into the silicon — a bet that a single frontier model would dominate long enough to justify custom hardware. That plan is no longer the product. Etched's current systems run any frontier model, and Wachen has been correcting the record with customers who still assume otherwise.
The investor list has become a who's-who of AI infrastructure bettors: Kleiner Perkins, Sequoia Capital, Andreessen Horowitz, Peter Thiel, Tiger Global, Bain Capital Ventures, Neo, Stripes, Primary, Positive Sum, Diffusion, Argo, and Blackstone. Blackstone's presence is notable — the asset manager has been building exposure to AI compute infrastructure through datacenter deals, and a chip-startup check fits that thesis.
The skeptical read is that a $21 billion valuation for a company shipping first-generation hardware to a small number of customers requires Etched to displace a meaningful slice of Nvidia's inference revenue within the next 24 months to justify the entry price. Nvidia's inference dominance is not built on silicon alone; CUDA, the software stack, and the developer ecosystem are the real moat. Etched will need to prove that customers care more about tokens-per-dollar than about the tooling they already know.
The market is now pricing at least one credible inference-specific challenger to Nvidia at eleven figures, and Jane Street's decision to move from evaluator to lead investor is the kind of signal that pulls other quantitative-heavy buyers — high-frequency trading firms, hyperscaler infrastructure teams, sovereign compute buyers — into the same evaluation loop. If the second and third Jane Street–equivalent customers show up in the next two quarters, the $21 billion mark will look conservative. If they don't, Etched will spend 2027 defending it.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
