Etched said Tuesday it raised $700 million at a $21 billion valuation, led by Jane Street after the quantitative trading firm tested and then purchased the startup's AI inference hardware. The jump is notable less for the dollar figure than for the pace: Etched sat at $5 billion in December, moved to $10.3 billion in July on a $300 million Series C, and has now nearly doubled again inside a single month.
What's driving the enthusiasm, according to co-founder and COO Robert Wachen, is architecture, not marketing. Etched built two components from scratch to accelerate inference - the compute-heavy process that kicks in after a user submits a prompt. Inference splits into two stages, Wachen told TechCrunch: a math-intensive "prefill" phase that parses the prompt and its context, and a memory-intensive "decode" phase that generates the actual output tokens a user reads. cannabis dispensary management software washington
Etched's prefill chip runs at low voltage, which lets it pack in more transistors without the heat penalties that typically come with high-end AI silicon - meaning it can push more tokens through faster. For the decode side, the company built what it calls cluster-scale memory, a new memory type paired with an interconnect that lets many chips share a single memory pool at low latency. Wachen says the combined effect is higher throughput at lower cost, though outside verification of those claims is still limited to early customers like Jane Street.
What Jane Street's Endorsement Actually Signals
Quant trading firms are notoriously demanding about latency; milliseconds move money. Jane Street's public statement - "We tested the chip and are pleased with the early results" - reads less like a courtesy quote and more like a firm that put the hardware through its own workloads before writing a check. That's a different kind of validation than a typical venture round, where backers often bet on a roadmap rather than a working rack already installed in a datacenter.
The "Etched" Name Still Confuses the Market
Etched has spent much of its short life correcting a misconception baked into its own branding. Early on, the company's premise was that it etched a specific frontier model directly into silicon - a chip built for one model, permanently. That's no longer how the systems work; Etched's current hardware, sold as full "frontier inference clusters," can run any frontier model, putting it in more direct competition with Nvidia's so-called AI factories.
Why the Valuation Math Looks So Aggressive
Investors including Kleiner Perkins, Sequoia, Andreessen Horowitz, Peter Thiel, Tiger Global, Bain Capital Ventures, and Blackstone are effectively wagering that inference - not model training - is where the next spending wave lands, and that a chip built specifically for that workload can undercut general-purpose GPUs on cost per token. That's a real bet, not a settled fact. Valuations moving this fast, on this little public performance data, deserve scrutiny rather than applause; the hardware is new, the customer base is small, and the real test comes when clusters run at scale under sustained commercial load rather than a single quant fund's benchmark.