Etched, a startup building specialized chips for AI inference, announced Tuesday it raised $700 million at a $21 billion valuation, led by quantitative trading firm Jane Street. The jump is notable less for the dollar amount than for the pace: Etched was valued at $5 billion in December, hit $10.3 billion in a Series C just last month, and has now roughly doubled again in a matter of weeks.
What changed investors' minds wasn't a pitch deck. Jane Street tested Etched's hardware directly, bought it, and is now running it in its own datacenter. That's a meaningfully different signal than a venture firm writing a check on projections - a demanding, latency-sensitive trading operation put real workloads on the chips before committing capital. Nevada dispensary POS platform
Etched's pitch centers on splitting inference into two distinct problems. The "prefill" stage, where a system parses a prompt and its context, is compute-heavy; the "decode" stage, where the model actually generates output tokens, is memory-heavy. Etched built a low-voltage prefill chip that packs in more transistors without the heat penalties typical of high-end AI silicon, and paired it with what it calls cluster-scale memory - an interconnect that lets many chips share a single fast memory pool. The company sells these as complete "frontier inference clusters," its answer to what Nvidia markets as AI factories.
Shedding the One-Model Myth
Etched has spent much of its short life correcting a misconception baked into its own name: that each chip is "etched" to run one specific frontier model and nothing else. That was the original plan. It isn't anymore. Current systems are designed to run any frontier model, which matters enormously for buyers - locking a very expensive rack of hardware to a single model that could be obsolete in a year is a nonstarter for most enterprise infrastructure planning.
Why the Valuation Math Moves So Fast
Inference, not training, is where AI compute spending is increasingly concentrated, since every user query against a deployed model consumes hardware cycles indefinitely, unlike training runs that happen once. Investors betting on Etched are effectively betting that specialized inference silicon can undercut general-purpose GPUs on cost and speed at scale. Whether that thesis holds against Nvidia's dominant position and its own aggressive product cadence is the open question the market is pricing in real time - and pricing it, apparently, several billion dollars higher each month.
A Crowded, High-Stakes Cap Table
Etched's backers now include Kleiner Perkins, Sequoia Capital, Andreessen Horowitz, Peter Thiel, Tiger Global, Bain Capital Ventures, Neo, Stripes, Primary, Positive Sum, Diffusion, Argo, and Blackstone. That roster spans traditional venture, growth equity, and now a working quant trading firm as both customer and lead investor - a combination that gives Etched credibility beyond the usual "trust us, it scales" claims common in early-stage AI hardware.
