Etched

Etched develops what it calls frontier inference clusters, a category of AI hardware in which chips, racks, software and manufacturing methods are co-designed so that frontier models can be served with high throughput, low latency, low cost and high power efficiency across both prefill and decode workloads. The systems are targeted at many-trillion-parameter mixture-of-experts models, long-context inference and agentic workloads. The company's A0 silicon returned from TSMC's N4P process earlier in the year, and Etched is validating its first rack-scale product with customers.

The company describes two core technical approaches. Low Voltage Inference (LVI) runs the chip's math blocks at under half the voltage of typical AI chips to increase FLOPs density and sustain 80%+ of peak FLOPs on trillion-parameter sparse MoEs without thermal throttling; it requires co-design spanning splittable math arrays, circuit techniques, tiling and scheduling algorithms, power delivery networks, VRM architectures, advanced packaging and cold plate design. Cluster Scale Memory (CSM) creates a lower-latency shared memory pool across the scale-up domain using a proprietary ultra-low-latency, high-bandwidth interconnect, with a hybrid HBM/SRAM design intended to address both memory capacity and memory-to-memory latency while avoiding tradeoffs the company attributes to SRAM-only chips, 3D DRAM or optics.

Etched is vertically integrated, with math block designers, inference engineers, thermal specialists and supply-chain managers co-located. It operates a factory in Taiwan and has built a data center, test house and NPI prototyping lab at its San Jose office.

Founding story

The three co-founders — Gavin Uberti, Robert Wachen and Chris Zhu — are Harvard Thiel Fellows; Uberti is a Math 55 alumnus and AI compiler specialist who developed the Cortex-M backend for TVM, Zhu is a Math 55 alumnus and high-performance computing and combinatorics researcher, and Wachen previously co-founded Prod and Mentor Labs (acquired by Crimson Education). [0]

Business model

Etched designs and produces AI inference hardware, selling rack-scale inference systems to customers under contracts; the site references over $1B in customer contracts with first racks shipping in the summer. [0]

Traction

A0 silicon returned from TSMC N4P earlier in the year; first rack-scale product in customer validation; production started against over $1B in customer contracts, with first racks scheduled to ship in the summer. Early customer tests are described by the company as showing state-of-the-art throughput, latency and power efficiency. The company reports 400+ engineers and total funding of $800M. [0]

Latest developments

The company announced a $700M raise at a $21B valuation led by Jane Street and disclosed $800M raised across four previously unannounced financings, including a strategic investment from VentureTech Alliance. It also disclosed the Low Voltage Inference and Cluster Scale Memory architectures, said first racks ship in the summer, and said it would share further performance and roadmap updates that summer. [0]