❯ AI Chip Startup Etched Closes $700M Series D, Valuation Doubles to $21B

ORDERS FIRSTEtched closed a $700 million Series D at a post-money valuation of $21 billion, led by quantitative trading firm Jane Street. It sells its in-house Sohu inference chip as a complete rack — cabinet, software, and liquid cooling bundled in — so customers can run large-model inference as soon as it lands in their own data center. Founded in 2022 by three Harvard dropouts, the company has raised $1.9 billion cumulatively, with signed customer contracts now exceeding $1 billion.

8-MONTH CLIMBLast December, Etched was still valued at $5 billion. This July, when it raised $300 million, the figure was $10.3 billion. A month later it jumped to $21 billion — up more than fourfold in eight months. The inflection point is clear. Etched only emerged from stealth at the end of June, and its public scorecard at the time showed $800 million raised and $1 billion in orders — with not a single machine delivered. What actually rewrote the valuation curve was last month’s delivery: the first rack went into Jane Street’s data center. After testing, the firm called the results “satisfactory” and immediately routed its own live production load through the machine. A customer tests the box, then turns around and leads the round — in chip startups, that is about as hard a vote of confidence as exists. The money is not backing a roadmap; it is backing a machine already working in someone else’s data center.

POSITIONINGEtched’s bet sits on two pillars. First, low-voltage inference (LVI): it pushes the operating voltage of its compute units below half that of rival AI chips, buying several times the compute density per unit area — the company says sparse trillion-parameter models can sustain more than 80% of peak throughput. Second, cluster-scale memory (CSM): it mixes HBM with SRAM and adds a proprietary low-latency interconnect so chips across the whole rack share a single memory pool, purpose-built for long-context and multi-trillion-parameter mixture-of-experts models. The original Sohu chip hardwires Transformer computation directly into silicon; Etched’s published figure is 500,000 tokens per second on Llama 70B, roughly 20x a full eight-GPU H100 system. The A0 tape-out runs on TSMC’s N4P process. More than 400 engineers come from Nvidia, Google’s TPU team, Broadcom, SK hynix, and TSMC. Among inference-chip challengers, Groq has already veered into cloud, and Cerebras mostly sells chips and cloud services; Etched sells the whole rack from day one and won’t let customers assemble their own systems. For Nvidia, the trouble is that these dedicated machines only compete with it for orders in the inference segment.

WHAT CAPITAL BUYSA $21 billion valuation against $1 billion in contracts on hand — a contract multiple of just over 20x — sits in an extremely aggressive tier for a hardware company. But the pricing logic of this round is not complicated: investors are paying for inference capacity that is already installed, not for a chip roadmap. Over the past two years, most inference-side challengers died at the same spot: great numbers on paper, but customers would not move production workloads over. Etched cleared that hurdle with one rack inside Jane Street’s data center. The real pressure now falls on inference-chip companies still stuck at the sample and white-paper stage; the yardstick has switched from benchmark scores to installed base, and the fundraising bar has risen accordingly. What to watch from here is delivery cadence: whether the $1 billion in contracts becomes revenue depends on TSMC capacity and rack yield — not on issuing yet another set of benchmark scores.

▪ SIGNALA chip company’s most expensive valuation jump came from a customer plugging the machine in, not from another round of benchmark scores. Pricing power in inference hardware is shifting from performance claims to machines actually installed.