Etched's $300M Series C at $10.3B: Sequoia Doubles Down on Inference Silicon

Superseded mark. Etched’s live valuation is $21 billion after a $700 million round on August 18, 2026 (Jane Street lead and first named customer). Full memo: Etched $21B. The July Series C below is the prior print.

Etched just did what skeptics said a transformer-specialist chip startup could not: double to a $10.3 billion valuation on a $300 million Series C led by Sequoia Capital, with Andreessen Horowitz back in the round (July 23, 2026).

Deal snapshot

Who uses the product

Etched sells rack-scale inference systems, not loose ASICs on a brochure. Buyers are large AI labs and platforms that need cheaper, faster serving than general-purpose GPU fleets for production token volume.

Access has been gated — investors and early customers first, including technical luminaries who sat through office demos (Karpathy, Noam Brown, Hinton named by the company). That scarcity is both go-to-market discipline and a diligence risk: few public benchmarks for outsiders.

The technical bet (plain English)

Inference splits into prefill (understand the prompt — compute-heavy) and decode (emit tokens — memory-bandwidth-heavy). Etched built silicon for both:

• Low-voltage prefill → less heat → denser transistors.
• Cluster-scale memory for decode → many chips share a fast memory pool.

Company claim: systems run transformers and MoE models (DeepSeek, Qwen) plus non-transformer designs like Mamba — pushing back on "etched for one LLM" skepticism.

Why Sequoia leads — portfolio logic

Sequoia takes legendary-company shots from idea to IPO. Partner presence (Sonya Huang / Abishek Malani photographed with founders) signals hands-on conviction. Leading the highest-valuation Sequoia Series C (per company) after TSMC silicon and order bookings is a manufacturing-and-delivery bet, not a paper architecture bet.

a16z returning keeps Etched inside the firm's AI infrastructure cluster — compute scarcity is still the binding constraint for model companies a16z already backs.

Likely reasons Etched took this syndicate:

1. Brand + LP patience for multi-year hardware cycles.
2. Strategic memory partner (SK Hynix) beside pure financial capital.
3. Trading-firm sophistication (Jane Street) that understands performance economics.

Competitive map

When the $10.3B mark looks wrong

• Order book slips into multi-year delivery risk.
• General GPUs keep winning on software ecosystem even at worse TCO.
• Model architectures shift away from Etched's silicon assumptions faster than tape-outs.

Practical takeaway

Operators: Demand workload-specific $/token and time-to-first-token on your models — private demos are not a purchase order. Investors: The next diligence gate is delivered racks, not booked LOIs.