AI
Etched AI Chip Startup Reaches $10.3B Valuation, Backed by Sequoia and Top-Tier Investors
The three-year-old company founded by Harvard dropouts has doubled its valuation in seven months and booked $1 billion in orders for its custom AI inference chips.
TechCrunch — AI
Etched, the chip startup launched by three Harvard dropouts in 2022, has secured $300 million in Series C funding that values the company at $10.3 billion. This round, led by Sequoia Capital with participation from Andreessen Horowitz, SK Hynix, Jane Street, and Diffusion Capital, marks a significant acceleration for a company skeptics once dismissed as pursuing a far-fetched vision.
The valuation doubling — from $5 billion just seven months earlier — underscores rapid momentum in a crowded field of AI chip makers. Notably, Sequoia's commitment brings the highest Series C valuation the firm has led, according to Etched's co-founder and COO Robert Wachen. The round also attracted marquee angels and early believers, including Peter Thiel, Andrej Karpathy, Dylan Field, and Amjad Masad.
From Skepticism to Silicon#
When Etched launched, the notion of building processors optimized exclusively for transformer-based AI models seemed quixotic to many. Yet the company has steadily validated its technical thesis: it already booked $1 billion in customer orders and, last month, announced the successful manufacturing of its chips through TSMC and began testing complete systems with clients.
The core skepticism stemmed partly from supply constraints — access to working systems remained limited to investors and early customers, limiting visibility into actual performance. That exclusivity became a feature: Wachen notes that high-profile technical leaders like Andrej Karpathy (Anthropic) and Geoffrey Hinton evaluated the hardware directly and became enthusiasts based on hands-on experience, not marketing claims.
A Different Architecture#
Contrary to early perceptions, Etched's systems run far more than a single proprietary model. The platform supports Mixture of Experts architectures (such as DeepSeek and Qwen) as well as non-transformer designs like Mamba, a state-space model variant. Google has since pursued a conceptually similar path with its Frozen v2 chip, suggesting the idea of embedding model-specific optimizations into silicon has shed its "wacky" reputation.
Etched's differentiator lies in two custom components designed to accelerate inference—the compute stage where models respond to user queries. Its prefill chip operates at unusually low voltage, reducing heat and enabling denser transistor packing, while a novel "cluster-scale memory" interconnect allows multiple chips to share a memory pool at high speed and low latency, promising both performance and cost efficiency.
