Skip to content
NewsFunding & M&A

Etched closes $300M Series C at $10.3B as inference startup doubles its valuation in 7 months

· by Pondero Newsdesk

The short version

Etched closed a $300 million Series C on July 23 led by Sequoia Capital, doubling its December valuation to $10.3 billion with $1 billion in customer orders already booked.

Etched closes $300M Series C at $10.3B as inference startup doubles its valuation in 7 months

A $10.3 billion valuation is notable. Reaching it while carrying $1 billion in booked orders, at Series C, separates Etched's July 23 announcement from the typical AI hardware funding release.

What

Etched closed a $300 million Series C on July 23, 2026, led by Sequoia Capital, per TechCrunch. The round valued the company at $10.3 billion, up from $5 billion in December 2025, when the company raised a $500 million round. Etched describes this as the highest valuation ever for a Sequoia-led Series C, per TechCrunch's coverage of the announcement.

Andreessen Horowitz, SK Hynix, Jane Street, Diffusion Capital, and Argo co-invested. Individual backers include Peter Thiel, Andrej Karpathy, Dylan Field, and Amjad Masad. The $1 billion in booked orders was disclosed June 30 and reflects demand for the company's first rack-scale inference system, which is currently in customer testing, per Etched's progress page.

The company was founded in 2022 by CEO Gavin Uberti, COO Robert Wachen, and Chris Zhu, all three of whom left Harvard before graduating. The team has grown to more than 400 engineers. Etched recently opened an 80,000-square-foot, 10-megawatt validation facility in Milpitas, separate from its main San Jose office.

Two proprietary technologies anchor the hardware. Low Voltage Inference runs the chip's compute blocks at under half the voltage of conventional AI accelerators, reducing thermal throttling and allowing sustained FLOPs throughput above 80 percent of peak on large mixture-of-experts models, per the June 30 technical disclosure. Cluster Scale Memory uses a proprietary interconnect to pool HBM and SRAM across multiple chips, cutting decode latency across a scale-up domain. The architecture supports transformer, MoE, and state-space model workloads rather than a single model class.

Why it matters

The $1 billion order book is a harder signal than benchmark claims. It means real AI labs are committing production budgets to an Etched system they have physically tested, not a roadmap. SK Hynix's role as a strategic investor ties the startup directly to a leading memory supplier, reducing the supply-chain concentration risk that shaped Nvidia's H100 allocation dynamics over the past two years.

COO Wachen told TechCrunch that researchers including Karpathy, Noam Brown of OpenAI, and Geoffrey Hinton were among those who tested the hardware in private demos before the round closed. The list carries weight because each of those names has a track record of evaluating compute independently of financial incentives.

For operators planning inference infrastructure for late 2026 and 2027, the key figure to track is not valuation but tokens per dollar versus Nvidia H100 and H200 clusters. Etched's Low Voltage Inference and Cluster Scale Memory claims would need to close that comparison by a margin large enough to justify rebuilding a software stack on non-Nvidia silicon. Those numbers are not yet public.

What to watch next

Etched expects first production shipments in Q3 2026. Initial benchmark disclosures on tokens per dollar and tokens per watt against Nvidia H100 and H200 clusters will arrive alongside those early deployments. Customer names disclosed in the first public deployment reports will be the clearest signal of where the system actually fits in production inference stacks.

Sources