Etched Raised $300M From Sequoia and SK Hynix. The Hard Question Is the $1B in Pre-Orders.

Etched closed a $300 million Series C on Tuesday at a $10 billion pre-money valuation, led by Sequoia Capital with SK Hynix, Andreessen Horowitz, Jane Street, Blackstone, and Diffusion Capital on the cap table. The round lifts the AI inference chip startup's total funding to $1.1 billion, according to the company's announcement and an EE Times report that visited Etched's San Jose lab.

The figure that has to justify the new valuation is $1 billion in pre-orders from customers president Robert Wachen won't name, paired with performance benchmarks Etched declined to publish. Both decisions are the company's own. Wachen, on the record at the lab, did not disclose customers or performance numbers. That refusal sits in plain view of a Series C cap table that, on paper, looks like the most serious assembled for a chip startup that has not shipped product.

Etched makes chips for the inference phase of AI. Inference is the part of the pipeline that runs an already-trained model to produce answers, as opposed to the training phase that builds it. The pitch has changed since 2024. Two years ago, founders Gavin Uberti and Chris Zhu, both Harvard dropouts, were selling a chip that "burns the transformer architecture into hardware" and claimed an order-of-magnitude throughput advantage over Nvidia's Blackwell generation. The current pitch has broadened to flexible inference across mixture-of-experts (MoE) models, diffusion models, and state-space models. That is a wider net, and it walks back the original "one architecture in silicon" framing.

A hardcoded transformer assumes a specific mathematical pattern. Flexible inference is a bet that the underlying model architecture will keep diversifying, and that a chip optimized only for transformers will become a specialty product as MoE and diffusion workloads grow.

The investors signal seriousness in different ways. Sequoia's lead is a vote from one of the firms that has made more money on AI infrastructure than almost any other. SK Hynix is the more interesting name on the list. The South Korean memory maker is a primary supplier of high-bandwidth memory (HBM), the stacked DRAM that sits next to Nvidia's GPUs in modern AI servers. SK Hynix joining the cap table is consistent with a memory supply partnership as much as a financial bet, though the press release does not characterize it that way. Jane Street, the quantitative trading firm, is a recurring check-writer for AI infrastructure and rarely puts capital into companies without a working product roadmap.

The verification gaps are not subtle. The press release from Etched's June 2026 stealth exit called the chip "working" and described $1 billion in "customer contracts" without naming any of the counterparties. The current Series C announcement does the same. Demos have been observed on-site in San Jose, where the company runs a small cluster, but no third party has published independent benchmarks against Nvidia Blackwell or any other inference chip. Without those, the order-of-magnitude advantage Etched originally claimed is exactly that, a claim.

Three things would move the $1 billion in pre-orders from company statement to falsifiable demand. A named customer with a signed purchase commitment. A ship date with a deliverable silicon timeline. An independent benchmark, run on a known workload, against a known Nvidia part. Etched has so far offered none of the three, and the Series C announcement does not introduce any of them.

Etched's new Milpitas R&D site, with a 10-megawatt data center and a surface-mount technology line for assembling circuit boards, is a real physical footprint, not a slide. That is one more data point than most pre-revenue AI chip startups offer. The rest of the story is the same kind of disclosure choice any inference chip announcement will have to make. A real cap table can coexist with an unverified demand number, and the reader has to hold both at once.