AI chip startup Etched doubles its valuation to $21 billion in a month

Etched, a San Jose startup that builds specialized computer chips for running artificial intelligence models rather than training them, said it has raised $700 million in a funding round led by the trading firm Jane Street, valuing the four-year-old company at $21 billion. The round, detailed in an announcement posted to the company's own site, marks the second time in a single month that investors have roughly doubled their estimate of what Etched is worth, and it comes with an unusual twist: Jane Street is not only the company's newest investor but also its first paying customer.

According to the funding announcement published on Etched's website, the company's valuation has climbed from $5 billion in December to $10.3 billion after a $300 million Series C round in July, and now to $21 billion following the new raise — an increase of nearly $11 billion in the space of about four weeks. Participants in the latest round include Kleiner Perkins, Sequoia Capital, Andreessen Horowitz, Tiger Global, Bain Capital Ventures, Blackstone and investor Peter Thiel, among others.

A month of doubling valuations

The pace of Etched's fundraising reflects how aggressively venture investors are chasing companies that claim to have solved a narrower, more tractable problem than building general-purpose AI chips: making inference — the process of actually running a trained model to answer a prompt — faster and cheaper than it is on Nvidia's graphics processors. Etched's pitch is that by designing chips specifically for the transformer architecture underlying most modern large language models, rather than for a broad range of computing tasks, it can strip away silicon Nvidia's chips carry for workloads Etched has no interest in serving.

That focus has produced two components the company says are novel: what it describes as a low-voltage prefill chip and a cluster-scale memory interconnect that lets many chips draw on a single, fast, shared pool of memory rather than each holding its own separate allocation. Co-founder Rob Wachen described the significance of that design choice in comments reported by TechCrunch.

Cluster-scale memory "allows many chips to connect together and use a shared memory pool at a very, very fast, low latency," Wachen said, explaining the interconnect technology at the center of Etched's pitch to customers.

Etched shipped its first production rack to Jane Street last month, and the quantitative trading firm has since begun running its own workloads on the hardware in its data center — the kind of real-world validation that startups building physical infrastructure, rather than software, typically need years to obtain. Beyond Jane Street, the company says it has secured more than $1 billion in customer contracts spanning public and private AI companies and cloud providers, though it has not named most of those customers.

Betting on inference over training

Etched's technical claims have not gone unchallenged, and independent, audited benchmarks comparing its chips directly against Nvidia's latest hardware are not yet public. But the company's own technical materials describe an approach built around co-designing chips, circuit boards and cooling systems together as a single system, rather than selling a chip and leaving customers to build everything else around it — a strategy the company summarizes internally as "production is the product." Etched says it delivers more tokens of AI output per watt of electricity consumed than general-purpose alternatives, though it has not published a specific multiplier against any named competitor's hardware.

The company's chips are fabricated by Taiwan Semiconductor Manufacturing Co. on a 4-nanometer-class process, and Etched has said it completed a working test of its inference software just 44 days after receiving its first chips back from the foundry, a pace it has used to argue that a small, tightly integrated team can iterate faster than larger, more bureaucratic rivals.

Building a team by hiring from Nvidia

Much of that engineering talent has come directly from the company Etched is trying to unseat. Roughly 15 percent of Etched's approximately 400 employees previously worked at Nvidia, according to a report from Tech Startups, which described how a systems engineer who spent nearly 23 years at Nvidia joined Etched in 2024 and has since helped recruit roughly a dozen more Nvidia veterans, some of whom turned down counteroffers to make the move. Etched was founded in 2022 by Gavin Uberti and Chris Zhu, two Harvard dropouts in their twenties who have leaned heavily on more experienced hires — including a former chief technology officer from a $9.4 billion acquisition and an executive who helped launch the original iPhone's supply chain — to compensate for the founders' relative youth in an industry historically defined by decades-long engineering careers, according to the company's own careers page.

Whether that team can keep pace with its fundraising is likely to be tested soon. Etched has said it intends to use the new capital to move toward what it calls "gigawatt-scale" production, which the company says will require new factories, expanded global supply chains, custom fleet-management software and what it terms self-improving kernel agents to keep its chips running efficiently as it scales. SiliconAngle reported that the company has already opened a factory in Taiwan and an 80,000-square-foot prototyping facility near its San Jose headquarters to support that expansion.

What comes next

Etched's rapid valuation growth mirrors a broader surge of investor enthusiasm for companies positioned around AI inference rather than training, as cloud providers and AI labs increasingly look for ways to run existing models more cheaply at scale rather than simply building ever-larger ones. Whether Etched's chips deliver the efficiency gains the company claims once they are deployed more broadly — and whether more than a handful of large customers are willing to bet critical workloads on a four-year-old chip supplier rather than an established vendor — remains to be demonstrated outside of Jane Street's data center. The company has not disclosed a timeline for when its next round of customer deployments, or independent performance benchmarks, might become public.