Etched's $21B AI chip startup lures Nvidia talent, signs Jane Street

Etched, a $21 billion AI inference chip startup, has signed Jane Street as its first customer while recruiting engineers from Nvidia, escalating the race to challenge the incumbent's grip on AI computing.

The company, founded by Harvard dropouts, operates its own in-office data center — a capital-intensive approach that gives it direct control over chip testing and verification, the Wall Street Journal reported Aug. 18.

Jane Street, a quantitative trading firm known for its technology-driven strategies, becomes Etched's first named customer, a win that extends beyond the hyperscaler market. The startup's in-house data center allows engineers to test inference chips in real-world conditions rather than relying on external cloud infrastructure. Etched has been actively recruiting from Nvidia, targeting engineers with expertise in AI accelerator design and high-bandwidth memory integration.

The moves come as the AI inference chip market becomes a battleground between Nvidia's dominant GPU line and a wave of specialized challengers. For investors, Etched's $21 billion valuation and its ability to attract both Nvidia talent and a sophisticated customer like Jane Street suggest the market is pricing in meaningful alternatives to Nvidia's data center dominance.

Inference silicon heats up

AI inference — the process of running trained models to generate outputs — has become the fastest-growing segment of AI computing as companies deploy large language models at scale. Unlike training chips, which prioritize raw compute throughput, inference chips must balance latency, power efficiency, and cost per query. Etched's focus on this segment positions it against Nvidia's H100 and B200 accelerators, which currently dominate both training and inference workloads.

The startup's decision to build an in-house data center is unusual for a chip company of its size. Most fabless semiconductor startups rely on cloud providers or third-party facilities for testing. By owning its infrastructure, Etched can iterate faster on chip designs and demonstrate real-world performance to prospective customers — a key selling point when competing against Nvidia's CUDA software platform and developer tools.

Jane Street's adoption is notable because quantitative trading firms demand extremely low latency and high reliability from their computing infrastructure. The firm's willingness to deploy Etched's chips in production environments provides independent confirmation of the startup's performance claims.

The recruitment of Nvidia engineers is part of a broader talent war in AI semiconductors. As startups like Etched, Cerebras, and Groq raise billions to challenge Nvidia, they are competing for a limited pool of engineers with expertise in accelerator architecture, high-bandwidth memory, and advanced packaging. Nvidia's stock-based compensation packages have historically made it difficult for startups to compete on pay, but the prospect of equity in a $21 billion company is proving attractive to some engineers.

For investors, the key question is whether Etched can convert its valuation and talent into meaningful market share. Nvidia's data center business generates tens of billions of dollars in quarterly revenue, and its CUDA software platform creates significant switching costs for customers. Etched will need to demonstrate that its specialized inference chips offer enough performance-per-dollar advantage to justify the migration. Nvidia shares have priced in continued dominance of the AI accelerator market; any meaningful erosion of that position could trigger a sector-wide revaluation of AI chip stocks.

This article is for informational purposes only and does not constitute investment advice.