Etched Emerges to Challenge NVIDIA: Meet the Startup Aiming to Change the Future of AI

The U.S. startup Etched has officially announced its exit from "stealth," presenting itself to the world with a complete hardware platform dedicated to the inference of artificial intelligence models. Founded in 2022, the company claims to have raised over $800 million in funding and secured more than $1 billion in contracts for its systems, despite its first product still being in the validation phase with customers.

The company's stated objective is not merely to create a simple AI accelerator but to develop a complete infrastructure comprising processors, racks, software, and jointly designed production technologies. According to Etched, this approach would allow for superior performance in terms of throughput, latency, power consumption, and operational cost while executing large language models, a phase known as inference.

"We're coming out of stealth. We've built our first racks after a successful A0 tapeout, $1B+ in customer contracts, and $800m raised. Early customer tests show us achieving SOTA throughput, latency, and power efficiency on inference workloads. Our first racks ship this summer." pic.twitter.com/FLccrkLTza — Etched (@Etched) June 30, 2026

The company claims that the first chip, identified as A0 silicon, has been successfully produced by TSMC using the N4P manufacturing process. The component is currently being used in the validation of the first rack systems intended for customers, while commercial deliveries are expected this summer. The new platform, named Frontier Inference Clusters, has been designed to handle large AI models, including those based on Mixture of Experts (MoE) architectures with trillions of parameters, applications with very extensive contexts, and workloads related to AI agents. To achieve these targets, Etched states that it has developed two proprietary technologies. The first, named Low Voltage Inference (LVI), is focused on increasing throughput. According to the company, traditional AI accelerators reduce their operational frequency when power consumption and temperatures rise, resulting in sustained utilization of computing capacity below 50% of the theoretical peak. In contrast, Etched's new architecture allows mathematical units to operate at less than half the voltage normally used by current AI chips, enabling higher computing density and maintaining up to 80% of theoretical performance without thermal throttling, even when executing MoE models with trillions of parameters. To achieve this, the company claims to have redesigned numerous elements of the infrastructure, from transistors to power systems, advanced packaging, dissipation, interconnections, and scheduling algorithms. The second technology, called Cluster Scale Memory (CSM), tackles the latency issue in the decoding phase of AI models. According to Etched, accelerators based exclusively on HBM memory cannot reach the speeds offered by SRAM, while fully SRAM solutions sacrifice memory capacity and computing power. For this reason, the startup has developed a hybrid solution that combines HBM and SRAM through a proprietary low-latency, high-bandwidth interconnection, creating a shared memory pool across multiple chips. The goal is to simultaneously reduce data access latency while maintaining high memory capacity, avoiding some trade-offs related to cost, reliability, yield, and thermal dissipation that characterize other architectures. Etched reports that the first tests conducted with customers have shown industry-leading results for throughput, latency, and energy efficiency in inference tasks. The company plans to publish additional performance data and the product roadmap over the summer. To support production, the startup has opened a facility in Taiwan dedicated to engineering activities and built a datacenter in San Jose with about 2 MW, a prototyping lab, and infrastructure intended for system validation before large-scale production. The company now employs over 400 engineers from companies like NVIDIA, Google, Broadcom, TSMC, and SK Hynix.

Among its investors are Jane Street, Hudson River Trading, Two Sigma, Ribbit Capital, and several notable names in the AI industry including Andrej Karpathy, Geoffrey Hinton, Fei-Fei Li, Arthur Mensch, and Scott Wu, along with investors Peter Thiel and Stanley Druckenmiller.