Etched AI develops a specialized transformer-focused inference chip for high-performance AI workloads

Etched AI, a startup founded by former Harvard computer science students, has emerged from stealth to introduce a specialized inference chip designed exclusively for transformer models. The company aims to achieve superior performance and efficiency by moving away from general-purpose architectures like GPUs in favor of a hard-wired, transformer-specific design. The chip utilizes low-voltage inference, operating at approximately 450 millivolts, to optimize power consumption and frequency. Additionally, the architecture incorporates "cluster-scale memory," which functions as a large, shared SRAM scratchpad accessible by compute units through high-speed interfaces. This design choice suggests a focus on addressing bandwidth limitations rather than just raw compute power. While the company has secured significant funding—raising $1.8 billion in six weeks across three rounds—the specific technical implementation, such as memory addressing and handling of transcendental functions, remains undisclosed. The startup is currently positioning itself as a challenger to established players like NVIDIA and AMD, with plans to provide more details on its roadmap later this year.

Etched AI is developing a specialized chip architecture designed specifically to accelerate transformer-based AI models. The company has raised $1.8 billion in funding over a six-week period across three investment rounds.

The chip utilizes low-voltage inference, operating at roughly 450 millivolts to improve power efficiency. A key architectural feature is cluster-scale memory, which provides a shared SRAM scratchpad for compute units.

The design prioritizes high-bandwidth data movement to address potential bottlenecks in transformer workloads. Etched AI is a top sponsor for the Hot Chips 2026 event, though they are not yet listed as a speaker.