Etched Secures $120M to Build Specialized Transformer ASIC Chips

AI hardware startup Etched has raised $120 million to manufacture Sohu, a specialized ASIC chip designed exclusively for executing Transformer models at unprecedented speed.

Silicon Valley artificial intelligence hardware startup Etched has announced a successful $120 million funding round. The capital will be utilized to manufacture Sohu, the company's highly anticipated application-specific integrated circuit (ASIC) designed exclusively to execute Transformer-based models. By hardwiring the mathematical structures of transformers directly into silicon, Etched promises to deliver inference speeds that outperform general-purpose GPUs by orders of magnitude.

#

The Paradigm Shift: Hardwiring the Transformer

Currently, almost all state-of-the-art AI models—from OpenAI's GPT-4o to Google's Gemini and Meta's Llama 3—are built on the Transformer architecture. While NVIDIA's general-purpose GPUs (like the H100 and Blackwell series) are incredibly powerful, they must remain flexible enough to run legacy code, graphics tasks, and alternative machine learning architectures.

Etched is placing a massive bet on a single architectural concept: Transformers will remain the dominant AI model for the foreseeable future. By dedicating 100% of the silicon real estate to transformer-specific math (like attention mechanisms), Sohu eliminates the overhead associated with general-purpose instruction sets.

#

Performance Claims and Technical Design

According to Etched, a single Sohu server can replace multiple GPU servers while achieving significant latency reductions: * Attention in Silicon: The attention mechanism, which allows AI models to contextualize relationships between words or tokens, is implemented directly as physical logic gates. * High-Bandwidth Memory Integration: Sohu is paired with ultra-fast memory stacks to feed the high throughput demands of massive model parameters without bottlenecking. * Extremely Low Latency: For real-time applications such as voice assistants and interactive coding agents, Sohu reduces time-to-first-token generation to milliseconds.

#

Venture Support and Market Dynamics

The funding round was led by prominent venture capital firms and strategic angel investors, including founders of major AI labs. The capital influx will fund the first tape-out and commercial manufacturing runs with major semiconductor foundries.

If Etched can deliver on its performance metrics at scale, it will offer a compelling alternative for enterprise customers who are currently bottlenecked by the high cost and low availability of NVIDIA's graphics cards. This marks a new phase in the AI startup ecosystem, where specialized hardware is emerging to challenge established chip giants.

Explore more AI hardware startups and compare funding data on the Startup Compare Tool and AI News.