Etched: From Skepticism to a Billion-Dollar AI Chip Powerhouse
Innovative AI Inference: Redefining Performance
Etched, established in 2022 by Harvard dropouts, recently completed a Series C funding round, securing an impressive $300 million and catapulting its valuation to $10.3 billion. This round saw lead investment from Sequoia, alongside significant contributions from Andreessen Horowitz, SK Hynix, Jane Street, and Diffusion Capital, among others. The company's valuation has more than doubled in just seven months, marking it as potentially the highest Series C valuation ever led by Sequoia. Etched's success is rooted in its ability to produce its own custom chips and its system's capacity to process $1 billion in orders, confirming its strong market position.
Beyond Transformers: Broad AI Model Compatibility
Initially met with skepticism for focusing on chips designed specifically for transformer-based AI models, Etched has proven its technology's versatility. Co-founder and COO Robert Wachen clarified that their systems are not limited to specific Large Language Models but can efficiently run a wide array of AI architectures. This includes Mixture of Experts models like DeepSeek and Qwen, which distribute tasks across specialized sub-models, and non-transformer designs such as Mamba, based on state-space models. This broad compatibility addresses critical concerns and expands the potential applications of Etched's technology.
Accelerating Inference with Novel Hardware
Etched's core innovation lies in its two custom-designed components tailored to speed up AI inference, the crucial processing phase after a user's prompt. Wachen explains that inference involves two stages: the computationally intensive 'prefill phase' for understanding prompts, and the 'decode phase' for generating output tokens, which requires substantial memory. To optimize these, Etched developed a prefill chip that operates at a significantly lower voltage, enabling more transistors and higher speeds with reduced heat generation. For the decode phase, they engineered a novel memory and 'cluster-scale memory' interconnect technology, allowing multiple chips to share memory rapidly and efficiently, thereby reducing overall costs.
Overcoming Early Doubts and Proving Innovation
The journey for Etched's founders—CEO Gavin Uberti, Robert Wachen, and CTO Chris Zhu—has been marked by persistent challenges and doubts from the tech community. When the company began, the concept of dedicated AI chips was largely unrecognized outside of Nvidia's domain. Despite manufacturing their first silicon batch with TSMC, skepticism remained, primarily due to limited access to their systems. However, private demonstrations to influential figures like Andrej Karpathy, Noam Brown, and Geoffrey Hinton, along with key investors, garnered enthusiastic support, validating Etched's innovative approach.
From Garage to Data Center: A Growth Story
The founders' beginnings were humble, starting with Wachen sleeping on a friend's floor and servers being rebooted by an employee's wife in a garage. Today, Etched has grown into a significant enterprise, employing 400 people and operating a 2-megawatt data center. The company is actively collaborating with leading AI firms, running tokens in their labs, and fulfilling orders. This remarkable transformation from a nascent startup to a major player in the AI chip market exemplifies their resilience and commitment. Wachen humorously notes his current comfort, symbolizing the progress made. He emphasizes that unwavering belief and persistent effort can overcome monumental challenges, leading to significant achievements in the face of adversity.
