Etched, a company founded by three Harvard dropouts, has recently secured $120 million in funding from investors including Peter Thiel. The company aims to take on Nvidia in the field of AI chip design by creating a new chip called Sohu that specializes in transformer inference. By burning the transformer architecture into its chips, Etched claims to have created the fastest transformer chip ever. This is significant considering Nvidia’s recent rise as the world’s most valuable company, surpassing Microsoft with a valuation of $3.3 trillion. Despite this, Etched believes its team of 35 employees can outperform Nvidia.
### The Rise of Transformer Models in AI
Etched’s bet on transformer models is backed by the observation that these models have the potential to scale unlike any other AI model. While other models like CNNs, RNNs, and LSTMs have their limitations, transformers have shown the ability to handle larger AI models effectively. CEO Gavin Uberti explains that by scaling AI models and feeding them more training data, they become significantly smarter. This trend is evident in the increase in FLOPS used to train state-of-the-art models over the past five years. Uberti believes that this scaling will continue, with major companies like Google, OpenAI, Amazon, and Microsoft investing billions of dollars in AI data centers.
### The Need for Specialized Chips
Etched argues that as AI models continue to scale and become more complex, the demand for specialized chips will grow. General-purpose chips like GPUs have limitations in handling transformer inference efficiently. Etched believes that specialized chips are inevitable and that they will outperform GPUs in transformer inference tasks. The company compares this trend to other computing markets that eventually led to the development of specialized chips, such as Bitcoin mining. With trillions of dollars at stake in the AI market, Etched is confident that specialized chips will dominate the industry.
### Etched’s Strategy for Success
Etched acknowledges that its specialized chips may not outperform GPUs in all workloads. However, the company is confident that its chips will excel in transformer inference, which powers most generative AI products. By focusing on this specific task, Etched’s chips are claimed to be an order of magnitude faster than Nvidia’s next-generation GPUs. The company believes that as the AI market consolidates around transformer models, the demand for specialized chips will increase, giving Etched a competitive advantage.
### The Potential Impact of Sohu Chips
If Etched’s bet on transformer models and specialized chips pays off, it could have a significant impact on the AI industry. Today, AI models are expensive and slow to build, hindering their widespread use. However, Etched claims that with its Sohu chips, real-time video, audio, agents, and search will become possible. This could lead to a complete inversion of the unit economics of AI products, making them more accessible and efficient. While the challenge of surpassing Nvidia may seem daunting, Etched is confident that its unique approach to parallel processing and specialization will position it as a formidable competitor in the AI chip market.
