The artificial intelligence sector is witnessing a strategic pivot towards specialized hardware, particularly for AI inference—the process of running trained models to generate responses. While Nvidia's general-purpose GPUs have long dominated both AI training and inference, startups like Etched are betting on application-specific integrated circuits (ASICs) to deliver superior efficiency and cost-effectiveness for specific AI tasks. This specialization aims to address the growing need for optimized compute power as AI models become more complex and widespread.

Etched, a San Jose-based AI chip startup, is currently negotiating two venture funding rounds that could dramatically increase its valuation. One round, led by existing investor Jane Street, targets a $20 billion valuation, quadrupling its prior assessment. Concurrently, Sequoia Capital is leading a separate round aiming for a $10 billion valuation. These deals, while not yet closed, highlight intense investor confidence in Etched's approach and the broader market's hunger for alternatives to existing AI hardware solutions.

The company's flagship product, the Sohu chip, is a specialized ASIC designed specifically for transformer architectures, which underpin most modern large language models. Etched claims its architecture can run math blocks at significantly lower voltages, leading to multiple times the FLOPs density of current AI chips without thermal throttling. This focus on optimizing for transformer workloads promises best-in-class throughput, latency, and power efficiency for inference tasks.

Despite not yet achieving large-scale commercial deliveries, Etched has garnered substantial interest, with its website indicating approximately $1 billion in expressed customer demand. The startup, founded in 2022 by Harvard dropouts Gavin Uberti, Chris Zhu, and Robert Wachen, views its strategy as building a new category of AI hardware: frontier inference clusters. Their vertical integration, from chip design to manufacturing methods, aims to scale quickly to meet the demands of next-generation AI.

This aggressive fundraising and valuation target underscore a critical trend in the AI industry: the move beyond general-purpose computing towards highly specialized, efficient hardware. As AI applications proliferate, the economic and performance advantages of purpose-built chips for inference are becoming increasingly apparent, fostering a competitive landscape where new players can challenge established giants by focusing on specific, high-value workloads.