Chip startup Etched draws offers at $40-50 billion

TechCrunch reported on October 5 that AI chip startup Etched is reviewing offers valuing it at $40 billion from top-tier investors and up to $50 billion from lesser-known backers. Talks are early and terms may change.

Etched was valued at $10.3 billion in July in a Sequoia-led round and at $21 billion in September when it raised $700 million. It reports about $1 billion in secured customer orders and a 10 MW data center in Silicon Valley; around 15% of staff are ex-Nvidia.

Its chip, Sohu, is an ASIC that runs only transformer models, the architecture behind nearly all of today's language models. Nvidia's GPUs are general-purpose; Etched bets that a single-purpose chip does the job faster and cheaper, and claims Sohu processes more tokens at lower cost than Nvidia hardware. It is made at TSMC.

If a new architecture replaces transformers, a transformer-only chip loses value fast. Nvidia, AMD and cloud companies' own chips such as Google TPU and Amazon Trainium compete in the same market.

A GPU is a flexible chip that can run any model. An ASIC is built for one job, like bitcoin mining rigs, trading flexibility for efficiency. Inference is a trained model producing an answer, which happens billions of times a day, so cutting its unit cost is hugely valuable.

AI spending is shifting from training to inference, since models are trained once and run billions of times. Cheaper inference hardware feeds directly into the price of AI services.

A $40-50 billion valuation would make Etched worth more than many listed chipmakers while its secured orders total about $1 billion. Investors are pricing the future size of the inference market, and such valuations can retreat quickly if growth disappoints.

Competition in inference hardware lowers the unit cost of AI. Recheck the cost of heavy-use automations each quarter, prefer usage-based pricing over long fixed contracts, and look at real-time products that faster inference makes possible.