AI chip startup Etched issued a progress report on June 30 saying it has achieved first-pass manufacturing success with TSMC on its Sohu inference chip and has now booked more than $1 billion in customer contracts, as the company prepares to ship its first rack-scale systems this summer. Etched is building Sohu as a transformer-only ASIC — a chip whose circuitry is hardwired specifically to run transformer-based AI models, rather than the general-purpose architecture Nvidia GPUs use to handle a broad range of workloads.

Betting specialization beats general-purpose GPUs

Founded by Harvard dropouts, Etched has raised roughly $800 million to date, including a $500 million round in December 2025 that valued the company at $5 billion. The company's central wager is that as the industry standardizes on transformer architectures for large language models, a chip built to do nothing else can outperform general-purpose GPUs like Nvidia's on both raw throughput and performance-per-watt for inference workloads. Etched has also said it designed the surrounding rack infrastructure itself — circuit boards, cooling plates, and networking — rather than relying on off-the-shelf systems, and has described using a "low-voltage inference" approach intended to reduce overheating and extract more performance from the silicon.

First silicon success clears a major milestone

Reaching first-pass (A0) silicon success on TSMC's N4P process is a significant technical checkpoint for any chip startup — it means the initial manufacturing run functioned close enough to specification that Etched did not need a costly and time-consuming re-spin before moving toward production. Combined with the $1 billion in contracted orders, the update signals Etched is moving from a funding-and-benchmarks story toward an actual shipping product, with its first customer racks scheduled to go out this summer.

Still unproven at scale

Etched's performance claims — reported elsewhere as 10-20x better throughput and performance-per-watt than Nvidia's H100 on transformer inference — have not yet been validated by independent third-party benchmarks, since Sohu has not previously shipped in volume to external customers. The chip market Etched is entering remains dominated by Nvidia, and rival specialized-inference and custom-silicon efforts from companies including Cerebras, Groq, and the major cloud providers' in-house chips are also competing for the same demand. Whether Etched's transformer-only bet holds up once systems are in customers' hands — and whether the transformer architecture itself remains AI's dominant paradigm — will determine if this summer's shipments turn booked contracts into a durable business.