AI designed an AI chip itself: two weeks from a text prompt to a working prototype

The Architect Labs Platform ran performance simulations itself, generated the circuit description, tested it, and built the firmware with software cores. In less than two weeks, the Redwood Nano accelerator came to life on a programmable board: 12.1 tokens/s on the Qwen3-0.6B model at 250 MHz. Updating the spec produces a new board version in under two days.

In 2024, 14% of chip projects achieved working silicon on the first try, while 75% fell behind schedule (VentureBeat, 18.09.2026). A cycle the industry measured in quarters took two weeks here.

Architect Labs was founded by Ibrahim Hussein and Aaditya Subedi, aged 20 and 19. The 25-person team includes chip-industry veterans with 20–30 years of experience. In June 2026, the company raised $24 million in a seed round.

What is measured and what is calculated. On the board, Redwood Nano delivers 12.1 tokens/s, while Nvidia Jetson Orin Nano reaches 28 tokens/s on the same model. The promised 1.75× speed and 3.4× performance per watt apply to Samsung 8 nm silicon that does not yet exist. The forecast for it is 49 tokens/s at 1.335 W.

The company claims 95% test coverage and zero bugs in the first version of the circuit. Mike Bartley, who reviewed the report for Alpinum Consulting, focuses on that figure: 95% coverage is not 95% correctness until the coverage models themselves are disclosed. When one system writes both the circuit and its tests, an error in understanding the spec propagates across every artifact at once.

You cannot reproduce it yourself. The platform is closed, with access available through partnership. What is public is a technical report on arXiv and live demos at DAC 2026.

Anthropic's Claude Opus 4.6 runs inside the pipeline. The system performs about 90% of checks on its own simulators, using industrial ones for the final verification.

Until Redwood's first silicon, all comparisons with Nvidia remain calculations on paper.