Startup Architect Labs says its AI-designed chip beats NVIDIA’s Jetson Orin Nano

Palo Alto–based startup Architect Labs, which emerged from stealth in June with a $24 million seed round, is boasting that its AI-designed and validated Redwood chip beats one NVIDIA edge product, the Jetson Orin Nano, in a specific AI-inference comparison.

The company projects that a Samsung 8 nm implementation of Redwood would deliver 1.75× the inference throughput while consuming about 47% less power than a measured Jetson Orin Nano baseline.

The flex is based on a projection. Redwood currently exists as a design deployed on an AMD Versal FPGA at 250 MHz. Architect Labs calibrated its Samsung 8 nm estimate using measurements from that FPGA implementation, then compared the modeled results with a measured Jetson Orin Nano reference.

In related news, NVIDIA announced the Jetson Orin Nano 2 earlier this week, which is scheduled to arrive in the first half of 2027. NVIDIA says the new module will deliver twice the inference performance of the Jetson Orin Nano Super and consumes 40% less power at equivalent performance. The current Nano Super had already raised the original Nano’s peak AI performance from 40 to 67 TOPS through a software upgrade.

In a press release, Architect Labs said its AI system generated the design in less than two weeks from an architectural specification drafted by two human architects.

Architect Labs is pitching Redwood for power-constrained inference in robots, drones and other edge devices. It says the accelerator could also be integrated into a larger data-center SoC or manufactured as a standalone chiplet.

Redwood would enter a crowded field. NVIDIA says more than 3 million developers build on its robotics stack, while Ambarella reports an installed base exceeding 46 million units for its edge AI platform. Qualcomm offers Dragonwing processors with up to 100 dense TOPS, AMD markets production-ready Kria modules for robotics, and Hailo offers a 40-TOPS INT4 accelerator for edge LLMs and vision-language models.