Designing the Physical World with AI

What It Covers

A16z General Partner Aaron Price Wright speaks with Alex Modin of Unlimited Industries and Davide Asnagi of Diode Computers about applying AI to physical world design — specifically large-scale construction engineering and custom PCB manufacturing — and the timelines, data challenges, and vertical integration strategies required to automate both industries.

Key Questions Answered

• •Code-as-hardware abstraction: Diode Computers bypasses the scarcity of PCB training data by reframing circuit board design as code generation. Their open-source compiler (diodeinc/pcb on GitHub) gives AI models enough structural hints that designing a circuit board resembles writing Python — leveraging the vast existing code training data LLMs already possess.
• •Design-for-manufacturing as the automation bottleneck: Electronics manufacturing is already 80% robotic via surface mount technology, but the remaining 20% — oversized transformers, board-to-enclosure assembly — stalls full automation. Diode's strategy is to constrain AI-generated designs so tightly to manufacturing specs that existing robotic lines can handle 100% assembly without waiting for robotics hardware improvements.
• •Parametric modeling eliminates construction redesign costs: Unlimited Industries builds all project designs as fully parametric code models, meaning a change six months into a year-long design process updates as a variable rather than triggering a full restart. This directly compresses project schedules by three to six months, which materially improves project IRR for infrastructure investors.
• •Vertical integration as the entry strategy for entrenched industries: Both companies conclude that selling software tools to traditional industries fails due to high switching costs and misaligned incentives. The viable path is owning enough of the end-to-end workflow — design through manufacturing or procurement through construction — to deliver a finished product customers already know how to buy, faster and cheaper.
• •Data generation, not model architecture, is the primary constraint: Asnagi argues the core barrier to fully automated PCB design is not a missing AI breakthrough but a lack of training data, currently siloed at companies like Apple, Meta, and SpaceX. His strategy is to become the open manufacturing infrastructure so that designs flowing through Diode's platform generate the proprietary dataset needed to compound model accuracy over time.