Written by: Techub News Compilation
In a16z's deep dialogue program, two entrepreneurs dedicated to bringing AI into the physical world—Diode Computers CEO Davide Asnaghi and Unlimited Industries CEO Alex Modon—shared their insights. They explored how AI is changing industries that traditionally rely on deep expertise and manual processes from two starkly different scales: microscopic (circuit boards) and macroscopic (large infrastructure projects). This conversation revealed that AI is not just about software and data; it's about how to move atoms, build physical entities, and reshape the future of American industry.
From Code to Circuit Boards: How AI Becomes a Hardware Designer
Davide Asnaghi's company Diode Computers is leveraging AI to design and manufacture custom circuit boards. His goal sounds ambitious: “I want to be able to start a hardware company as easily as my friends start B2B SaaS companies.” This means making hardware design—a task traditionally seen as extremely complex and specialized—accessible like writing software.
However, the field of hardware design suffers from a lack of data, and engineers are accustomed to specific tools and processes. Asnaghi's solution is clever: “We cheated.” They circumvented the challenge of directly training AI to understand the physical design of circuit boards by utilizing a domain where AI already excels—code. They built a compiler that provides AI models with sufficient prompts to make them feel like they are writing a Python program rather than designing a circuit board. Essentially, they have restructured the physical design problem into one that an AI model already has massive training data for (code).
“Traditional electronic design is not done in code at all,” Asnaghi explained, “and many sophisticated electrical engineers disdain the idea of coding. That's fine; we don't want those who don't want to code to code. Let the models code. The model's ability to write code is remarkable; it understands these concepts.”
This "everything is code" approach allows AI to generate hardware designs that are directly manufacturable. Asnaghi's predictions regarding the timeline for automation are quite aggressive: “I once thought it would take 5 years, but now I believe it may only take 2 years.” Of course, he refers to specific types of electronic design (a subset he focuses on) achieving fully automated design within two years.
Automation is not just about design. On the manufacturing side, circuit board assembly (SMT) already has a lot of robotics automation, but about 20% of the work remains difficult to automate, such as soldering large transformers and assembling housings. This part typically relies on human labor. Asnaghi points out that simply relying on labor in the U.S. cannot rapidly scale the production of data centers or reduce construction times from four years to two years. Therefore, their core objective is not to completely replace design work with AI but to automate those design types that produce "highly manufacturable" outputs. If the designs themselves are constrained (i.e., optimized for manufacturing), then 100% automated manufacturing can be achieved today—“the robots are already here.”
Asnaghi's vision is to expand the definition of “software engineer”: “You want to give everything that can generate code the same ability to generate hardware.” This includes not only skilled software engineers but also AI agents. Their open-source compiler toolchain (diode.link/pcb) is the infrastructure built for this purpose, aiming to provide a seamless track for AI and humans to design hardware and integrate into the manufacturing process.
Regarding the data challenge, Asnaghi believes that the field of circuit board design lacks sufficient public data to train a truly powerful foundational model. Their strategy is to first bootstrap using code methods to generate a verified library of design modules that will serve as training data for the next generation of models, creating a compounding effect. Meanwhile, if they can become a free design platform for the masses (rails), data will naturally converge. He believes that the biggest barrier is a lack of data, not model architecture. As long as profitable manufacturing can occur in the U.S. at costs competitive with Asia, sufficient data can be generated to spike the model's accuracy.
Infrastructure Automation: Redesigning Large Projects with AI
Alex Modon's Unlimited Industries targets large-scale infrastructure projects—power plants, hospitals, and large facilities. He made a bold assertion: “Within 10 years, all construction will be fully automated.”
Currently, a large infrastructure project begins when a developer has an empty plot of land. Following that, it may take up to a year or even a year and a half for planning, with hundreds of engineers (mechanical, process, electrical, civil) involved, collectively completing a massive "construction release package" (IFC). This document is subsequently handed over to the general contractor for procurement and construction.
Modon described their vision for automating this front-end process: inputting site information and construction requirements, AI will explore thousands of different design permutations and generate a globally optimized IFC package with a single click. The optimization goals can include capital expenditures (CapEx), but a better approach would be the project's total cost of ownership, including operational maintenance and construction convenience.
Similar to the parameterized, flexible design approaches in the software industry, Modon hopes to bring this model into infrastructure design. “For our clients, the greatest value is that if you spent 6 months designing something... If you want to change something 6 months later, oh my gosh... you have to start all over again. This is completely a nightmare.” However, under their AI-driven model, everything is just a variable for updating and can iterate efficiently.
Back-end construction involves more robotics, from more immediately feasible autonomous bulldozers to a future with numerous humanoid robots and drones on construction sites. Modon believes that with the right incentives, this future is destined to manifest within ten years.
Like hardware design, the infrastructure industry also faces challenges of data scarcity and traditional working methods. Modon pointed out that the incentive structures in infrastructure are closely tied to risk aversion, where capital sources dictate the incentive mechanisms for the entire project, often suppressing the adoption of new technologies. “You walk into these companies, look at people's computers, and you feel like you are stuck in the late 90s.” Therefore, Unlimited Industries chose a vertical integration strategy, “We must have enough parts to establish a clear interface with the industry.” Instead of trying to penetrate a small segment and force people to change.
In team building, Modon's experience is: “Teaching a cross-disciplinary person to use the latest and greatest AI tools is much easier than the reverse (teaching AI experts to become domain experts).” Their team is primarily composed of mechanical, electrical, civil engineers with multi-disciplinary backgrounds, along with some talents in AI and software.
Regarding AI's understanding of physical world constraints, Modon similarly adheres to the “everything is code” philosophy. They embody complex system relationships within a robust ontological model, providing a coding environment framework for AI agents and LLMs that can utilize various deterministic tools, just like ordinary engineers. This creates a parameterized relationship that eases optimization and iteration.
Challenges, Opportunities, and Future Outlook
Both CEOs acknowledge that achieving complete automation still has the final gap to bridge. For Diode, AI currently can accomplish about 90% of the design work, leaving 10% that requires human engineers to review and refine. Asnaghi believes that “the final frontier is that we don't have enough data.” Data is something society needs to generate. He personally believes that the existing architectural components are sufficient, but data is key. His co-founder has a different view, believing that many problems are very suitable for Monte Carlo tree search reinforcement learning-style approaches. In any case, their strategy is to focus on building end-to-end systems and be ready to leverage any breakthroughs in architecture or data generation at any time.
For Unlimited Industries, Modon believes the data sparsity of their problems is higher, but most issues can be bounded, and there are numerous standards governing how to build within the industry. He emphasizes the system design philosophy: “Ensure that the system you design is genuinely autonomous, rather than requiring human intervention.” This drives a very different architecture. They also bet that models will get better.
Regarding broader physical automation such as humanoid robots, both are optimistic. Modon believes that the efficiencies and learning rates enabled by mass manufacturing around a single design will be enough to offset the losses in efficiency due to customization, making humanoid robot forms very important. Asnaghi jokingly stated that he “loves all robots equally” because they all have circuit boards inside. He is more optimistic about robotic arms with visual language action models (VLA) to tackle that remaining 20% of work that is difficult to automate.
Both mentioned the existence of “tacit knowledge” in the industry—those intuitions and tricks possessed by experienced engineers and electricians. Modon pointed out that the average salary of electricians in the U.S. may have already surpassed that of software engineers in Silicon Valley, with high demand. Asnaghi believes there is a larger cultural disconnect in the U.S. circuit board industry regarding the disconnection between design and manufacturing. Designers often design in an “ivory tower” and then outsource manufacturing to others, leading to a lack of visceral connection for “design for manufacturing” (DFM). In China, even if designers know manufacturing is not handled by them, they will design carefully to make production easier for their friends' factories. One of AI's goals is to automatically generate these DFM-ready designs.
Finally, regarding the second-order effects of their work and mission, Asnaghi hopes to inspire the next generation of Americans to build tangible things. “I hope American engineers, American teenagers are passionate about building tangible things and say, ‘I want to build a cubic star and send it into orbit, do something cool,’ and it's easy because I can quickly launch and manufacture my circuit board the next day.”
Modon is concerned about the decline of the U.S. infrastructure capacity. He comes from the software world, where everything is advancing, while metrics such as labor productivity in the infrastructure sector have been deteriorating over the past 50 years. “The charts I see when I close my eyes at night look just like that.” Their goal is to improve one aspect of the lifecycle of constructing these projects by an order of magnitude or even multiple orders of magnitude, thus having the capability to tackle the whole problem from incentivizing projects to constructing many more—everything from the data centers' energy needed for AI success to all advanced manufacturing companies required for re-industrialization, along with critical minerals. All this constitutes the core framework of the world we see, and we are becoming worse, which is deeply concerning.
The conversation between the two entrepreneurs reveals a common theme: the application of AI in the physical world is not just a technical challenge, but also a reshaping of industry culture, incentive mechanisms, and the data ecosystem. By encoding physical problems, they are building a bridge towards a more efficient, creative, and future-ready physical industrial landscape.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
