Alex Modon | The a16z Show Summary

1. Key Themes

Construction Is a Software Problem in Disguise

Both guests converge on the insight that the key to automating physical world design is reframing it as a code generation problem. Construction and PCB design share a structural parallel: complex systems of interrelated variables that, when expressed as code, become tractable for AI agents.

"Everything is code. And if you can play in a language that the models already understand, your life is like a bajillion times easier." — Alex Modon [00:23:01]

"We basically built a compiler that gives the model enough hints that it feels like it's writing a Python program instead of designing a circuit board." — Davide Asnagi [00:20:44]

The Incentive Structure of Capital Projects Is the Root Cause of Construction's Stagnation

The reason construction has resisted technology adoption is not cultural stubbornness — it is structural. The capital stack rewards risk removal and stable IRR, not upside capture, which eliminates any incentive for innovation at every level of the project hierarchy.

"The source of capital sets the incentives all the way down to the project... When you get to someone who might have some new piece of technology, it's just super unincentivized because there's no upside. Like no one actually wins from that environment." — Alex Modon [00:12:09]

"If you look at basically any construction metric — labor productivity or adjusted CapEx numbers over the past 50 years — we're getting worse." — Alex Modon [00:47:21]

Vertical Integration Is the Only Viable Go-to-Market for Physical AI

Neither company can sell software tools to incumbent workflows. Both have concluded that they must own the full stack — design through delivery — in order to offer a clean interface to the market and actually demonstrate value.

"Vertical integration is like, we have to own enough of it that we can actually do a clean interface to the industry rather than trying to pick off one small part and then force people to change and adopt that technology." — Alex Modon [00:13:02]

"You don't want to convince people to buy your software. You want to convince people to buy the end product. And this is a much harder company to build." — Davide Asnagi [00:16:34]

Data Scarcity Is the True Final Frontier for Physical AI

Unlike software domains, circuit board and construction design data is sparse, siloed, and not available on the open internet. The path to fully autonomous physical design runs through data generation strategies, not just model architecture improvements.

"The last frontier standing is we don't have enough data. The data is the thing that we need to generate as a society if we want circuit boards to be automated by AI. The data exists. It's usually siloed into the Apples, Metas, SpaceXs of the world. And they will not obviously fork it over." — Davide Asnagi [00:28:44]

Designing for Full Autonomy From Day One Creates a Different Architecture

Both companies treat full autonomy as a design requirement, not a future aspiration. This shapes every architectural decision they make — and is distinct from the typical "AI-assisted" product that keeps humans in the loop indefinitely.

"Making sure that you design a system to actually be fully autonomous and to not be human in the loop — for us at least, it feels like it's driven a very different architecture." — Alex Modon [00:34:19]

The 80/20 Automation Gap in Electronics Manufacturing Is an Underappreciated Bottleneck

Surface mount technology has automated the majority of PCB assembly, but the remaining 20% — chunky transformers, board-to-enclosure assembly, non-standard components — remains manual and is the critical constraint on U.S. domestic manufacturing scale.

"There's been companies like Foxconn, for example, for Apple or Pegatron that have solved that problem with labor. And that makes complete sense for certain segments and in certain geographies. But you're not going to double the production capacity for data centers in the U.S. by just relying on labor alone." — Davide Asnagi [00:07:38]

The Engineering Value Prop Is Speed, Not Cost

In large capital projects, engineering fees are a tiny fraction of total cost. Compressing design time by months has an outsized impact on project IRR — and that is the true lever AI companies in this space must pull.

"The value prop for us is not that it's cheaper from an engineering perspective. Engineering is such a small percentage of the costs anyway... If you can bring in a schedule by three months or six months, that actually materially impacts the finance ability." — Alex Modon [00:19:28]

Open-Sourcing the Design Toolchain as a Data Acquisition Strategy

Diode Computers open-sourced its core compiler toolchain specifically to generate a flywheel of validated design artifacts — each output manufactured through their ecosystem becomes training data and a manufacturing lead.

"Our core compiler tool chain is open source. If any electrical engineer wants to take it and run with it, please be my guest. Because what we did is once you generate an artifact, it will work in our ecosystem. We will be able to take it and send it for manufacturing. We want to own the infrastructure, not the core design primitives." — Davide Asnagi [00:17:28]

The Design-for-Manufacturing Muscle Has Atrophied in the U.S. Due to Offshoring

The cultural and cognitive disconnect introduced by decades of sending designs abroad has degraded the U.S.'s DFM intuition. This is not recoverable by hiring alone — it requires AI that enforces manufacturing constraints at the design stage.

"A lot of the result of being able to just send your designs, like ivory tower — you are designing in the U.S. and then sending to manufacturer somewhere else. You're abstracting the manufacturing somewhere else. You kind of don't feel that pain. And that's why the design for manufacturing muscle kind of atrophies." — Davide Asnagi [00:42:33]

2. Contrarian Perspectives

Construction Will Be Fully Automated End-to-End — No New Breakthrough Required

The conventional view is that construction is too complex, too variable, and too human-dependent to automate. Alex argues the opposite: the domain is actually highly bounded by standards and first-principles engineering, making it more tractable than it appears — and the status quo bar is so low that even incomplete AI wins decisively.

"Most all those problems can be bounded... the benefit of this space is there is an incredible amount of standards that govern how something should be built... to beat status quo, the bar is so unbelievably low. So yeah, I would say take the under on it." — Alex Modon [00:33:26]

Simulation Should Be a Training-Time Tool, Not an Inference-Time Tool

The industry assumption is that simulation is the gold standard for validating designs. Davide argues that at scale, simulation is too slow for inference, and the goal should be training models to develop the intuitive taste that engineers currently carry in their heads — so they rarely need to simulate at all.

"Simulation needs to be a training tool. And then you kind of need to get physics to tell you you're right or you're wrong. So reducing the training time is probably the most important part." — Davide Asnagi [00:29:41]

The Right Hire for an AI-Native Physical Company Is a Domain Expert, Not a Software Engineer

Counter to the default Silicon Valley instinct of hiring software engineers and teaching them the domain, Unlimited Industries found the reverse is far more effective: take the domain expert and teach them AI tools.

"Most of our team is actually not software engineers. In our experience, it's been much easier to teach a multidisciplinary person the latest and greatest AI tools than the other way around." — Alex Modon [00:13:52]

The U.S. Already Has World-Class PCB Manufacturing Capability — Just at the Wrong Tier

The common narrative is that the U.S. has lost PCB manufacturing capability to Asia entirely. In fact, U.S. contract manufacturers have elite capabilities — but only because they have been forced into high-end military contracts. The mass production tier has simply evaporated due to economics.

"All the contract manufacturers for PCBs in the U.S. are very talented. The capabilities of contract manufacturers in the U.S. are super high because they only bid on military contracts, which require the highest possible capabilities. But what you want to build is this new set of mass production capabilities, which has been kind of evaporated by the industry because the economics didn't make sense." — Davide Asnagi [00:44:46]

Hardware Startups Should Be as Easy to Spin Up as B2B SaaS

The prevailing assumption is that hardware is inherently slow and capital-intensive, an asset class structurally different from software. Davide argues this is a tooling problem, not a physics problem — and that solving it unlocks a Cambrian explosion of hardware companies.

"I want to be able to spin up a hardware company the same way that my friends spin up B2B SaaS. Like, you should be able to say, I want to do something that's considered very hard and just go and do it." — Davide Asnagi [00:45:54]

3. Companies Identified

Diode Computers

AI-native circuit board design and manufacturing company. Uses a code-first compiler approach to make AI agents design PCBs as if writing Python. Building toward cost-competitive U.S.-based manufacturing at scale.

"We are 90% more efficient at building boards than we were in a world without our tools." — Davide Asnagi [00:18:52]

Unlimited Industries

AI-native firm vertically integrating design, engineering, procurement, and construction for large infrastructure projects. Generates optimized "issued for construction" packages parametrically using AI agents.

"AI is going to explore tens of thousands of different permutations about how to optimally design that facility, a button click. And then what you get back from that is a globally optimized issued-for-construction package." — Alex Modon [00:03:51]

Anthropic

AI research company whose Claude models are a core part of Diode Computers' design workflow. Mentioned for the pace of capability improvement between model generations.

"We do a lot of work with Anthropic, and the jump that we see in design capabilities between each model tier — like publicly available model — is wild. We thought it would be five years. I think that I can probably say two." — Davide Asnagi [00:06:00]

Foxconn / Pegatron

Contract electronics manufacturers cited as the incumbent solution to the 20% manual assembly problem — using concentrated labor in specific geographies — and implicitly the model Diode is working to displace domestically.

"There's been companies like Foxconn, for example, for Apple or Pegatron that have solved that problem with labor. And that makes complete sense for certain segments and in certain geographies." — Davide Asnagi [00:07:38]

ANSYS

Commercial electromagnetic simulation software used in professional PCB design workflows. Mentioned as an existing tool that Diode intends to shift from inference-time use to training-time use.

"We've had tools in electrical engineering, like Spice at the schematics level. And then electromagnetic simulation kernels, like OpenEMS on the open source or ANSYS on the actual board level." — Davide Asnagi [00:27:03]

Microsoft

Cited as a striking example of the skilled trades shortage bottleneck facing U.S. data center construction.

"I was talking to the CTO at Microsoft and he was telling me that one time Microsoft employed a third of the electricians in the state of Georgia when they were building a big data center there." — Erin Price-Wright [00:41:26]

Apple / Meta / SpaceX

Named as the holders of siloed, proprietary PCB design data that would — if aggregated — be sufficient to train high-quality models for circuit board automation.

"The data exists. It's usually siloed into the Apples, Metas, SpaceXs of the world. And they will not obviously fork it over." — Davide Asnagi [00:31:01]

4. People Identified

Lenny (Co-founder, Diode Computers — last name not given)

Described by Davide Asnagi as "the smarter one." Holds a competing thesis to Davide on AI capability: that Monte Carlo tree search / reinforcement learning style formulations could solve PCB design without requiring large new datasets.

"My co-founder, his name is Lenny. Him and I have a philosophical disagreement. His take is that a lot of these problems are really well structured for Monte Carlo tree search reinforcement learning style — where you can basically formulate a problem with two players playing against each other and they get better just by nature of improving recursively." — Davide Asnagi [00:30:32]

David Hansen

Friend of Davide Asnagi, described as building high-quality motors using rare earth / Western Magnetics materials. Mentioned as a collaborator and example of the physical manufacturing ecosystem Diode wants to support.

"I have a good friend, David Hansen, who builds beautiful motors with Western Magnetics materials. And we want to build them." — Davide Asnagi [00:39:06]

Dan Wang

Technology analyst, cited by Erin Price-Wright for his concept of "process knowledge" — the tacit, embodied industrial knowledge that China has accumulated through decades of hands-on manufacturing.

"Dan Wang calls it the process knowledge, which China has in spades, which the U.S. has to some degree." — Erin Price-Wright [00:39:51]

5. Operating Insights

Sell What the Customer Already Knows How to Buy

Both companies found that selling new tooling to entrenched industries fails. The winning motion is to fit the incumbent purchasing model — take specs in, deliver a physical product out — and keep the AI-native method as an internal implementation detail.

"You are used to effectively working with companies that deliver the exact same product, which is I give you my specifications, I get back a physical product. We fit into that mold. And then how we do it is an implementation detail to the company, but we do it faster and we do it cheaper." — Davide Asnagi [00:16:34]

Build Compounding Data Flywheels Into the Business Model From Day One

Diode Computers' open-source toolchain is not just a developer relations tactic — it is a data acquisition architecture. Every design produced on open rails feeds validated blocks back into training data, compounding model accuracy over time.

"If you become the rails that everybody can design for free on, that is data that comes your way. As long as you give enough incentives and say I'm going to manufacture in the U.S. at a cost competitive with Asia, you will be able to actually generate the amount of data that makes the model just skyrocket in accuracy." — Davide Asnagi [00:22:48]

Design the Product Architecture for Full Autonomy, Even Before the Models Are Ready

Rather than incrementally adding automation to a human-in-the-loop workflow, designing the system architecture to be fully autonomous from the start forces different and ultimately more defensible architectural choices.

"It's actually an important design paradigm — making sure that you design the system to actually be fully autonomous and to not be human in the loop. For us at least, it feels like it's driven a very different architecture." — Alex Modon [00:34:19]

Use Parametric / Model-Led Design to Turn Iterations From Restarts Into Variable Updates

In traditional construction, a design change six months in means starting over. A model-led parametric approach means every element is a variable — changes cascade automatically and the system re-optimizes globally rather than locally.

"If you spend six months designing something and you want to change something six months in, you'd start over. In this version, it's like everything's just an updated variable." — Alex Modon [00:24:19]

6. Overlooked Insights

The Data Center Build Cycle Is a Direct Market Forcing Function for Domestic PCB Manufacturing

This was mentioned only briefly but is extremely significant. Davide argues that the only path from a four-year to a two-year data center build cycle in the U.S. runs directly through redesigning and manufacturing circuit boards domestically at scale — making AI-native PCB manufacturing not a nice-to-have but a critical infrastructure bottleneck.

"You're not going to be able to reduce the time cycle that it takes to bring up a data center from four years to two years if you're not able to redesign the boards, redesign them for manufacturing, manufacture them at scale in a constricted timeline." — Davide Asnagi [00:08:02]

This makes Diode Computers not just a PCB company but a direct enabler of U.S. AI infrastructure scaling — a strategic position that is almost entirely absent from public discourse about the data center supply chain.

Diffusion Models Are a Stealth Moonshot for Physical Layout Problems

Davide mentioned almost in passing that he is "very bullish on diffusion as an architecture" for the specific problem of physical PCB layout — and immediately pulled back, saying he would discuss it "next time." This is a non-obvious technical bet: diffusion architectures have shown remarkable capability in spatial, generative domains (images, protein folding), and applying them to the 2D/3D placement problem of circuit board routing could represent a step-change over current agent-based approaches.

"I will just say I am very bullish on diffusion as an architecture. That's what I will say for this specific problem. And the thing is that you need to bootstrap it somehow." — Davide Asnagi [00:21:53]

No one in the conversation followed up on this, but if correct, it implies that whoever cracks diffusion-based PCB layout first — with sufficient bootstrapped training data — may leapfrog all current agent-based approaches in both speed and design quality.