Two founders bet code-first AI can automate boards and construction

1. Design before build — ▶ 2:13

   Alex Modon splits a construction project into design and build. Design alone can take up to 1 to 1.5 years and involve hundreds of engineers producing an issued-for-construction (IFC) package. He says automating that stage end to end is within line of sight: feed in a site and requirements, and AI explores tens of thousands of design permutations, optimizing for total cost of ownership rather than only CapEx. The build phase is robotics, from autonomous earth movers in the short term to humanoids and drones, which he expects over roughly a decade if incentives align.

2. Two-year board timeline — ▶ 5:22

   Davide Asnaghi says the capability jump between each publicly available Anthropic model is "wild." His team once thought full automation of board design would take 5 years and now says 2. He limits the claim to a subset of electronics design and rejects blanket statements about all of it. He also concedes that his timelines keep "getting shorter and shorter," which flags the figure as a forecast that has already moved once.

3. The 80/20 assembly gap — ▶ 6:26

   Surface-mount robots place components and an oven bakes the board, but roughly 20% of the work resists automation: large transformers needing other soldering processes, components that hug the board, and enclosure assembly. Foxconn and Pegatron solved this with labor, which Asnaghi says cannot double US data center production or cut bring-up time from 4 years to 2. His bet is AI that designs manufacturable boards. If the design is constrained, he says, 100% automation is possible with robots that already exist.

4. Agents as hardware engineers — ▶ 8:51

   Asnaghi reframes his goal of turning software engineers into electrical engineers. The population that produces code now includes agents, so the aim is to give anything that can generate code the ability to generate hardware. He cites coding agents exploiting the structure of code for tasks that are not code at all. Agent-designed boards are limited in complexity today, but he says extrapolating the improvement curve gets to the target in a year or two. Skilled engineers keep building complex designs alongside.

5. Capital sets incentives — ▶ 11:04

   Modon argues that construction resists new technology because of who funds it. Projects pass through stage-gated risk removal until they are fundable by investors who want a set rate of return, with no upside, so no one gains from adopting new tools. He compares it to defense before Anduril. His answer is vertical integration: own enough of the workflow to offer a clean interface, then expand. The team is mostly mechanical, electrical, civil, simulation and controls engineers, because teaching domain experts AI tools is easier than the reverse.

6. Sell outcomes, not software — ▶ 14:54

   Asnaghi rejected building CAD software, citing high switching costs. Diode sells the finished board against a customer's specifications, including to Fortune 100 firms, with real-time visibility similar to a GitHub pull request. Its compiler toolchain is open source. He claims 90% efficiency gains over a world without the tools, with 10% still done by Diode's engineers. Modon adds that engineering is a small share of project cost; the value is compressing schedules by 3 to 6 months, which affects IRR and financeability.

7. Code sidesteps data scarcity — ▶ 19:56

   Board data is two to three orders of magnitude too small to train a foundation model, Asnaghi says, so Diode reframes the problem as code. The compiler hints make the model feel it is writing Python, and a library of validated blocks becomes training data for later models. If Diode becomes the rails others design on, and manufactures in the US at a cost competitive with Asia, he expects data to flow to it. He is bullish on diffusion architectures for the harder physical problem, without elaborating.

8. Parametric projects, not snowflakes — ▶ 22:56

   Modon says every construction project is an n-of-one snowflake, which he identifies as the core reason the industry does not get better at building. His fix is an ontological model of the whole system, within which agents write code and call deterministic engineering tools. Everything becomes parametric. A change six months into design, which today cascades into a restart, becomes an updated variable. He adds that projects are "Legos on hard mode": valid designs must match inputs and outputs, so most answers are calculable.

9. Simulation as training tool — ▶ 27:09

   Asnaghi notes that human electrical engineers rely on intuition and use simulation only to verify at the end. He wants simulation used at training time to ground the model's physical sense and build that taste, since it is too slow at inference. He claims taste emerges with enough data and covers 95% of cases. He still argues physics must be the final judge: the best engineers he knows spend 3 weeks simulating and still distrust a design until it is built in its enclosure.

10. Data versus algorithms — ▶ 30:10

    Asnaghi says data is the last frontier for circuit boards. It exists but is siloed at companies such as Apple, Meta and SpaceX, and he considers open sourcing it unlikely. His co-founder Lenny argues that problems formulated for Monte Carlo tree search and reinforcement learning may need no more data, and Asnaghi defers to him. Modon says construction data is even sparser, but standards bound the problem and the status quo is a low bar. He designs for full autonomy, with no human in the loop, which he says drives a different architecture.

11. Designer proximity problem — ▶ 42:12

    Asnaghi argues that US designers are insulated from manufacturing pain, so design-for-manufacturing skill atrophies. In China, he says, designers know the person on the line: a friend avoided double-sided boards because it meant 2 SMT passes for someone he knew. He doubts this closeness can be hired in or recovered from retirees. His alternative is to have models iterate cheaply, for example working through 150 manufacturability checks, to rebuild mass-production capability that he says US contract manufacturers lost because the economics stopped working.

Two founders working at opposite ends of the physical economy, circuit boards and large infrastructure, converge on one strategic bet: the fastest route to physical-world AI is to turn design into code, the medium models already handle fluently. Davide Asnaghi of Diode Computers describes a compiler that gives a model "enough hints that it feels like it's writing a Python program instead of designing a circuit board." Alex Modon of Unlimited Industries states the same philosophy: if you play "in a language that the models already understand," life is "a bajillion times easier." The shared claim concerns representation more than robots. Once a hard physical artifact can be expressed as code, general-purpose agents can work on it.

Asnaghi extends the logic to the labor pool. Software has structure that agents exploit even on tasks that are not code, so he wants to give anything that can generate code, including agents, the ability to generate hardware. Diode open-sources its compiler toolchain because it wants to own the infrastructure rather than the design primitives. Anything generated will work in its ecosystem and can be sent to manufacturing. That is a notable choice. Instead of building a CAD business with high switching costs, Diode sells the finished board against a customer's specifications and treats how it is made as an implementation detail.

Modon applies the same idea to buildings, where the obstacle is that every project is a one-off. His remedy is an ontological model of the whole system, a simple representation of a complex one, inside which agents write code and call deterministic tools. The payoff is parametric design. Today a change six months into design can cascade into a restart, whereas in his version it is an updated variable. The platform can explore tens of thousands of permutations and optimize for total cost of ownership rather than only construction cost. He argues the problem is tractable because valid designs must match inputs and outputs, which makes answers calculable.

The economic case rests less on cheaper engineering than on time. Modon says engineering is a small share of project cost, so the value lies in pulling schedules in by 3 to 6 months, which changes financeability. Asnaghi makes a parallel argument for manufacturing. Robots already place most components, and he sees the obstacle in the roughly 20% of work, such as large transformers and enclosure assembly, that has been handled with labor. His claim is that designs constrained to be manufacturable can run at 100% automation with today's robots. Without redesign, he says, cutting data center bring-up from 4 years to 2 is out of reach.

Both founders also see incentives, not only technology, as the constraint. Modon argues that construction capital comes from investors seeking a set return with no upside, so no one in the chain gains from adopting new tools. His response is vertical integration: own enough of the workflow to give the industry a clean interface, then expand. Asnaghi's version is selling outcomes, with the pitch that Diode is faster and cheaper, plus real-time visibility into the process. Both choices are partly hedges. If the models are not fully capable yet, a company that delivers the finished product can still be useful, and Diode's engineers close the remaining 10%.

The hardest problem is data, and here the account is least settled. Asnaghi says board data is two to three orders of magnitude too small for a foundation model. His bootstrapping plan is a library of validated blocks that becomes training data, plus data drawn from designers who use Diode as free rails if it can manufacture in the US at a cost competitive with Asia. Simulation plays a specific role in his scheme: a training-time tool for building taste, not an inference-time step. He argues this because the best engineers he knows will not trust a simulation until they have built the hardware.

Asnaghi is candid that this strategy rests on an open dispute. He believes data is the last frontier, much of it locked inside companies such as Apple, Meta and SpaceX. His co-founder Lenny argues that Monte Carlo tree search and reinforcement learning, in a self-play formulation, may need no more data. Asnaghi says he usually defers to Lenny, and Diode will build end to end meanwhile. Modon is firmer. He says construction is even sparser in data but bounded by standards, and he designs for full autonomy rather than a human in the loop. If models stall, he admits, they are "back to the drawing board."

The evidence behind these claims is thin. The 90% efficiency gain is Asnaghi's own figure, with no baseline given. The 2-year timeline has already moved from 5, and he admits his estimates keep shrinking. Modon's 10-year construction horizon bundles an automatable design phase with a robotics phase that depends on incentive changes he says he has not yet seen. These are founder forecasts from companies selling the technology. They explain mechanisms well, notably code as a bridge, parametric models and training-time simulation, but they do not show that the last 10% yields to scale rather than to a research breakthrough.

The human side of the argument deserves equal attention. Asnaghi says US designers are insulated from the cost of manufacturing, while a designer in China may know the person on the line and avoid a double-sided layout to spare that person two passes. He doubts that closeness can be hired in or recovered from retirees. His workaround is to have a model iterate through manufacturability checks, which costs little in effort. Modon notes a tacit-knowledge shortage in the trades, citing an anecdote from a Microsoft CTO relayed by the host about Microsoft employing a third of Georgia's electricians during one data center build.

The stakes Modon names are larger than efficiency. He says US construction labor productivity and adjusted CapEx have gotten worse over about 50 years, and that extrapolating the line means losing the muscle to build large projects, from energy for data centers to critical minerals. Asnaghi's counterpart is cultural: he wants a teenager to build a CubeSat and have boards manufactured the next day. Both rest on whether models can absorb what manufacturing pain once taught designers. The unresolved question is whether that comes from more data, better algorithms, or a pipeline that keeps feeding models the results of real manufacturing.