Christian Keil, who worked on satellite programs at Astranis, thinks AI will become a better hardware designer than any human soon, and that this will settle very little. In an essay on a16z’s newsletter, he argues that design is only the first step of making a physical product, and the later steps depend on a connection between computers and the real world that loses information in both directions. a16z is a venture investor, so read the piece as an investor-side view of the market.
Keil’s evidence that design is nearly solved is the benchmark record. He points to CAD Arena, EE Bench, and his own LuxoBench, a test that had three models, Astra from OpenAI, Fable 5.1 from Anthropic, and Grok 4.6, help build an interactive desktop lamp. His verdict is that the models cannot yet do hardware engineering unaided, but the gains are steady and he would not bet against them.
The trouble starts when a design has to become an object. Keil walks through five levels of difficulty. Buying off-the-shelf parts that fit together is the first. Custom parts come next, where a CAD file must pass through a machinist who picks the stock, the tooling, and the cutting strategy. Then come assemblies, where tolerances stack up so that every part can match its drawing and the whole still fails to fit. After that come outside vendors; Keil notes that SpaceX, a symbol of vertical integration, lists 591 current employees with supply chain titles and is hiring 128 more. The last level is volume, where building one unit and building a million are different jobs.
He then explains why the physical side resists automation, in five reasons. Real objects vary and drift. Data from them is sparse, since a thermal team cannot put sensors everywhere. Precise measurement is expensive. Physical processes take fixed time, so a test campaign cannot be compressed however fast the redesign happens. And some feedback is qualitative, such as whether a device feels good to hold. His anecdote makes the point: a friend who buys 45 million chickens a week saw the saleable yield swing by as much as 20 percent from week to week, even with identical suppliers and machines.
From this, Keil predicts where the work will go. Better simulation tools would let engineers iterate at software speed, leaving physical builds to confirm results rather than discover them. Demand should move toward raw materials and components, which he says will favor suppliers of wire harnesses, machined parts, electronics, and copper. Companies armed with cheaper design talent may also start making more of the components they now buy from others.
The scarcest input, in his account, is what he calls process knowledge: the machinist who hears a problem in the sound of a CNC machine. Citing Eric Jang’s 2026 estimate, Keil describes a steep funnel for robotics training. Egocentric video runs to millions of hours, gripper recordings to several hundred thousand, teleoperated robot sessions to a few tens of thousands, and high-quality demonstrations for hardware-specific work to just hours. He cites Periodic Labs, which runs an autonomous lab for physical experiments, as one way to generate that data directly.
The hardest step, he suspects, is learning what customers want. An AI could draw a million versions of a smartwatch and have them made cheaply, but it would still need people to tell it which one they like.
Keil’s case is a hypothesis, and he offers no measurement of how fast any of these barriers is falling. The chicken yield and the stacked tolerances are persuasive illustrations, yet they are one person’s experience. His logic does carry a testable claim for buyers, though. If design costs collapse while testing, sourcing, and customer feedback stay slow, the money in hardware shifts to whoever owns those slow steps, so a startup pitching “AI hardware engineer” in the coming year deserves one question first: where does its own data from the physical world come from?
Based on an essay by Christian Keil published in a16z’s newsletter on 20 October 2026.
