This is the tenth part in a series on how AI is crossing out of software and into the physical world. This time, Jan Bosch looks at AI designing the hardware that AI runs on, and what happens to engineering when iteration becomes almost free.
In 2021, a Nature paper from Google described a reinforcement learning agent that could place the macros on a chip in under six hours, with better power, timing and area than human engineers achieved. The self-improving loop had apparently closed: Better chips train better models, which design better chips. Five years on, that result is the most disputed claim in the field, and the shape of the dispute turns out to be more instructive than the result would have been.
This Alphachip controversy) is worth recounting precisely, because it’s easy to tell badly in either direction. A Google engineer, Satrajit Chatterjee, produced an internal analysis arguing that conventional methods did better; Google declined to publish it, his employment ended and he sued, alleging misconduct. Chung-Kuan Cheng and Andrew Kahng at UC San Diego re-implemented the method and reported that it didn’t beat simulated annealing, an academic placer, a commercial placer or human placement. In a 2025 update, it was confirmed that for full routing, it never produced shorter wire length than the commercial tool while using far more computation. Igor Markov published a critique in Communications of the ACM alleging cherry-picked benchmarks and selective baselines; ACM attached an expression of concern to it.
Google’s authors reject all of this, and not without argument. Their position, set out in a paper pointedly titled “That chip has sailed,” is that critics skipped the pre-training step that makes the method work and used insufficient compute. Nature published an addendum in 2024 adding missing methodological detail. The critics reply that the compute they used matched the paper.
I’m not going to adjudicate it, and I don’t think anyone reading this can. That’s the point. Because notice what’s actually missing. The training data was proprietary. The pre-training inputs were never released. There’s no common public benchmark on which the competing claims can be run side by side. Five years, a lawsuit, a journal addendum, an ACM expression of concern, and the reason it can’t be settled isn’t that the science is subtle; it’s that the evidence required to settle it doesn’t exist in a form anyone outside one company can use.
And this isn’t confined to the disputed case. Cadence states that Cerebrus has been used on more than a thousand production designs. Synopsys reported a hundred commercial tapeouts for DSO.ai, with up to 15 percent power reduction and better than threefold productivity gains, and says nine of its top ten digital-implementation customers are adopting its AI technology. These are probably all true. They’re also not comparable: The definitions of “design,” “tapeout” and “AI-driven” differ between vendors and public evidence correlating these claims with production results at advanced nodes is, as far as I can find, absent. A Synopsys executive remarked this year that reinforcement learning in core EDA algorithms, meaning placement and routing itself, “hasn’t really panned out,” which, from a vendor selling AI-driven EDA, is a usefully candid distinction between AI in the flow and AI replacing the optimizer.
So, let me separate what happened from what was announced, because something real did happen. AI didn’t arrive in design as a better decision-maker; it arrived as a cheaper search. The optimizers that place and route a chip were already extremely good; what was expensive was exploring the space of possible configurations: trying a hundred floorplans, sweeping constraint sets, testing architectural variants before committing. That exploration used to cost engineer-weeks. It now costs compute. Siemens has just agreed to buy Precision Innovations, the commercial company around the open-source Openroad flow, and the strategically interesting piece is a fast physical-estimation layer that could return power-performance-area feedback before RTL (register transfer level) decisions harden. That’s the actual shape of the revolution: not a machine with better taste, but feedback arriving early enough to change your mind.
Which brings me to what I think is the real consequence, and it isn’t the one in the press releases. When iteration is expensive, engineering is a sequence of choices. You think hard, you commit, you live with it. When iteration is nearly free, engineering becomes a search and a search optimizes exactly, and only, what you wrote down.
That relocates the scarce human skill. It used to be the ability to design the thing. It’s becoming the ability to say precisely what good means: which constraints are hard, which are soft, what you’re trading against what and what you forgot to mention. The objective function has quietly become the design. Everything else is execution, and execution is the part that got cheap.
This is a harder discipline than it sounds, and most organizations are bad at it, because for a century, they didn’t have to be. An under-specified objective used to produce a mediocre design that an experienced engineer would notice and fix. An under-specified objective plus a million iterations produces something confidently optimal and quietly wrong: optimal against the metric you named, pathological against the three you assumed. The search has no opinion about what you omitted. It can’t have one.
There’s a second cost, less discussed and in my view more serious for the industry’s long run. A human-designed circuit or structure carries a rationale. There were reasons, the reasons are recoverable and in five years, when someone must modify it, the rationale is the asset. A searched design carries a score. It’s defensible as an outcome and opaque as an argument. If you can’t reconstruct why a part is shaped the way it is, you can’t safely change it, review it or certify it. Which is the problem of the fourth part of this series arriving at the design stage rather than the operating one.
For a startup, the honest read is that the optimizer layer is closed. Synopsys, Cadence and Siemens own the tools, the flows, the foundry relationships and the correlation data, and they’re consolidating rather than fragmenting.
What’s open is everything either side of the optimizer. Specification tooling: capturing intent, constraints and trade-offs in a form a search can consume and a human can audit. Verification and correlation: The whole field has a proof gap, and a company that made AI design claims testable would be selling to every vendor and every customer at once. Public benchmarks are, genuinely, a business. The Alphachip saga is five years of evidence for how much their absence costs.
The broader opportunity is in physical products generally. If the cost of trying a design variant collapses, the economics of hardware startups change materially, because the expensive part was never manufacturing a prototype so much as being confident enough to commit to one. That advantage is real wherever the objective is crisply statable, such as for a bracket, heat sink or lattice, and thin wherever it isn’t.
The EDA incumbents are in a stronger position than the disruption narrative suggests, and their vulnerability isn’t where it’s usually placed. They’re not going to be displaced by a startup with a better placer. But if AI-driven design depends on pre-training across many prior designs, then the differentiating asset is a corpus of proprietary designs and that corpus belongs to their customers, not to them. Every vendor is therefore negotiating for rights to data that the chip companies have excellent reasons to protect, and whoever resolves that tension first gets a compounding advantage that’s very hard to copy.
The second vulnerability is the proof gap itself. In a market where every vendor’s AI claims use private definitions and private data, the first one to publish correlated production results on common public benchmarks would reset the terms of competition. It’s a strange situation: The strongest available competitive move is to make your own claims falsifiable.
For society, cheap design cuts two ways at once, and both are real. The democratizing case is strong. Open flows like Openroad lower the floor for university groups, small teams and countries without a domestic EDA license budget, and the ability to iterate a physical product without a large engineering organization genuinely widens who gets to build. That matters, and it’s the first unambiguously good news in several posts.
The concentrating case is just as strong and operates through data rather than tools. If the quality of AI-assisted design scales with how many prior designs you trained on, then advantage accrues to whoever has already designed the most things. The tools get cheaper and more available; the input that makes them work gets scarcer and more private. We’ve seen that pattern before in this series, and it doesn’t resolve itself.
And there’s a civic version of the rationale problem. Our mechanisms for holding engineering accountable, such as safety certification, failure investigation and professional liability, all assume a design can be explained. Not reproduced, explained. A world in which the honest answer to “Why is it like that?” is “The search preferred it” is a world where those mechanisms don’t have anything to grip.
Which is why I want to end on the oldest warning in this field, written before anyone could have known how specific it would become. Norbert Wiener, in a 1960 Science paper on the moral and technical consequences of automation: “We had better be quite sure that the purpose put into the machine is the purpose which we really desire.”
Sixty-six years later, that sentence has become a work instruction. The loop of the substrate designing its own successor is genuinely powerful, and it’s also strictly bounded by the objective we hand it. A self-improving design loop doesn’t get better at things we didn’t think to ask for. It gets relentlessly, impressively better at what we specified, which means the ceiling on the whole enterprise is the quality of our specifications.
So, the uncomfortable question for anyone running a design organization is not whether to adopt these tools. It’s this: Who in your company writes the objective function, how senior are they and when was the last time anyone reviewed it as carefully as they would review the design it now determines?
