What Breaks, What Validates, What Automates: What I Heard at DAC 2026

“Exciting times. Let’s talk next year.” That was how one audience member closed out his exchange with Vinci co-founder and CEO, Hardik Kabaria, after pressing him on the two hardest questions in this space — whether you can ever let AI be your calculator, and how physics AI works when the most valuable design data in the industry sits in a safe with three keyholders. When the skeptical questions end in “let’s talk next year,” something has shifted.

At DAC 2026 in Long Beach, Hardik shared a stage with Marc Halpern, VP and analyst at Gartner, and Adrian Lew, Stanford professor of mechanical engineering and Vinci technical advisor, moderated by Semiconductor Engineering editor-in-chief Ed Sperling. The session put three hard questions right in its title: Deterministic, Solver-Accurate Physics Intelligence for Chips-to-Systems Design: What Breaks, What Validates, What Automates. I sat in that room for the full session, and I want to set down what I heard — because the conversation was candid, technically serious, and willing to take the hard questions head-on.

Ed put a number on the table early, citing Harry Foster’s most recent study at Siemens: first-pass silicon success now sits at 14%. Massive simulation and effectively unlimited cloud compute were supposed to prevent exactly that. So Ed opened with the right question: what went wrong?

What breaks: the industry fixed the wrong constraint

Hardik’s answer was direct. “The algorithms became faster, the compute is easily accessible, but the human bottleneck that really puts the whole physics into the loop — we haven’t really addressed it. We are still requiring humans to perform manual work before the simulation even begins.”

That sentence carries more weight than it may appear to. The industry spent two decades accelerating solvers and scaling compute, and the constraint did not move — because the constraint was never inside the solver. It is the specialist-gated, manual workflow wrapped around it. I wrote about this in The Physics Bottleneck in Modern Engineering: hardware development is no longer limited by what teams can design, but by how quickly physical systems can be modeled, validated, and iterated.

Marc Halpern has spent decades in simulation, across civil and mechanical engineering before covering chips-to-systems design at Gartner, and he described where the bottleneck leads. When simulation takes too long, “the designers are just going to make their best choices and move forward — and unfortunately, sometimes it’s a little too late when they’ve made bad decisions, and you have to make corrections downstream, where it’s a lot more expensive.” His summary was the sharpest line of the afternoon: slow simulation forces design without truth.

Hardik described the same failure from inside the workflow — the over-the-wall pattern I examined in The End of Physics Silos in Engineering AI. Someone creates a design, hands it to a specialist with a different tool, and waits. “Forget daily — it’s barely weekly cycles.” Meanwhile, the industry is racing toward agentic design, where ideas are generated far faster than any specialist-gated process can evaluate them. Hardik stated the consequence plainly: “If we cannot evaluate design ideas at the same speed we are going to create them, we are not really going to be able to choose which is the better design.”

What validates: "if it fails nicely, you wouldn't tell"

The middle of the panel produced the exchange I keep thinking about, and it deserves to be quoted at length.

Halpern brought the discipline the question needs: AI in design works best when designs are largely repeatable, and subtle changes can send AI in the wrong direction — so teams have to be discerning about where they rely on it, with a capable human in the loop. That caution set up the deeper question: what happens when even the capable human can’t tell?

Lew took the point a level deeper, from thirty years of teaching exactly this. “I always tell my students: if a simulation fails catastrophically, you know it’s wrong. But if it fails nicely, you wouldn’t tell. This idea of the human being able to detect that the AI solution is wrong is not real. I myself have believed a lot of results that in the end were wrong, just because they looked good.”

Read that again. The plausible-but-wrong answer is precisely the one that even an expert reviewer will not catch — which means expert review alone cannot be the validation strategy.

So validation has to be built into the system itself, and that is the bar Hardik proposed. Physics has equations, so a system can verify that its answer actually satisfies them — “and if it doesn’t, be humble and say: I couldn’t solve this problem.” The same standard applies to repeatability. Same inputs, same outputs, run after run. “The moment we detect non-determinism, we start reducing our trust in the system. And the moment there is lack of trust, we are back to square one. No matter how good the system is, it’s not going to get plugged into the design workflow.” It was a measure of the panel’s quality that this landed as consensus — Halpern agreed that guardrails built into the system, surfaced for humans to inspect and override, are how trust gets earned over time.

Lew supplied the working definition the panel converged on: whatever the system produces has to satisfy the partial differential equations that define the model, match analytical solutions where they exist, and show error decreasing at known rates as resolution increases. “Whatever system comes next cannot leave that certainty aside — because otherwise, we don’t know what we get.”

In hardware, plausible is not enough. That is not a slogan. It is the adoption criterion.

What automates: from 10 ideas a day to 10,000

No one on stage argued for removing engineers. Hardik drew the parallel to what large language models did for software: “Some software engineers have become super-duper software engineers — they’re not doing less, they’re doing a lot more of it. The goal of continuous physics is that a specialist won’t be evaluating 10 ideas a day; they might evaluate 10,000. And on the other end, everybody becomes an okay analyst for a physics problem.”

The part of the automation argument I find most underappreciated is that speed buys back fidelity. If automated analysis returns in seconds what used to take two hours of manual setup and simplification, you can spend that dividend in the other direction. As Hardik put it: “Now instead of two hours, let’s run an analysis for seven days — but resolve every via in your high-bandwidth memory stack.” No simplification, no homogenization, no discarding the geometry that actually determines where the design fails. This is how Vinci’s platform is being used today for thermal and thermo-mechanical analysis across full chip–package–board assemblies.

Lew saw the same future from the research side: simulation speed has reached the point where massive design-of-experiments becomes practical at near-perfect resolution. “No more response surfaces. We can learn a lot more about these complex systems right now. Somebody’s going to do that,” he said — then added, “maybe Vinci.” I will take that.

On the data question — foundry and fabless design data locked in the safe — Hardik was unambiguous that physics AI cannot be a fine-tuned language model fed proprietary designs. The right analogy is weather models: systems that predict physical fields directly, verified against the governing equations, with customer-specific material data supplied at run time so it never leaves the firewall. That architecture — no customer-data training required — is foundational to how we think about a foundation model for physics.

The honest caveats

One of the things I appreciated most about this panel is that nobody oversold. Halpern reminded the room that every simulation approximates true physics, “and when we forget that is when we make mistakes” — his cautionary tale involved a North Sea oil platform lost to a misunderstood finite-element model. And when Ed asked whether this works for all physics today, across all scales, Hardik answered with one word: “No.” Then the condition: “But if we don’t lower the bar — deterministic, able to self-verify, and able to say ‘I didn’t solve this problem well, don’t trust that’ — we will see much broader adoption for physics.”

That is the version of this technology worth taking seriously. Not a black box that answers everything, but a system that predicts, checks, verifies, and tells you when it is outside its lane. In hardware, that honesty is the product. Software gets patches. As Hardik reminded the room: in hardware, there are no patches in production.