HWNEWS
Batteries, advanced chip packages, and airframes couple thermal, mechanical, and electrical physics in real time. Per physics AI models can't, and a few teams want to fix that.
In an advanced 2.5D chip package, thermal, mechanical, and electrical failures happen in the same cubic millimeter, on the same microsecond timescale, and they reinforce each other. The engineering AI trained to predict them is still organized the way CAE, or computer-aided engineering, was in the 1980s: one model per physics, no coupling, and no built-in way to check the answer against the real solver it claims to replace.
That inheritance is the next bottleneck in engineering AI, and it is not model quality. The per-domain surrogates in use today are genuinely fast. They are also deaf to the coupling that defines every modern engineered system.
Consider a 2.5D package, the kind that stacks high-bandwidth memory (HBM) on a logic die in today's AI accelerators. Current crowding in the through-silicon vias heats the silicon. That heat warps the package, which changes the via resistance, which changes the current. The warp also stresses the solder microbumps, which can crack. None of these are separate problems. They happen together, and a single-physics surrogate handles them one at a time. Stitch three surrogates together and you still get a stitched answer, not a coupled one.
Modern batteries and airframes are the same problem at larger scales. A battery cell's thermal runaway is electrochemical, electrical, and mechanical at once. An airframe's fatigue life is structural, thermal, and aerodynamic. The reason CAE ran in silos for forty years was that compute was scarce and each physics needed its own numerical method. That reason is gone. The silos survived anyway, and engineering AI is now rebuilding them in software.
A recent Semiconductor Engineering analysis argues the field is starting to move past this. Per-domain models are giving way to a single cross-domain physics intelligence layer that reasons across thermal, mechanical, electrical, and material physics in one pass, with outputs that are deterministic and can be checked against a real solver. The framing is borrowed directly from where machine learning has been for a decade: stop training hundreds of small specialists, train one foundation model.
The first concrete signal that the thesis is more than a slide deck is from Vinci, which emerged from stealth on December 2, 2025 positioning a "foundation model for physics." Vinci's product framing is a single model that sits above the solvers and returns coupled, deterministic answers at solver accuracy, and the company has published a definition and minimum-criteria paper on continuous physics reasoning and a deterministic thermal-and-warpage study on 2.5D HBM packaging as the first falsifiable evidence. The platform describes the same stack.
The HBM paper is the interesting artifact, not the slogan. It claims solver-accurate thermal and warpage answers at the manufacturing resolution that advanced packages actually need, which is the unit of work where the old single-physics surrogate breaks. A vendor blog can be a vendor blog. A paper with a defined mesh, a defined solver, and a defined comparison is testable. Whether the numbers hold up under independent reproduction is a separate question, and right now the only available reference set is Vinci's own.
That is also the skeptic's case, and it is not new. Multi-physics simulation has tried to break out of silos for thirty years. The history is full of "single environment" CAE platforms that promised to fuse thermal, structural, and electromagnetic solves into one coupled run and ended up too slow to use, too narrow to trust, or both. The reason was not laziness. The underlying numerical methods are different, the time scales are different, and validating a coupled result requires validating each component plus the coupling itself. A foundation model that promises all three at once is making a much bigger bet than a per-domain model, and the only honest measure is whether the outputs can be checked against a real solver, on a real geometry, at production resolution.
The risk in the current wave is not that the bet fails. It is that vendors ship plausible outputs in the meantime. The first high-profile failure, an AI-assisted design that passed review and failed in qualification, will set the field back more than a year. The second-order risk is consolidation. If a single cross-domain physics stack becomes the default, the dependency on the team that built it is total, because no one else has the bench data to reproduce it.
What to watch in the next eighteen months: independent reproduction of Vinci's 2.5D HBM results on a non-vendor geometry, the first named customer production deployment of any cross-domain physics model on a safety-critical program, and whether any major EDA or CAE incumbent acquires a foundation-model-for-physics startup rather than building one. The first two settle whether the architectural shift is real. The third settles who captures it.