Physics aware AI, trained on past simulation data to predict aerodynamics, crash safety, or thermal performance on new 3D shapes in seconds rather than days, is moving the bottleneck in product design from compute time to human judgment about which shapes are worth exploring.
When Jaguar Land Rover's aerodynamics team could only run about 50 design candidates a day through full-fidelity simulation, the engineer's hardest job was guessing which shapes were worth trying. When a physics-aware AI surrogate model collapsed that evaluation to seconds, the job changed: now the team has to decide which 1,500 shapes to generate. That inversion, from a bottleneck in compute time to a bottleneck in human attention, is what makes Neural Concept's pitch of "AI-native" product design worth taking seriously.
The mechanism is concrete. Numerical simulation remains the gold standard for predicting how a car body handles airflow, how a battery pack sheds heat, or how a chassis crumples in a crash. A single high-fidelity run can take days on a large compute cluster, so even well-staffed teams only iterate through tens to hundreds of designs per project. Neural Concept's approach, articulated by co-founder and U.S. managing director Thomas von Tschammer on The Cognitive Revolution podcast, is to train neural networks on past simulation and test data so they can predict performance on new 3D shapes in seconds. The company calls these models "physics-aware" because they learn from physics outputs, not just from rendered pixels or designer labels.
Von Tschammer frames this as the third engineering revolution: physical prototypes gave way to computer-aided design and finite-element / computational-fluid-dynamics simulation, and now simulation is being compressed by learned surrogates. His headline number, offered on the same podcast, is that physics-aware AI lets a team generate more than 1,000 early-stage design candidates per day. That figure is the founder's framing, not an independent benchmark, and the word "design" here is loose: a candidate is a shape the surrogate thinks is plausible, not a part a human has signed off on.
Two case studies anchor the claim with concrete before-and-after numbers. The first comes from a presentation by Jaguar Land Rover's aerodynamics team at NVIDIA's 2026 GTC conference (the chipmaker's annual AI-focused gathering). The team moved from roughly 50 aerodynamic design evaluations per day on full-fidelity simulation to about 1,500 per day with the surrogate, a 30x throughput jump. The second is the battery cool-plate case that von Tschammer also discusses: a supplier reportedly cut development cycles by about 80%, improved cooling performance by roughly 20%, and reduced battery weight by around 15% on the same kind of workflow.
These numbers are vendor-reported and not independently verified in the public sources so far. They are also consistent with what Neural Concept's other public customers say. GM and Nissan have both used the company's tools for early-stage car design, according to The Verge, and Formula 1 teams have leaned on the same aerodynamic AI for downforce and drag work, as TechCrunch reported in 2024. At CES 2026, Neural Concept introduced a physics- and geometry-aware AI Design Copilot that wraps this surrogate core into a tool the company says engineers can query conversationally. The firm, spun out of Switzerland's EPFL engineering university, separately announced a $100 million Series C round (the third major private funding stage), led by Goldman Sachs Growth Equity to scale what it calls "AI-native engineering."
The harder question is where the surrogate is wrong, and who notices. Surrogate models in computer-aided engineering have disappointed before: they interpolate cleanly inside the distribution of shapes they were trained on, and they quietly regress toward a mean shape outside it. A car body with an unusual drag-reducing feature, a battery pack with a nonstandard cell layout, or a chassis geometry outside the training envelope can elicit confident predictions that are physically nonsense. That gap, between evaluation cost (the surrogate's strength) and judgment cost (the engineer's new bottleneck), is where the real engineering problem now sits. The 30x speedup does not change how many hours a senior engineer spends asking whether a design is well-posed, and where full-fidelity simulation would still catch a failure the model cannot see. It changes what those hours are spent on.
Von Tschammer's analogy on the podcast is the cleanest framing of the trend. Computer-aided design did not kill physical prototypes; it pushed them later in the loop, into roles where they earn their cost. Physics-aware AI, on the same logic, does not retire full-fidelity simulation. It moves simulation later, more surgically, and asks engineers to spend their judgment where the surrogate is least trusted: at the boundaries of the training distribution, and on the candidates the surrogate ranks highly but a human reviewer has reason to doubt.
What to watch next is concrete. The vendors that survive this transition will be the ones who can show, with independently benchmarked numbers, that their surrogates do not just generate more shapes but also catch failures that engineers would otherwise miss. The companies that buy them will want to see how Neural Concept and its peers handle out-of-distribution geometry, how the new Design Copilot hands off to a full-fidelity solver when confidence is low, and whether the 30x and 80% numbers hold up outside carefully selected reference projects. Until then, "1,000 designs a day" is a real workflow change, but it is also a vendor's promise, not a settled engineering fact.