From 50 Designs a Day to 1,500: How AI Changed What Jaguar Land Rover Can Explore

Nathan Labenz opens from Detroit, where his father started at General Motors in a drafting department of pencils, big tables, and slide rules, then spent the rest of a career moving into CAD. By the time he retired, designs were routinely tested in physics-based digital simulations before anyone cut metal. Labenz has hosted more than 350 episodes of *The Cognitive Revolution* and admits AI for physical engineering has been a gap. His guest is Thomas von Tschammer, co-founder and US managing director of Neural Concept, a Swiss company that trains specialist models for aerodynamics, heat dissipation, and collision safety.

Thomas walks through three eras of car design. Forty years ago, crash safety meant building a prototype, driving it into a wall, and checking whether passengers and pedestrians survived; if not, you rebuilt. A team might explore five to ten prototypes a year. CAD and numerical simulation changed the scale: engineers could draw on a computer and run finite-element crash analysis instead of destroying hardware, reaching perhaps 50 to 100 designs a year. The new bottleneck was compute. A single crash simulation can still take a day or two on a large cluster, and Thomas says that wait has been the main limit on iteration for three decades.

AI is the third jump. Results that used to take days now come back in minutes, which means thousands of options instead of dozens a year. Labenz hears an echo of protein folding and other domains where models learn a kind of intuitive physics and knock a couple of orders of magnitude off the compute. The physics list on a modern car is long: aerodynamics (especially EV range), pedestrian and occupant crash, battery and cabin thermal management, electromagnetism in electric motors, and structural dynamics for the chassis. A GM-scale OEM employs thousands of domain experts iterating on those assemblies. Neural Concept’s job is to give each of them a faster loop.

The specialist models train on both simulation data and physical tests. Some phenomena still cannot be captured accurately by traditional solvers, so teams measure them in a wind tunnel or on a test bench and mix those measurements with numerical data—what Thomas calls hybrid training. Inputs are CAD geometries; outputs are predictions of drag, deformation, temperature, and similar quantities. The architecture Neural Concept started with in 2019 came out of computer vision, ingesting 3D geometry directly rather than routing everything through a language model.

Thomas is careful not to oversell replacement. The industry never fully dropped prototypes; it builds fewer of them, later. He expects the same pattern for numerical simulation: AI explores a richer space early, narrows the candidates, and high-fidelity solvers validate before anyone cuts a prototype. Aerodynamics is still too specific for one foundation model to solve every car off the shelf. Neural Concept is doing that research, but today every OEM still retrains on its own simulations and tests.

He tells Labenz to think about data as knowledge. The company does ship pretrained models for some applications, then fine-tunes them on each customer’s data so the model matches that company’s best practices and requirements. Every new data point triggers a retrain, covering a broader space and raising accuracy, which makes the next development cycle faster. Decades of trial and error become something a model can actually inherit.

The public metric Thomas leads with is Jaguar Land Rover. At NVIDIA GTC in March 2026, the UK OEM described its work with Neural Concept on external aerodynamics. A couple of years earlier, even after a highly parallelized, highly optimized solver pipeline, the studio and aero teams could evaluate about 50 designs a day—a daily trade-off between the best-looking car and the one with the most range. With AI they went to 1,500 designs every day, in production. Other suppliers designing battery cold plates cut development cycles by 80%, with examples of 20% better cooling and 15% less weight. On aero, engineers have found designs that are 2, 3, even 5% more aerodynamic.

Those numbers are not a black box that swallows a spec sheet and returns the perfect car. Thomas describes AI-driven workflows as an assistant: the model ingests requirements, sets up a base, and the engineer validates and steers along the way. Engineering is trade-offs among disciplines, cost, styling, and performance, so domain experts stay in the loop. The shift is from a veteran tweaking a point cloud by intuition to AI proposing a space of options the engineer can explore. In that loop the agent already talks to CAD, pushes geometries, and sends promising candidates to high-fidelity solvers automatically.

Labenz asks whether frontier “reasoners” will sit in the copilot seat, or whether Neural Concept will train specialist generators on the validator feedback that encodes company know-how. Thomas reaches for Jensen Huang’s five-layer cake: the application layer on top is the one that matters, and many firms now specialize in giving generic models domain skills. Neural Concept wraps state-of-the-art general models with tools that read and write 3D geometry, know what a millimeter means, and know injection-molding constraints. A plain video model will not solve fluid dynamics. The winning pattern, he says, is capable general reasoners equipped with the right tools, including the physics-aware predictors Neural Concept has built since 2019.

Labenz’s software-shaped intuition is that a clear spec is now enough for models to deliver, and that automotive RFQs must be even cleaner because you cannot patch a car in production. Thomas says the opposite is still true. OEM requests for quotation are not fully streamlined, automated, or defined. Standards exist, but human interpretation remains, especially on a car with an almost infinite set of coupled constraints. Change the thickness of one under-hood part and you may move the whole engine block. Starting from a spec and translating it into 3D is already happening; turning that into a software-style black box is not, because the problem is much higher-dimensional and there is never only one path to a solution.

Low-value work is already disappearing: manually setting up a simulation, manually rolling a new CAD design. Thomas does not think the engineer who makes the final design call is going away. Labenz pushes further: from product definition through subsystem split, 3D, and simulation setup, where is the step an agent society could not take for next year’s model? Thomas grants that those tasks can already be automated for specific products—battery cold plates, aerodynamics, crash. What he will not treat as a black box is stacking every dimension of a full vehicle. That is where engineers should keep the last trade-off, and where differentiation between OEM 1 and OEM 2 will live: whoever encodes their engineering know-how most deeply into the AI workflow.

Pricing will have to follow the same shift. Seat-based licenses made sense when value sat with an individual user; now value sits with agents. Thomas expects the industry to move toward pure value-based pricing, with the hard part being how each sector measures that value. He also expects a widening gap. Century-old OEMs have built cars the same way for twenty years, with the same teams and tools, so the ask is not only new software but a change in thinking and governance. Digital-native hardware companies—consumer electronics firms that have been building product for ten years at most—pick up the workflows much faster. In a year, he has seen them speed development across many applications, while larger, older engineering companies move more slowly.

The competitive clock is blunt. A Western European or US OEM still takes 48 to 60 months from deciding to launch a new car to manufacturing it. In China the same trip is 18 to 24 months. Everything those Western companies care about, Thomas says, is how to get from 60 to 24, because that is the next durable advantage. Labenz calls it a sobering deficit: iteration count in a given window is hard to overcome for long.

Labenz wonders whether super-optimized designs simply dump pain onto the factory—tighter tolerances, harder machining. Thomas’s answer is to optimize for manufacturing at the same time. Additive manufacturing promised freeform design because you could print anything; it did not scale on cost. If a plant stamps metal, the stamping rules have to be available to the model so every explored design is manufacturable at reasonable cost. Embedding those rules, and the cost physics, early in the search is a large part of Neural Concept’s work.

Asked to split China’s speed between design and plant ramp, Thomas sides with the plant: more agile about rolling out a new factory, and highly automated once it is up. European and US engineering executives are already traveling to China, taking notes, and bringing practices home. On design, the edge is less legacy. Teams without processes frozen in the 2000s can pick today’s best tools and take more risk—the same agility he sees in digital-native hardware companies. A general aerodynamics foundation model would be another unlock, letting a team jump from its historical sedan line toward something as different as a Cybertruck. He expects aero foundation models first, because the physics, while hard, is relatively consistent across components. Neural Concept may not ship the first one; its bet is still Jensen’s application layer—domain tools, visualization, geometry edits—so a 100,000-person engineering org can actually use whatever foundation model arrives.

Neural Concept also serves Formula 1 teams, which Labenz had not realized operate under an explicit compute cap. Teams are limited in the CPU hours they can spend on aerodynamic simulation, and the allotment depends on last year’s ranking: do better, get less compute next season. The point is to stop a pure budget-and-cluster arms race. Labenz analogizes to the NFL’s strength-of-schedule and the NBA’s draft, then notes F1 goes further—it changes the rules of *preparation* from race to race, not just who you play or who you can sign. For AI-governance listeners, it is a rare live example of reallocating compute to keep a contest competitive.

That constraint is why F1 is a proving ground. The car changes every week, in details spectators cannot see, which forces an extreme automation of design iteration. Thomas treats F1 as the state of the art every OEM is trying to approach, and as a stress test: if the workflow holds for a race team, it can be built for a traditional manufacturer. He wants those teams to push until something breaks, because the break is the product roadmap.

What does an F1 “token maxer” actually do? Look at next week’s circuit—more straights, more overtaking—and turn that into aerodynamic requirements. Overnight, an AI workflow that already knows the 3D geometry generates hundreds to tens of thousands of configurations, evaluates them, and lands an interactive dashboard on the aerodynamicist’s desk in the morning. The engineer inspects the cloud of points, inspects the shapes, and picks what goes to Saturday. That, Thomas says, is token-maxing in engineering: thousands of designs explored overnight, fully automated.

Labenz asks for the engineering analog of AlphaGo’s Move 37. Thomas says it is one of his favorite parts of the job. Engineers come in the next morning and report that the model proposed a design they would never have rated as good. Show it to them a few years earlier and they would have said to scrap it. The design outperforms anything they would have drawn, and now they have to go back to the dashboard to understand *why*. Physics is still the floor—AI will not emit something that breaks the equations—but it will walk outside human intuition and force the expert to update. Some of those moments felt like blunders until they were not.

On willingness to pay, Thomas keeps the math concrete. A supplier that sells battery cold plates to big OEMs and wins one extra program a year is looking at millions to tens of millions of dollars. Speed has a second path to cash: if a good design used to take six months and now takes three, the other three months go into manufacturing process, a part that is 10% cheaper, and then one, two, or three more programs worth tens to hundreds of millions. On the OEM side, developing a car from scratch is typically on the order of a billion dollars; even a 20% cut is large, quickly.

Pressed for a CEO-level milestone at a company like GM, Thomas splits the next two years. In year one, put crash safety, aerodynamics, and powertrain on AI-led iteration: a workflow that orchestrates the tools on every cycle. That alone can mean a 20, 30, or 40% speedup in those flagship disciplines. In year two, orchestrate *across* disciplines—break the walls between crash, aero, and thermal—so an agent optimizing aero is also holding safety and manufacturing constraints. That is where gains compound, and where he has already seen 50 to 60% cycle cuts on specific multidisciplinary workflows. No OEM has done it at full-company scale yet.

Engineers sit on a spectrum. Some are artists of intuition who have worked the same way for thirty years and reasonably resist a workflow that looks like it might flatten a GM car into a Ford or a BYD. Internal methods and machine-learning teams are at the other end: first to adopt, first to benchmark frontier models. The conversion Thomas keeps seeing is hands-on. Once veterans actually query the model, the resistance flips, because they stop waiting on setups and solver queues and get to the part they liked all along—interactive what-if exploration.

Labenz notes that a lot of cars already look alike. Thomas thinks autonomy will push that further: once driving is solved, the car becomes more of a commodity and form factors converge. The window *before* that is when winners encode brand know-how into the AI workflow. Most “autonomous” cars today are still driveable vehicles with autonomy bolted on—Waymo taking a Jaguar I-Pace and making it autonomous, in part because regulation still wants a human fallback. Zoox is the counterexample: built autonomous-first, marketed as a robotaxi designed around the passenger, and it does not look like a car. As the technology matures, he expects more of those native concepts.

Labenz sketches an end state where every seat in the company is an agent, virtual customer panels score products the way aero models score drag, and a complicated electromechanical device can be spoken into existence the way a video can. Thomas does not see a fundamental capability gap. The blockers at large companies are infrastructure, governance, and getting data flowing to the right places. The physics-grounded solvers stay in the loop as fuel for the models; they will not vanish. He does not think the vision is five years of missing science so much as adoption.

Humanoid robots, in his ranking, are second-order for plant speed. Robots will matter on the factory floor; whether they need to be humanoid is another question, because humans may not be the right form factor for a plant. Two breakthroughs are still missing for scale: AI models that can interact efficiently with the physical world, and hardware—heat, endurance, and a hand as agile and durable as ours. First-principles automation and general intelligence can remove a lot of bottlenecks before anyone waits on a general-purpose mechanic that can walk in and apply a wrench.

His closing note is that outsiders still underestimate how far this already is. Engineers are designing products AI-first today; the acceleration is just starting. Watch automotive, aerospace and defense, and consumer electronics for the next few years. The next breakthroughs in design, performance, and product, he says, will have an AI-driven workflow somewhere in them.