1000 Designs a Day: Neural Concept's Thomas von Tschammer on AI-Native Engineering
Summary
- Neural Concept’s near-term wedge is a step-change in iteration economics: learned physics predictions arrive in minutes instead of days, expanding the feasible search from dozens of designs to thousands. CAD and finite-element analysis already lifted automotive development from roughly 5–10 physical prototypes to 50–100 simulated designs a year; AI adds another order-of-magnitude shift. Thomas von Tschammer stresses, “We are not fully replacing numerical simulation”: expensive solvers and physical prototypes move later, validating the most promising candidates.
- Jaguar Land Rover offers the clearest production proof, increasing external-aerodynamics evaluations from 50 to 1,500 designs per day. Other Neural Concept customer work has shortened battery-cooling-component development by 80%; separately, Thomas cites examples of designs that cool 20% better and weigh 15% less. Faster iteration is therefore not merely a labor-saving story—it can generate performance improvements that determine which supplier wins a program.
- The emerging stack combines frontier reasoning models, company-specific engineering knowledge, CAD control, numerical solvers, and specialist physics models. A plain LLM cannot accurately solve external fluid dynamics, but an agent “equipped with the right tools” can edit geometry, call high-fidelity simulation, and use Neural Concept’s fast predictors inside an automated loop. Fine-tuning on company-specific simulation and test data turns proprietary information into accumulated know-how: “Think about data as knowledge.”
- Nathan Labenz pushes Thomas beyond the comfortable copilot framing, and Thomas concedes that every individual step from specification to design and simulation “can be automated and should be.” His remaining boundary is the full vehicle’s enormous trade-off surface—safety, aerodynamics, thermal performance, manufacturing, cost, and other coupled constraints—where engineers retain final judgment. As agents become a potential unit of work alongside or instead of seats, Thomas expects pricing to move “purely on the value” delivered.
- AI adoption could create “exponential gaps” between manufacturers because Western product cycles are already dramatically slower than China’s. Thomas puts a new vehicle at 48–60 months for a US or Western European OEM versus 18–24 months in China, with much of China’s advantage coming from agile, highly automated manufacturing and less process legacy. Digital-native hardware companies adopt new workflows faster because they can choose “what’s best out there today and not what was best yesterday.”
- Formula 1 is a high-pressure demonstration of making scarce compute more productive and exposing designs humans would reject. Teams face ranking-adjusted CPU-hour caps for aerodynamic simulations, yet AI workflows can generate and evaluate 100,000 configurations overnight. Engineers sometimes discover that an apparently bad design is better than anything they devised themselves—“thought it was a blunder”—while remaining grounded in physics.
- Thomas’s achievable roadmap is 20%–40% acceleration in selected disciplines during year one, followed by 50%–60% reductions in iteration cycles as AI breaks silos across aerodynamics, crash, thermal, and manufacturing. Specific multidisciplinary workflows are already showing roughly 60% speedups, although no OEM has applied the approach across an entire car at scale. He sees infrastructure, governance, and data flow—not a missing fundamental AI capability—as the chief barrier to broader reinforcement-learning-style product loops: AI-first engineering is “already happening,” and “the acceleration is just starting.”
Deep dive
1. Engineering’s first digital leap created the bottleneck AI now attacks
Nathan grounds the transition in his father’s GM career: hand-drawn plans and slide rules gave way to CAD, but engineers still manually changed geometry, submitted it to a solver, and waited before trying the next idea.
Thomas’s automotive history begins with building and crashing physical prototypes. That expense limited a program to perhaps 5–10 candidate designs a year; CAD and finite-element analysis raised the range to roughly 50–100 by replacing many physical crashes with numerical ones.
The remaining constraint is computation. A single high-fidelity crash simulation can occupy large clusters for one or two days, and similarly expensive solvers govern aerodynamics, thermal management, electromagnetism, and structural dynamics.
AI changes the cadence from “results in days” to results in minutes. The consequence is not simply doing the old workflow faster: engineers can examine thousands of options, search a much richer design space, and reserve expensive validation for finalists.
2. Physics-aware models learn company know-how without retiring solvers
A car requires distinct physics across external aerodynamics and EV range, passenger and pedestrian safety, battery and engine cooling, cabin ventilation, electric-motor electromagnetism, chassis durability, and road-induced structural dynamics. Large OEMs employ thousands of specialists across those components and subassemblies.
Models can learn from numerical simulations, physical tests, or hybrids of both. Where traditional solvers poorly capture a phenomenon, wind-tunnel or test-loop measurements provide experimental data; low-fidelity simulation can still be combined with those measurements in hybrid training.
Thomas rejects a clean replacement narrative: “We are not fully replacing numerical simulation,” just as simulation never eliminated prototypes. Current models are fine-tuned on each company’s data and requirements, then retrained as new results arrive—capturing institutional knowledge so the next development cycle starts from a stronger base.
3. JLR moved external-aero throughput from 50 to 1,500 designs a day
Jaguar Land Rover’s published Neural Concept workflow had already parallelized and optimized conventional aerodynamic simulation to evaluate about 50 designs daily. Its production AI workflow raised that to 1,500, accelerating the negotiation between studio aesthetics and aerodynamicists optimizing EV range.
A supplier designing battery-cooling components reduced its development cycle by 80%. Thomas separately cites examples of solutions with 20% better cooling and 15% less weight. He cautiously suggested that aerodynamic exploration might likewise yield 2%, 3%, or 5% improvements, which could be “game-changing.”
The mechanism is breadth, not magic: more candidates increase the chance of finding performance combinations outside an engineer’s intuition. The human shifts from manually proposing each geometry to navigating a model-generated “space of solutions.”
4. The copilot is a tool-using system, not a plain LLM
Today’s workflow is “not a black box” that accepts a specification and returns the best possible car. The system can interpret requirements and set up a base model, but engineers validate intermediate steps, guide exploration, and decide among competing objectives.
Automation closes the iteration loop: an agent can operate CAD, generate geometry, send candidates to a numerical solver, call a faster specialist predictor, and return the results without an engineer repeatedly preparing each tool by hand.
Thomas adopts Jensen Huang’s “five-layer cake” framing: general models supply reasoning, while the application layer adds engineering context and skills—what valid geometry means, how injection-molding rules constrain a part, and which physical or manufacturing objective should be optimized.
Neural Concept’s original 2019 architecture was not an LLM but a computer-vision-derived model that ingested 3D geometry and predicted aerodynamics, deformation, or temperature. Those physics-aware models now become tools inside the broader agent rather than competitors to general-purpose reasoners.
5. Every task may be automatable even when the whole car is not a black box
Nathan suggests engineering specifications may be firmer than many software requirements, but Thomas says RFQs and specifications are still not fully streamlined or defined. Changing one component’s thickness can propagate through a vehicle’s interconnected constraints, so the problem remains less deterministic than it appears.
Thomas initially says AI will remove low-value setup, simulation, and drafting work without replacing engineers. Nathan’s pushback—worth keeping—is that models can already ideate, decompose systems, create 3D representations, and run solvers; why could a “little society of Fables” not produce the next model-year car?
Thomas’s honest concession is that all those individual tasks “can be automated and should be.” His reservation concerns combining every vehicle dimension into one black box: engineers should retain final trade-offs because their embedded domain expertise and choices are also how one OEM differentiates itself from another.
6. Value pricing follows agents, while adoption speed widens corporate gaps
When value is no longer attached to an individual assisted seat but potentially to autonomous agents, Thomas expects seat-based pricing may give way. The destination is pricing based on delivered value, with the industry and application determining how that value is monitored.
He forecasts “exponential gaps” between companies that adopt AI-driven engineering and those that do not. Legacy OEMs must change teams, tools, governance, and habits that have remained broadly stable for 20 years—an organizational transformation, not a software installation.
Digital-native hardware companies, often with no more than a decade of history, can scale new workflows across applications within a year. Thomas sees their lack of entrenched process as an opening for faster new entrants, while older organizations vary widely in adoption speed.
7. China’s 18–24-month cycle exposes the West’s 48–60-month handicap
Thomas estimates that a new car takes 48–60 months from launch decision to manufacturing at a US or Western European OEM, compared with 18–24 months in China. Nathan calls the resulting speed gap “sobering.”
Thomas believes more of China’s advantage lies between completed design and factory rollout. Plants are highly automated and operated more agilely; European and US engineering executives are already visiting China, studying those systems, and taking practices home.
China also benefits on the design side by carrying less legacy. Companies can take more risk and select “what’s best out there today and not what was best yesterday,” rather than fitting AI into processes and tool choices inherited from the 2000s.
Optimization must nevertheless include manufacturability from the beginning. The lesson from additive manufacturing’s promise 15 years ago is that geometric freedom does not guarantee economical scale; stamping limitations, design rules, physics, and cost must all constrain the AI’s search.
8. Aerodynamics is the first likely engineering foundation model
Nathan asks when per-company models become broader foundation models, invoking the multimodal integration associated with “Nano Banana Omni.” Thomas calls general aerodynamics the “low-hanging fruit”: its physics are complex but comparatively transferable across companies, so he expects such a model quickly.
Neural Concept is researching that direction but “may not be the first” to build the foundation model. Its intended position is the application layer that makes any such model usable inside a 100,000-person engineering organization—integrating domain-specific workflows, visualization, geometry editing, tools, and other capabilities.
9. Formula 1 turns constrained compute into an AI stress test
Formula 1 teams face explicit CPU-hour limits for external-aerodynamics simulation. The allowance varies with the prior year’s ranking, so stronger teams receive less compute—preventing the contest from collapsing into an arms race won by the largest budget.
That rule makes prediction efficiency directly competitive: more useful evaluations within the allowance mean more geometries explored and a better chance of improving the car before the next race.
Thomas describes F1 engineering as the industry’s most agile form. Cars change in fine detail every week, forcing extreme automation; Neural Concept invites teams to “push the limits” and break its workflows because failures reveal what must improve before broader OEM deployment.
A race profile becomes engineering requirements—more straight-line performance, for example—then an overnight system explores and evaluates 100,000 configurations. In the morning, aerodynamicists inspect an interactive dashboard, examine trade-offs and geometries, and select what can advance for the next Saturday.
10. Move 37 designs have measurable commercial value
Engineers sometimes tell Neural Concept that they would have immediately scrapped an AI-proposed shape, only to find it outperforms every human candidate. The result can look like a blunder without violating physical law, forcing the engineer back to the dashboard to understand and revise an intuition.
For a component supplier, one additional OEM program can be worth millions or tens of millions of dollars. Cutting development from six months to three also leaves three months to improve manufacturing; a part that becomes 10% cheaper can help win several programs worth hundreds of millions.
At the OEM level, Thomas says a car developed from scratch typically costs about $1 billion. Moving a cycle from 48 months toward 24, or saving even 20% of development cost, makes the willingness-to-pay calculation “pretty quick.”
11. The two-year roadmap starts within silos, then breaks them
Thomas’s year-one target is to make every iteration AI-enabled inside flagship disciplines such as crash safety and dynamics and powertrain. An AI workflow coordinating the relevant tools could produce 20%–40% speedups without waiting for whole-vehicle autonomy.
In year two and beyond, the objective is multidisciplinary optimization: aerodynamics changes should simultaneously account for crash, thermal, manufacturing, and ultimate design constraints. Breaking those team silos is where benefits compound toward 50%–60% shorter cycles.
Some focused multidisciplinary, automated workflows already show approximately 60% acceleration, but Thomas is explicit that no OEM has yet implemented this across an entire car at scale.
Adoption spans both extremes. “Engineers are artists in a way,” and some resist losing manual work they enjoy; once they use the system, however, waiting and setup give way to interactive “what if” exploration—the part many engineers find genuinely rewarding.
12. Physical abundance is gated more by governance than missing AI pieces
Thomas argues that autonomous driving may commoditize cars, yet AI can also preserve differentiation by encoding each manufacturer’s engineering know-how and product choices. New form factors should emerge as autonomy matures: Detroit companies currently adapt standard cars for autonomy, while Zoox’s autonomous-first robotaxi is deliberately “not a car.”
Asked whether an end-to-end reinforcement-learning loop could span virtual customers, strategy, design, and validation within five years, Thomas sees no fundamental capability gap. The pieces exist; large-company bottlenecks are infrastructure, governance, data understanding, and putting information in the right locations.
Numerical solvers remain the grounding mechanism because “you won’t break the physics.” Robots will matter in factories, but Thomas questions whether humanoids are the ideal form factor and highlights unresolved model performance, overheating, autonomy, dexterous hands, and hardware durability.
More general top-down intelligence can automate around those physical bottlenecks before general-purpose robots arrive. Thomas’s closing call is that AI-first products already exist across automotive, aerospace and defense, and consumer electronics: outsiders underestimate how close the transition is, and “the acceleration is just starting.”