Neural Concept: An EPFL Spin-out Betting on Neural Surrogates, with $138M Raised

Executive Summary

Neural Concept SA is a spin-out from EPFL's Computer Vision Laboratory that sells what it calls the intelligence layer for engineering. It does not sell a solver. It sells a neural surrogate trained on a customer's own CAD geometry and historical CFD/FEA results, plus the engineering workflow layer built around that surrogate, so that designers can evaluate design variants in milliseconds and spend solver time only where it is worth spending.

Four facts frame the investment case.

First, the technical pedigree is real and checkable, but the commercial model is a black box. Four papers are traceable: the ICML 2018 geodesic convolutional architecture (arXiv:1802.04016), DEBOSH for deep Bayesian shape optimization (arXiv:2109.13337), DeepGeo for differentiable geometric parameterization (arXiv:2305.02116), and ICML 2024 work on uncertainty estimation in iterative neural networks (arXiv:2403.16732). An official EPFL feature published in June 2026 confirms that an algorithm developed during CEO Pierre Baque's doctoral work was patented and licensed to Neural Concept, which is the most likely explanation for why early investors funded a CVLab spin-out. What is not public: the current production architecture beyond the name "Geometric Regressor", the composition of training sets, the weights, and the layer-by-line inheritance from the 2018 architecture.

Second, the capital position is strong and unusual for European industrial AI. On December 18, 2025 the company closed a $100M Series C led by Growth Equity at Goldman Sachs Alternatives, with Forestay Capital, Alven, HTGF, D. E. Shaw Ventures, Aster Capital and Anaïs Ventures participating. Together with a $27M Series B in June 2024, a $9.1M Series A in March 2022 and a $2M seed round in 2020, disclosed equity totals at least $136.1M, or $138.1M including the seed. Database totals range from $136M to $140M; the spread comes from grants, debt and rounding, not from disagreement about the venture rounds.

Third, every accuracy number is self-reported and none has been independently reproduced. On the MIT DrivAerNet++ benchmark the company reports surface pressure MSE of 3.98e-2 and drag R-squared of 0.978; on a synthetic Formula 1 dataset it reports surface pressure R-squared of 0.916; at JLR it reports daily design evaluations rising from about 50 to 1,500. MIT's Dr Faez Ahmed is quoted approvingly, but he is the dataset's author, not an independent auditor.

Fourth, customer disclosure is the weakest part of the story. The same coverage number has been rewritten repeatedly between 2022 and 2026: 40% of the largest European and Asian OEMs plus 25% of the world's top 100 Tier-1 suppliers in 2022, "70+ OEMs" in 2024, "70% of the world's largest OEMs and 40% of the top 100 tier-1 suppliers" in September 2025, and "50% of the world's leading OEMs and top-tier suppliers" on the 2026 homepage. Three separate customer-count conventions coexist alongside them: "50+ global companies", "60+ OEMs" and "70+ OEMs and Tier-1 teams".

Our view: technical credibility is higher than commercial credibility. The company has demonstrated, with papers and a public benchmark, that geometric deep learning can reach usable accuracy on industrial CFD data. That is a real result. It has not demonstrated three things: that accuracy transfers across solver toolchains, that it generalizes across geometry families, or what a single customer contract is worth. Contract values, ARR, gross margin and churn are all undisclosed. The valuation was set by growth equity rather than by delivered revenue, which is the single largest valuation risk attached to this asset.

Company Overview

Neural Concept SA is registered in Lausanne, Switzerland, at EPFL Innovation Park. It is legally a spin-out founded in 2018, with commercial operations beginning in 2019. This is not a contradiction but two different dates that appear in two different families of company documents: the technology and membership materials say 2018, while the financing and product boilerplate says 2019. The evidence chain is one-directional. Venturelab lists the company with a first contact point at Venture Kick 2018, HTGF announced it in spring 2018, four industrial pilot projects were completed inside EPFL between May and December 2018, and more than twenty official documents from 2022 to 2026 say "founded in 2018". No source has been found that denies a 2018 founding.

The founder is Pierre Baque, French national, PhD in computer science from EPFL in 2018 with a thesis on mean-field methods for structured deep learning in computer vision, supervised jointly by Pascal Fua and Francois Fleuret. He previously studied applied mathematics and operations research at Ecole Polytechnique in Paris and worked as an engineer at Credit Suisse in London building algorithmic trading tools. The claim that Barbara Marcora was his doctoral advisor appears nowhere in EPFL records or in his papers and is treated here as wrong.

The co-founder group has also drifted. The 2022 Series A release names Baque, Fleuret and Fua. The 2026 About page names Baque (CEO and Founder), Theophile Allard (CTO and Co-founder), Jonathan Donier (Chief Scientific Officer and Co-founder), Thomas von Tschammer (General Manager USA and Co-founder) and Philippe Cuendet (COO). Fleuret remains a co-founder and sits on the advisory board as a Meta research scientist and professor at the University of Geneva.

Two hires in July 2026 are worth reading closely because they define the competitive position. Brett Chouinard became Chairman of the Board after roughly three decades at Altair, latterly as COO, President, CTO and Chief Product and Strategy Officer. Pietro Cervellera became Chief Revenue Officer after senior commercial roles at Altair and Siemens Digital Industries Software. Both companies sell AI embedded in CAE, so Neural Concept has staffed its board and revenue leadership with people who know the incumbent's pricing and product roadmaps from the inside. That is a coherent commercial strategy and a disclosed conflict of interest at the same time.

Headcount is the number no one will settle. The official About page says "100+" in 2026; Startupticker said 65 in 2025; database aggregators cluster between 126 and 153 across 2026. The 2026 careers page shows 37 open positions across 11 offices, with the split going to Solutions Engineering (9), Forward Deployed Engineering (7), Sales (6), Product (5), Vertical Solutions (4), Marketing (3), Deployment (2) and Operations (1). The offices are Paris, Munich, Zurich, Lausanne as headquarters, Detroit, San Francisco, New York, Pune, Bengaluru, Seoul and Tokyo.

Technology

The 2018 architecture is the foundation

The original paper, *Geodesic Convolutional Shape Optimization* (arXiv:1802.04016, ICML 2018), is the clearest public statement of what the company actually does. It applies geodesic convolutions directly on 3D surface meshes, remeshing irregular triangular meshes into regular quadrilateral grids via polycube remeshing so the convolution becomes GPU-efficient. The network has a shared trunk feeding a local branch and a global branch, and it is differentiable with respect to vertex positions, which is what enables gradient-based shape optimization rather than only prediction.

The 2D airfoil results from the paper's own table: standard separate branch 68.77% pressure accuracy and 82.23% drag accuracy; dilated separate 74.42% and 84.06%; joint two branches 81.32% and 84.90%; joint four branches 82.03% and 86.01%. The paper claims a 5% to 20% improvement over the then state of the art, and it also contains the number that is most often dropped from secondary summaries: on 3D shapes, drag accuracy is 77.2% on validation, 51.5% on test and 70.3% on test with fine-tuning. That 51.5% is the original published value and it is the single most informative data point available on how hard generalization to unseen 3D geometry is. Any evaluation of this company's 3D capability should start there.

The optimization loop is discrete, not an in-the-loop differentiable solver

DEBOSH (arXiv:2109.13337) makes the structure of the loop explicit: run a physical simulation, train a graph neural network on the results, compute an acquisition function over unsimulated samples, optimize the geometry, add the new sample to the training set, run the simulation again. That is a discrete outer loop, not an online differentiable solver inside the training or inference path.

DeepGeo (arXiv:2305.02116, AIAA Aviation Forum 2023) adds a differentiable parameterization layer, with a Latent Space Model, a Direct Mapping Model and volumetric mesh deformation regularization, and the paper demonstrates that it can be plugged into either a surrogate model or an adjoint solver. It is a research framework, not a description of the shipping workflow.

This matters for how the technology should be categorized. No evidence has been found of a differentiable CFD or FEA solver running inside the loop, of physics-informed neural networks, or of hard physics constraints. The official material actually says the opposite: the platform is solver-agnostic, because the AI trains on simulation outputs, and "neither replaces the solver". The value proposition is stated as evaluating more designs before committing to one. "Physics-informed" in Neural Concept's marketing is a positioning statement about the training data, not a claim about the architecture.

Uncertainty is handled by convergence rate, not ensembles

The ICML 2024 paper (arXiv:2403.16732, PMLR 235) shows that the successive outputs of an iterative network converge at a rate that correlates strongly with final accuracy, so the convergence rate can serve as an uncertainty proxy. The method costs far less than deep ensembles and requires no modification to the model. It was applied to aerial road detection and to 2D and 3D aerodynamic property estimation. Jonathan Donier's affiliation on that paper is listed as Neural Concept SA, Lausanne, which makes it the clearest public link between the research lineage and the company.

What the production architecture actually is

The public technical description names the commercial model the Geometric Regressor; the earlier flagship was Neural Concept Shape (NCS). The current architecture is described as surface-oriented geodesic convolutions feeding two branches: a Global Feature Branch (geodesic convolutions, average pooling, two fully connected layers, predicting total aerodynamic loads, total drag and aerodynamic balance coefficients) and a Local Field Branch (additional geodesic convolutions plus pointwise operations, producing continuous pressure and velocity fields). It is parameterization-agnostic, meaning it can consume unparameterized CAD geometry directly.

PointNet, MeshGraphNet, U-Net-on-SDF and neural fields appear in the DrivAerNet++ comparison as baselines to beat, not as the production architecture. That distinction is easy to miss and is worth being explicit about.

Physics and CAD coverage

The official multi-physics list is external aerodynamics, thermal management, structural mechanics, electromagnetics, rotating machinery, fluid-structure interaction, injection molding and internal flows, mapped to application areas of aerodynamics, thermal, durability, e-motor, turbomachinery, crash, manufacturing and (for the last) internal flows. The CAD integration list is NX, CATIA, Solidworks, MotorCAD and CFX; the simulation tool list is StarCCM+, Fluent, Abaqus, Ansys Maxwell, Ansys Mechanical and Moldflow. The October 2026 MCP Bridge release adds connectors for CATIA and NX at the design stage, ANSA and HyperMesh at model preparation and Abaqus at simulation.

Performance Benchmarks

DrivAerNet++ (company-reported, September 2025)

The setup: 8,000 automotive geometries, 39 TB of CFD, geometry only with no parametric dataset inputs, trained on 4 NVIDIA A100 GPUs for 24 hours total with the best model emerging at 16 hours, deployable in real time on a single 16 GB GPU, and a full production pipeline from 39 TB of CFD in one week on default settings.

• Surface pressure: MSE 3.98e-2, MAE 1.08e-1, against GAOT's MSE 4.94e-2 and MAE 1.10e-1.
• Wall shear stress: MSE 7.80e-2, MAE 1.44e-1, against GAOT's MSE 8.74e-2 and MAE 1.57e-1.
• Volumetric velocity: MSE 3.11, MAE 9.22e-1, against TripNet's MSE 6.71 and MAE 1.52e0.
• Drag: MSE 0.8e-5, MAE 2.22e-3, max absolute error 1.13e-2, R-squared 0.978, against TripNet's MSE 9.1e-5 and MAE 7.19e-3.

The volumetric velocity MSE is more than halved against TripNet. That is the strongest single technical claim the company has made. The caveat is that the benchmark is self-run, and the corroborating quote comes from MIT's Dr Faez Ahmed, who built the dataset.

Synthetic Formula 1 dataset (company-reported)

1,644 unique Onshape geometries exported as STL with up to 56 shape variables, simulated in OpenFOAM at 250 km/h with a half-car symmetry model and SST k-omega RANS turbulence. Surface pressure R-squared 0.916 all-point mixed, and 0.967 on a section 1.7 m behind the wheel. Inference in under 0.1 seconds. The stated workflow change is from five CFD runs per week to several hundred evaluations.

JLR (company-reported, GTC 2026)

More than 20,000 simulations run in total, up to one billion data points in a single case, and daily design evaluations rising from about 50 to 1,500, aligned against CFD and wind-tunnel validation. The same 50-to-1,500 figure appears in Microsoft's customer story, which is at least a second published appearance of the number.

Company-reported business outcomes

Development cycles shortened by up to 75%; simulation accelerated by up to 10x; product performance improved by up to 30% on efficiency, safety, speed and aerodynamics; late-stage redesigns cut by 30% to 50%; time to market shortened by up to 2 years; $50M saved per year for customers. On General Motors pedestrian crash work, prediction time falls from 48 hours to 30 seconds, a 5,760x factor, and 10,000-plus variants become explorable against roughly 100 before. On an aerospace pressure field, one hour becomes 30 milliseconds, with some workflows at 200x. On an e-motor customer, mass falls 12%, torque rises 9% and development cycles fall 75%. On MAHLE, 30 million-plus virtual designs produced a 15% efficiency gain and a 4 dB noise reduction.

These acceleration factors are not comparable to one another. They are different cases with different baselines: a 1-hour CFD run, a 48-hour crash model, a full workflow including setup. The company itself flags this in a September 2025 article by CEO Baque in which he argues explicitly against the benchmark-first evaluation of surrogates, describing his own DrivAerNet++ numbers as "flattering" and calling for evaluation backwards from production value.

Product and Commercial Model

The product naming has been rewritten several times. NCS (Neural Concept Shape) was the early flagship, running in co-pilot mode where an engineer uploads a shape and the model proposes improvement paths. Engineering Intelligence became the primary positioning from 2024. Intelligence platform and Engineering Intelligence Layer are the current platform-level terms. Design Lab is the interactive exploration layer, Studio is where copilots are trained and deployed, and CodeLake is the trusted engineering knowledge base that supports retrieval for the platform's assistant.

Two 2026 launches define the current product direction. The AI Design Copilot, announced January 7, 2026 at CES, was the first enterprise-scale function combining spatial reasoning, physics awareness and CAD-ready geometry generation; it compresses weeks of manual modelling into minutes, lets engineers investigate 10 to 1,000 times more variants per iteration and millions in total, cuts manual workload by up to 90% and late-stage redesigns by up to 50%. The MCP Bridge, announced October 19, 2026, is built on the open Model Context Protocol and lets agents operate CATIA, NX, ANSA, HyperMesh and Abaqus across Windows workstations and Linux clusters. It benchmarks models on real engineering tasks so each step runs on the best-suited model. The quantified case is a wheel rim geometry cleanup in roughly three minutes against about a day manually, with TRATON and OPmobility named as users.

Commercial model facts: quote-based pricing with no public price list; on-premise, air-gapped on-premise, private cloud and managed SaaS deployment on Azure, AWS and GCP; Microsoft AD and SSO integration; two license tiers introduced in 2023, Expert for building and fine-tuning models and Production for consuming their outputs, with a stated go-live of under 24 hours in four steps. TechCrunch reported F1 customer annual fees in the range of EUR 100,000 to EUR 1,000,000 depending on team size and access level, against an F1 team cost cap of $135M per year. The company states that "the intelligence our customers build on the platform is theirs" and that its ARR growth significantly outpaces its burn rate.

Named customers

Confirmed with first-party or company-level evidence: General Motors (pedestrian crash, 48 hours to 30 seconds, 10,000-plus variants, models trained on data from 11 vehicle programs), Safran, Airbus (long-running relationship renewed in 2021), Bosch, Subaru (stamping and die-surface design, 3 hours to 2 minutes against a traditional 48 hours), Eaton (cooling modules, 30%-plus pressure-drop improvement and 10%-plus weight reduction), MAHLE (30 million-plus virtual designs, 15% efficiency gain, 4 dB noise reduction), Williams Racing, Visa Cash App Racing Bulls, Leonardo (aeroacoustics since 2023, roughly 10 engineers trained across three divisions, running on A100 GPUs and Omniverse libraries on Leonardo's own davinci-1 HPC cluster), JLR, OPmobility, TRATON, Renault Group, GE Vernova and General Electric, Plastic-Omnium, Mubea, Antolin, Hanwha Ocean, Sumitomo Wiring Systems, SPAL Automotive, Kautex, Multi-Wing, Cooper Standard (specialist knowledge codified into the design copilot, cutting custom tooling design time for moulded parts by 50%), DLR, LS Electric (99%-plus accuracy on injection moulding temperature and pressure), Volvo Cars, RWDI, Nexans and Cybernet.

Three names in circulation could not be verified: Yutong, Slam and GE Renewable Energy as a precise official customer name. Porsche AG appears in a single aggregator listing only.

Competitive Landscape

The direct AI-native competitors are PhysicsX (London, $300M Series C in June 2026 at about a $2.4B valuation, with NVIDIA, Applied Materials and Siemens following on, and roughly $50M of 2026 revenue), Luminary Cloud ($72M Series B in September 2025 led by N47 with NVentures participating, shipping SHIFT pretrained models), Emmi AI (acquired by Mistral AI in May 2026, terms undisclosed), Monolith AI (announced as acquired by CoreWeave on October 6, 2025 and already treated as acquired in CoreWeave's Q4 2025 filing on February 25, 2026) and SolverX, a Seoul company founded in 2021 whose AI Solver, AI Optimizer and AI Enhancer products occupy the same CAE space and which PitchBook lists Neural Concept as one of four competitors.

The more serious competition is inside the incumbents. Synopsys completed its acquisition of Ansys on July 17, 2025, giving the combined entity a stated addressable market of $31B and a portfolio that includes Ansys Mechanical, Maxwell and Moldflow, all of which Neural Concept integrates. Siemens embeds AI in Simcenter and BeyondMath while simultaneously being a Neural Concept partner through the STAR-CCM+ and NX integration announced in June 2024. Dassault Systemes embeds AI in SIMULIA and 3DEXPERIENCE. Altair ships PhysicsAI while having lost two of its senior executives to Neural Concept.

The structural point is that Neural Concept is not competing with the incumbents on their own terms. It is a neutral layer that sits above them. That is a real distribution advantage, and it is also why the incumbent's most natural response is to copy the layer rather than to lose the customer.

Funding and Team

• Seed, 2020, $2.0M, lead undisclosed, participants undisclosed.
• Series A, March 2022, $9.1M, led by Alven, with Aster, CNB Capital and HTGF.
• Series B, June 2024, $27M, led by Forestay Capital, with D. E. Shaw, Alven, CNB, HTGF and Aster.
• Series C, December 18, 2025, $100M, led by Growth Equity at Goldman Sachs Alternatives, with Forestay, Alven, HTGF, D. E. Shaw Ventures, Aster and Anaïs Ventures.

Total disclosed equity: at least $136.1M, or $138.1M including the $2M seed. Grants, the CHF 500K FIT debt facility and ESA BIC support are excluded from that figure. Anaïs Ventures confirmed its co-investment alongside Forestay on its own site in January 2026.

No valuation has ever been disclosed. Dealroom's funding history lists $400M while its own profile band lists $500M to $1B, and secondary reporting says "over $500 million". All valuation figures are database estimates or media reports, and the largest aggregator contradicts itself.

NVIDIA has not invested in Neural Concept. No evidence has been found of any NVIDIA or NVentures equity stake. NVIDIA appears consistently and only as a technical and ecosystem partner: accelerated computing, GPU and CUDA optimization, the Omniverse Blueprint for real-time digital twins at Supercomputing 24, the Omniverse Blueprint at CES 2025, a joint General Motors demonstration at GTC 2025, the JLR session at GTC 2026, and a NemoClaw combination for agentic engineering in May 2026. Secondary headlines that read "with Nvidia, Siemens and Microsoft as partners" are the source of most of the confusion.

Regional expansion has been the other 2026 theme: Seoul as the first direct Asia-Pacific office in June 2026, Bengaluru and Pune as the first dedicated India footprint in September 2026, and Cambridge as the first UK base in September 2026. The company received the World Economic Forum Technology Pioneer designation in 2025, one of only two Swiss companies that year, achieved ISO 27001 certification in June 2024, and was named in the McKinsey and Boardwave report on Europe's application-layer AI opportunity in September 2026.

Discrepancies

Twenty conflicts were identified and adjudicated. The most consequential:

1. Founding year, 2018 versus 2019. Resolved as legal founding and spin-out in 2018, commercial operations from 2019. The two dates appear in different document families rather than describing two entities.

2. Total funding. At least $136.1M, or $138.1M with the seed. The $136M to $140M spread across databases and media reflects grant, debt and rounding treatment, not disagreement about the venture rounds.

3. Series C date. December 18, 2025. The January 29 and February 9, 2026 dates in secondary write-ups are transcription errors.

4. Valuation. Not adjudicated, only labelled. No official figure exists and Dealroom contradicts itself.

5. Customer percentages, 50% versus 70% versus 70+ versus 40% versus 25%. The "70+ customers" and "70% of the world's largest OEMs" statements have different denominators and can both be true. The most stable and best-defined pair is 40% and 25%, which recurs from 2022 to 2026. The 50% on the 2026 homepage uses a broader combined denominator and conflicts directionally with the 70% used in the 2025 boilerplate. The move from 25% to 40% on the same top-100 Tier-1 denominator is a change in the company's own claim. The count conventions moving from 60+ to 70+ to 50+ reflect different definitions and should not be read as customer loss. A secondary claim of roughly 15 customers contradicts every official statement and is rejected.

6. Headcount. 65 in 2025 according to media, 100+ on the official About page in 2026, 126 to 153 across databases. No single figure is reliable; the official discontinuity most likely reflects a change in whether contractors, consultants and interns are counted.

7. NVIDIA investment. None found. NVIDIA has only ever been a technology and ecosystem partner.

8. SP80 record numbers, 65.45 versus 68.01 knots. The 68.01 knots figure is the boat's peak speed; the certified world record is 65.45 knots over 500 m. Not a contradiction, but SP80's own text is easy to misread. As of the cutoff date, no result from the 2026 Walvis Bay campaign has been published, which is not the same as a failure.

9. Pervis speed record. The 2025 result of 141.73 km/h, or 88.07 mph, is the latest valid record and also earned the 2025 world human-powered speed championship in the university category. Pervis and IUT Annecy appear in neither the entry list nor the event review for WHPSC 2026, so there is no 2026 result. The men's individual world record remains 89.59 mph, set in 2016.

10. LeadIQ internal errors. LeadIQ lists $27M as the latest round while simultaneously dating it December 18, 2025, which is the $100M Series C. Its funding detail is unusable; its $10M to $25M revenue estimate is retained as a database estimate only.

11. Dealroom revenue time series. Values ranging from $38.5M to $2.6B with a uniform negative growth rate are data errors and are not used.

12. Chairman identity. Not a contradiction but a handover: Florent Perrichon in September 2025, Brett Chouinard from July 2026.

13. EPFL Innovation Park building. Official company directory says Building A, TipRanks says Building C. The official directory prevails.

14. Series A amount. $9.1M is the figure; "$9 million" in the later Series B release is rounding.

15. Monolith and CoreWeave transaction dates. Announced October 6, 2025; already booked as acquired by February 25, 2026. One broker page shows October 2026, which matches neither and is treated as a bookkeeping field.

16. Which competitor "Solver" refers to. Most likely SolverX, the Seoul company, given the matching domain, CAD input, founding year, headcount, seed size and the fact that PitchBook lists Neural Concept as one of SolverX's four competitors. The mapping is retained as unverified because the original question did not specify a geography.

17. An "EMail" product. No product by that name exists; the reference is a transcription of Engineering Intelligence.

18. Barbara Marcora as doctoral advisor. No supporting evidence; judged wrong.

19. Acceleration factors, 1,000x versus 5,000x versus 5,760x. Different cases with different baselines, not a contradiction. Any citation must bind the number to its case.

20. The 51.5% figure in the 2018 paper. This is the original published 3D test-set drag accuracy and reflects real difficulty generalizing to unseen geometry. It should not be dropped from secondary summaries.

Risks

No independent reproduction of any benchmark. Every accuracy figure is self-reported. The MIT endorsement is from the dataset author.

The production model is opaque. The architecture is named but not specified, and its relationship to the 2018 paper is undisclosed. There are no public weights and no proprietary public dataset, so third parties cannot verify accuracy.

"Physics-informed" is not technically substantiated. No evidence of physics-informed neural networks, hard constraints or an in-the-loop differentiable solver.

Revenue has never been disclosed in absolute terms. The only official signals are a 4x enterprise revenue increase over 18 months and a statement that ARR growth outpaces burn rate. LeadIQ's $10M to $25M is the only point estimate available.

Valuation rests on third-party estimates. No official figure exists and the largest aggregator contradicts itself.

Material conflict of interest. The chairman and the CRO both come from Altair, a direct competitor in CAE-embedded AI. The CEO and adviser Francois Fleuret are both from EPFL's CVLab, and Fleuret remains a co-founder.

Brand exposure from record attempts. The recumbent bicycle and the SP80 sailboat are brand assets with no verifiable revenue attached.

Template-like rewriting of customer claims. Coverage percentages and customer counts are restated across years within the same boilerplate paragraphs.

Concentration of research assets. The deep learning capability sits inside the EPFL CVLab ecosystem around Fua, Fleuret, Donier and Baque.

Scoring

• Technology, 7.5. Four traceable papers including ICML and AIAA venues, a public benchmark beat with more than halved volumetric velocity MSE, a real differentiable parameterization line of work, and a cheap credible uncertainty proxy. Deducted for zero independent reproduction, an undisclosed production architecture, no public weights, and the paper's own 51.5% 3D generalization result.
• Commercialization, 6.5. More than 50 enterprise customers with strong names across automotive, aerospace, energy, electronics and defense, four F1 teams, a documented reference workflow and a per-deployment PR of 30x. Deducted for zero disclosed revenue, ARR, gross margin or contract values, a customer-concentration pattern typical of this category, and customer-percentage claims that have been repeatedly restated.
• Funding, 9.0. $138.1M disclosed including the seed, a $100M Series C from Goldman Sachs Alternatives growth equity, a syndicate spanning European venture, D. E. Shaw and Aster, and Anaïs Ventures independently confirmed. Deducted only because no official valuation exists and no revenue base yet tests the mark.
• Product maturity, 7.5. A clear naming and layer evolution from NCS through Engineering Intelligence to Design Lab, Studio, CodeLake, AI Design Copilot and MCP Bridge, with deployment across SaaS, on-premise, air-gapped, private cloud and Azure, and a community kit carrying the DrivAerNet++ workflow. Deducted for no published accuracy envelope, an accuracy model still described mainly in relative terms, and 2026 product shifts that remain unproven.
• Market position, 7.0. A credible share of European and Asian OEMs and top-100 Tier-1 suppliers, a neutral-layer position across every major CAD and CAE vendor, a WEF Technology Pioneer designation and explicit naming in the McKinsey and Boardwave European application-layer report. Deducted for a valuation far ahead of any evidenced revenue, a capitalised category that has already moved an order of magnitude, and a persistence moat that remains unquantified while four of the five largest CAE vendors ship embedded AI.

References

1. Neural Concept, "Neural Concept closes $100M funding round led by Growth Equity at Goldman Sachs Alternatives" (2025-12-18)
2. Neural Concept, "Neural Concept raises $27 million Series B to further accelerate global growth" (2024-06-03)
3. Neural Concept, "Neural Concept raises a $9.1m Series A to scale Engineering Intelligence" (2022-03-21)
4. Neural Concept, "From Raw Data To State-of-the-art Results On Aerodynamics ML" / DrivAerNet++ benchmark (2025-09-11)
5. Neural Concept, "Neural Concept Introduces a Physics- and Geometry-Aware AI Design Copilot" (2026-01-07)
6. Neural Concept, "Neural Concept Launches MCP Bridge" (2026-10-19)
7. Neural Concept, "Finding the Path to Value in Engineering AI - Why We Need to Rethink Benchmarking", Pierre Baque (2025-09-18)
8. Neural Concept, careers page: 37 open positions across 11 offices (2026-10)
9. EPFL, "Neural Concept's AI Engineering Redefining Modern Industry" (2026-06-16)
10. EPFL Innovation Park company directory entry (id 11049)
11. Baque et al., "Geodesic Convolutional Shape Optimization", arXiv:1802.04016, ICML 2018
12. Durasov et al., "DEBOSH: Deep Bayesian Shape Optimization", arXiv:2109.13337
13. Wei, Fua, Bauerheim, "Automatic Parameterization for Aerodynamic Shape Optimization via Deep Geometric Learning", arXiv:2305.02116, AIAA Aviation Forum 2023
14. Durasov et al., "Enabling Uncertainty Estimation in Iterative Neural Networks", arXiv:2403.16732, ICML 2024
15. Microsoft Customer Story 27098, Neural Concept on Azure (2026-08-21)
16. Forbes, "Establishing Secure And Sovereign Supply Chains For Physical AI", Sabbir Rangwala (2026-09-30)
17. McKinsey and Boardwave, "Europe's new AI edge? The emerging application layer opportunity" (2026-09)
18. TechCrunch, "How Neural Concept's aerodynamic AI is shaping Formula 1" (2024-04-14)
19. MAHLE newsroom, EV blower design release (2025-04-21)
20. Synopsys investor relations, completion of the Ansys acquisition (2025-07-17)
21. WHPSC.org 2026 Lookback and IHPVA records; SP80 official release on the 2026 Namibia campaign

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of October 21, 2026; please refer to official company announcements and filings.