NAVASTO: Moving Solver Insight Onto the Designer's Desk, and How Autodesk Bought That Entry Point

Executive Summary

NAVASTO GmbH was a Berlin engineering-software company founded in December 2014 by Stefan Blume and Matthias Bauer, who had started aerodynamics doctoral work at TU Berlin in 2005. It built machine-learning surrogates for engineering simulation, first as the NAVPACK API and later as two commercial SKUs. Autodesk agreed to acquire it in October 2024, the deal closed in December 2024, and the family now ships as Autodesk NavPack: NavPack Design runs as a plugin inside Alias, NavPack Core is a Python toolkit for training and applying prediction models.

Five facts frame the case.

First, the technical proposition is architectural rather than promotional. NavPack learns on geometry directly rather than on a parameterised design space, using graph neural networks and, alongside them, point cloud transformers. It accepts NURBS, SubD, mesh or a mixture, and it does not require watertight geometry. That last point is the design-side tell: the input is the surface a designer is sculpting, not a closed manifold prepared for a solver.

Second, the official position on solvers is explicit. The AIF 2025 session notes record the line "This is not a solver", and the 2026 webinar FAQ answers the question head on: does NavPack replace CFD or FEA? No. NavPack accelerates access to physics insight earlier in the workflow, while simulation remains essential for validation, edge cases and expanding the model's knowledge. Autodesk's framing is that predictive AI does not replace simulation, it left-shifts engineering insight.

Third, the evidence base is uneven. Only one set of accuracy figures comes with a complete experimental protocol, the DLR research-car study presented at CDFAM 2024. The three headline numbers in circulation, an error of 1.1 percent, accuracy above 98 percent and accuracy of 99 percent, carry no test conditions at all: no denominator, no geometry set, no indication of averages versus maxima, and no statement of whether training and evaluation data were separated.

Fourth, the company was self-funded. Matthias Bauer wrote in the acquisition announcement that "we bootstrapped our way here". The only identifiable money was EU Horizon 2020 public grants and commercial revenue. There is no trace of an institutional venture round and no record of a valuation.

Fifth, integration has moved faster than the market usually expects. NavPack 2026 reached general availability in December 2025 as a fully integrated Autodesk offering, with Autodesk reporting that adoption among automotive OEMs doubled over the previous twelve months. In October 2026 Autodesk announced at GTC Berlin that NavPack will connect to NVIDIA PhysicsNeMo and NVIDIA Warp, an integration not yet shipped at the collection date.

Our view. NavPack is a real technical asset built on an unusually clean architectural idea, and this looks like a purchase of a workflow position rather than of a solver. Autodesk bought a left-shift entry point into the designer's desktop, together with a decade of simulation workflow knowledge that is expensive to rebuild. The counterweights are equally clear. Financial transparency is close to zero outside the acquirer: the price is undisclosed, pre-acquisition economics are invisible, and no standalone financial statement has appeared. Every public accuracy claim is weaker than the single rigorous experiment supporting the technology. The patent moat is thin, since the known patents describe an aircraft flow-guide body filed in 2018 with no evident relation to the graph-network core. And dependency on the NVIDIA stack, while a genuine accelerator, is also a channel through which the differentiation can be diluted.

Company Overview

NAVASTO was founded in December 2014 with two people in Berlin by Stefan Blume (managing director) and Matthias Bauer (chief executive). The origin is academic: in 2005 both began doctoral studies in aerodynamics at the Technical University of Berlin, whose Chair of Aerodynamics has been led by Prof. Julien Weiss since 2018. After Blume's doctorate he joined the methods development team at AUDI AG, while Bauer worked on projects for Airbus. The two reunited to found the company in 2014. The registered address is Reuchlinstrasse 10-11 in Berlin, on the TU Berlin campus, with a second office in Hamburg.

The company stayed close to its academic origin without a formal institutional tie. The lineage is verifiable in the literature: a 2016 paper in Aerospace lists NAVASTO GmbH as first affiliation alongside the TU Berlin Chair of Aeronautics and Astronautics and the DLR Institute of Fluid Technology. No public source establishes an endowed chair or sponsored professorship; the relationship shows up as co-authorship and staff movement rather than equity.

The company was bootstrapped and operated at small scale. A 2024 conference talk described the organisation as having grown to about 20 engineers, AI scientists and programmers. LinkedIn still shows 11 to 50 staff and 7,048 followers, and the legal entity NAVASTO GmbH remains registered, with PitchBook describing it as an Autodesk operating subsidiary.

Early engineering activity was funded by EU Horizon 2020 grants. DECOROUS ran from June 2016 to May 2019 with a total cost of EUR 388,750 fully contributed by the EU, allocated to the TU Berlin side and therefore not NAVASTO revenue. The SME-Inst project InSoMa (grant agreement 876160) ran from July to December 2019 with a total cost of EUR 71,429 of which the EU contributed EUR 50,000, a 70 percent funding rate, with NAVASTO as coordinator. InSoMa's purpose was to lay the groundwork for NAVPACK: it evaluated technical, legal and economic feasibility, delivered an interface to the ESI PAM-CRASH solver for structural and crash work, produced a proxy-based optimisation application validated on a real automotive OEM use case, and settled the licensing model and pricing strategy.

Commercialisation milestones cluster in 2023 and 2024. NAVASTO partnered with SimScale in September 2023, which launched AI physics prediction on its platform with NAVASTO as model supplier. In October 2023 it announced an innovation partnership with INEOS Britannia, the America's Cup team, whose head of fluid dynamics said their CFD needs to be accurate, fast and reliable and that the tools delivered in all three respects. A month later it announced a strategic partnership with ENGYS for worldwide distribution.

The acquisition sequence is documented to the day. Autodesk announced on October 15, 2024 that it had signed a definitive agreement to acquire NAVASTO, describing the plugin as bringing wind tunnel simulation into the early creative design process and enabling designers to predict and optimise the drag coefficient over an automotive exterior, reducing the likelihood of redesigns during engineering. The same announcement carried standard pre-completion risk language: the acquisition had not been completed and the two companies continued to operate separately until then. Stefan Blume, on an Autodesk University 2025 course page, states completion happened in December. The price has never been disclosed, and no seller-side financial advisor appears in Autodesk's own material; trade press reports the transaction was advised on the financial side by Pava Partners, representing NAVASTO's shareholders.

Integration proceeded on an Autodesk schedule. At AU 2024 the company said the plugin would reach select customers in the fourth quarter of that year and be broadly available alongside Form Explorer in the Alias 2025 release. The Alias 2026.1 help documentation now contains a formal introduction page for NavPack Design, and the Alias 2027 feature list still carries the name "Navasto Plugin for Alias", described as bringing AI-powered aerodynamic insight into design with real-time airflow and drag evaluation.

The brand migrated without the legal entity disappearing. The company LinkedIn page has been renamed Autodesk NavPack, the product is sold as Autodesk NavPack, and NavPack 2026 was introduced as a fully integrated Autodesk offering. The NAVASTO name survives in three places: a learning section on NavPack product pages, course speakers who still describe themselves as NAVASTO co-founders, and the Alias 2027 feature name.

Technology

NavPack runs on a three-stage pipeline with a deliberate pre-check in front of it. Before any data work the team asks four questions: what simulation results already exist, in CFD, structural or thermal form; how many variants there are, tens, hundreds or thousands; which geometric representations the model must learn, NURBS, SubD, mesh or a mixture; and who will use the model, designers, engineers or both. The binding constraint is stated openly: NavPack models do not require watertight geometry, only that the representation of the training samples matches the representation of the geometry being evaluated. That single decision is what makes the tool usable inside a styling environment rather than only inside an analysis environment.

Pre-processing collects simulation results from the engineering environment, where one case can run to tens or hundreds of gigabytes, and converts them into an efficient snapshot format. A geometry-transfer tool makes sure outputs can be mapped onto the geometry types Alias uses. The earliest product form, documented in the 2019 CORDIS reporting, already included the interface to ESI PAM-CRASH.

Model training runs through the NavPack Python API or a browser tool, where engineers optimise the loss, tune hyper-parameters and validate accuracy. The documented workflow order is pre-processing, model training, hyper-parameter optimisation, then prediction with the model. Inference and deployment is where the latency claims live: a trained model runs on a laptop in seconds, and the integration points are NavPack Design inside Alias, ParaView, Blender, Python workflows, and custom integrations in tools such as Ansa, which was in beta at the time of the AIF 2025 session. The official learning library is a 40-minute-49-second webinar on training graph neural networks for CFD led by Jakob Lohse, a podcast on AI-accelerated engineering with Matthias Bauer, and a short interview on the role of AI in engineering.

The modelling route evolved from parametric machine learning to direct geometry learning. Autodesk describes the progression explicitly: graph neural networks let NavPack act directly on complex 3D geometry and give designers full freedom to push, pull, sculpt and refine, without being constrained by preset parameters. 2024 conference material adds that a point cloud transformer is used alongside the graph network, and that inputs may be non-parameterised shapes. Official 2026 material adds that geometric deep learning architectures are what make free-form surface manipulation, and therefore early-stage design, possible.

Uncertainty quantification runs on two mechanisms, because they answer different questions. The first is a global domain judgement: the new geometry and its boundary conditions are compared with training-data coverage to decide whether the case is interpolation or extrapolation, and a confidence value is displayed. Only when confidence clears a threshold is direct use recommended. In the demonstration, a roof profile rising gradually still looked like a normal car and stayed inside the training domain, while a flatbed or van-like shape sat outside it. The second is a local, field-level estimate produced by multi-head networks and deep ensembles, rendered as a colour map where red means the model considers itself uncertain and blue means it does not. The calibration check described publicly is straightforward: run a real physics simulation, subtract the prediction to obtain the actual error field, and compare it with the uncertainty map, which the team reports the two "correlate very well". The vocabulary was fixed in January 2026, when the official FAQ introduced the term "surface uncertainty" for highlighted regions that fall outside the training domain, and the guidance on when a team should add simulation.

The boundary condition on all of this is stated by the company itself. The AIF 2025 notes record the red line: "This is not a solver. The AI model will work on what it was trained to do, but it can't solve every problem, and it's not going to replace simulation." The founder is blunter about extrapolation, describing it as involving luck, and the same session concedes the trade explicitly: the team is trading the generalisability of a first-principles tool for the speed of an AI model. On domain coverage, Autodesk refers to CFD, crash, thermal and durability, and any other domain where geometry can be linked to performance; the 2026 FAQ names CFD and FEA explicitly.

The latency figures come from a 2024 conference talk and are worth keeping precisely because they carry their conditions. Ground truth from a GPU CFD solver takes about 8 hours per run. The interactive loop of modifying geometry, clicking and reading a result takes about 2 seconds. Full-field aerodynamic inference runs at 70 milliseconds per frame, the per-frame update rate of the Formula 1 car video used in the demonstration. A crash box runs at about 20 milliseconds per time step and a full vehicle crash at about 20 milliseconds per frame, against a benchmark of hours for conventional crash simulation. Structural analysis was described as fast enough to be unmeasurable. The geometry in play is of the order of 1 to 2 million nodes or points. A separate company post shows a 12-hour CFD run contrasted with a 20-millisecond AI prediction.

Data requirements are moderate and the company says so plainly. The general starting point is 50 to 100 engineering simulation samples; the demonstration case trained an aerodynamic graph neural network on about 350 samples, a figure cross-confirmed in a 2023 NAFEMS abstract describing a graph neural network trained on 350 car-shape samples. Transfer learning is where the economics get interesting: adding about 20 new samples adapts the model to a new vehicle while preserving predictions for the original DLR car and for the design space between the two, with published cases involving a DrivAer derivative with and without a rear wing and a Volkswagen Jetta. The official FAQ sets the upper end: simple variations may need small datasets, while complex geometry and fine-grained variation may require hundreds to thousands of simulations.

The accuracy evidence must be read carefully, because only one available statement is an experiment. The CDFAM 2024 protocol is complete: a DLR European research car, 500 CFD simulations, roughly 350 for training and roughly 150 validating on geometries the model had never seen, the target quantity the drag coefficient, ground truth from a GPU CFD solver at about 8 hours per run. The reported result is that the absolute error sits in the same order of magnitude as the error of the first-principles simulation itself, and that predicted points fall on the diagonal of the ground-truth comparison. That is a strong result in the right setting and a narrow one. The official qualitative claim is that predictions agree with CFD within the inherent noise of the simulations, with wake slices, in-plane velocities and drag curves described as near-identical. The three numbers without conditions are an error of 1.1 percent, accuracy above 98 percent and accuracy of 99 percent; the first appears only in a third-party rendering of the NAVASTO site and has never been repeated by Autodesk, while the other two come from an Autodesk employee post and a SimScale partner announcement.

The generative half extends the same loop. Rather than optimising toward a single best shape, the generative path proposes new geometric variants that are immediately scored by the predictive model, and the team is explicit that unconstrained optimisation converges on droplet-like shapes nobody wants. The recommended cadence is generate, predict, refine, repeat, and Autodesk frames a full design-evaluate-refine cycle as something that can run in the time of a lunch break. Inside Alias, sensitivity maps answer three questions: which regions dominate drag or lift, where pushing a surface has the largest effect, and how a design decision propagates into the performance result. The recommendation is to search for a favourable region satisfying brand identity, aerodynamics, packaging, manufacturability and cost, rather than a single optimum.

Product and Commercial Model

The product matrix has two SKUs and a clear division of labour. NavPack Core is the toolkit for training and applying prediction models for real-time simulation and analysis, positioned as a Python API. NavPack Design is the interactive design tool exposing near real-time performance predictions and design optimisations inside the design workflow, positioned as the Alias plugin. Design can load a trained model, select geometry, and overlay model outputs such as pressure fields, key performance indicators and sensitivity maps onto the surface model. Core has system requirements of Python 3.10 to 3.12 with the venv module, Linux or Windows, CUDA 11.8 or newer for GPU-accelerated machine learning, and optional tkinter for graphical applications.

Pricing is public for one SKU and deliberately closed for the other. On the Autodesk Store in the United States, NavPack Design is listed at USD 3,155 per year on a one-year term, and the Chinese storefront lists the design product at CNY 6,773 per year. NavPack Core is listed as "Contact Autodesk for pricing". For context, Alias Concept starts at USD 460 per month, Inventor is USD 2,585 per year, AutoCAD is USD 2,095 per year and Navisworks Simulate is USD 1,145 per year.

The licence model is a network subscription with floating seats, no per-user assignment and no sign-in requirement. One operational detail is a genuine first-hand indicator of transitional state: the purchase page states that software download continues to be available through pre-acquisition systems until NavPack is fully integrated into the Autodesk portfolio, which was still the wording in 2025. Autodesk also offers a NavPack Onboarding Service for ramp-up.

Distribution runs along four channels: the Autodesk global channel, ENGYS for CFD through HELYX and ELEMENTS, the SimScale platform, and Rescale. The pre-acquisition channel relationships are documented from the NAVASTO side, 2023 for both SimScale and ENGYS. Their persistence after the transaction is not documented, and ENGYS in particular has no public update after the 2023 agreement.

The customer roster must be read as layered in time rather than as one list. The current official roster, from the NavPack 2026 general availability announcement of December 8, 2025, names GM, Audi, Volkswagen, Seat and Lamborghini. The Autodesk Welcome NAVASTO post from December 2024 named VW, Audi, GM and Airbus, and the NAVASTO LinkedIn page carries an older line naming Airbus, AUDI, Volkswagen and MAN. Airbus and MAN therefore belong to the 2014 to 2019 early period and no longer appear in the current list. Named projects are more specific: GM Motorsports evaluates subtle aerodynamic changes to racing geometry and is reported to have dramatically reduced iteration time during competitive season development; Audi uses AI-driven sensitivity maps with fast drag prediction to guide surface optimisation and validate design directions earlier; Bentley appears in the October 2026 Autodesk news caption as a user of NavPack in its vehicle design workflow; INEOS Britannia remains a named partner with a published testimonial. On Formula 1, 2024 material describes NavPack as the go-to solution for AI-accelerated engineering across multiple teams.

The most useful adoption metric Autodesk has published is qualitative but specific: NavPack adoption among automotive OEMs doubled over the past twelve months, stated in the December 2025 general availability announcement. No absolute count accompanies it. Beyond automotive, product pages cite industrial simulation users including a valve surrogate model trained by an Indian engineering consultancy, and the fluid domain reaches hydrodynamics through a marine hull case.

Post-acquisition momentum is the strongest part of the story. On November 19, 2025 the AIF 2025 session delivered two talks, one by Bauer on AI-accelerated design and engineering and a deep dive by Lohse, with a detailed written account. On November 27, 2025 a further blog post set out the three conditions for left-shifting: integration into the design tool, intuitive visualisation, and education aimed at designers, together with confidence indicators and uncertainty visualisations. On December 8, 2025 came the general availability of NavPack 2026. On January 29, 2026 a joint Alias, VRED and NavPack webinar introduced the surface uncertainty terminology and the explicit statement that NavPack does not replace CFD or FEA. On October 21, 2026, at GTC Berlin, Autodesk announced that NavPack will integrate with NVIDIA PhysicsNeMo and NVIDIA Warp to extend model architectures and GPU-accelerated workflows, quoting the vice-president of automotive design at Autodesk. As of the collection date that integration is announced and not shipped, with no published timetable.

Competitive Landscape

NavPack competes where engineering insight is surfaced, and the comparison set is better drawn along that line than along a vendor list.

PhysicsX is an AI-native large-physics-model company that raised USD 300 million in a Series C in 2026 led by Temasek, with Atomico, General Catalyst, NVIDIA, M&G and Intrepid Growth Partners participating, at a reported valuation of USD 2.4 billion, on a team grown from about 150 to more than 300. Its stated proposition is a native architecture rather than AI attached to an existing solver. Neural Concept, an EPFL spinout, closed a USD 100 million Series C in December 2025 led by Goldman Sachs Alternatives Growth Equity, reports more than 50 enterprise customers and enterprise revenue growing fourfold in eighteen months, and sells a CAD-native physics-aware platform. Luminary Cloud is a cloud-native GPU multi-physics platform that raised USD 72 million in a Series B and has published an industrial aerodynamic foundation model in the SUV segment. SimScale, based in Munich, has raised more than EUR 27 million and has shipped AI capabilities powered by NAVASTO itself since 2023, making it both a distribution partner and a customer of the underlying models. Emmi AI appears in the public record only as a brief mention in a secondary roundup of AI-assisted CAE vendors. Among the vendors covered on this site, the Chinese CAE firm SimWE differentiates through owned solver kernels and a unified GPU multi-physics architecture, while Prometheus, Vinci, Mistral AI and World Labs occupy different segments again. Nothing in that set shares NavPack's specific combination.

The differentiating combination is fourfold. Geometry handling is design-side native, with NURBS, SubD, mesh and mixed representations supported and no watertight requirement, a different starting point from platforms assuming a mesh ready for analysis. The deployment point is inside the designer's own CAD environment, Alias, rather than in a separate analysis application the designer is asked to consult. The uncertainty treatment is designed for a designer audience, with surface uncertainty and confidence indicators framed as guidance on when to add simulation rather than as confidence intervals for a specialist. And distribution runs through a vendor with a global installed base, with three additional third-party routes alongside it.

The challenges are equally specific. PhysicsX and Neural Concept are independent platforms that do not depend on a CAD vendor, and both pursue the enterprise buyer with far larger disclosed capital than NAVASTO ever raised. If a customer wants an AI-native environment rather than a prediction embedded in an existing CAD seat, NavPack is competing for a different budget line. Channel relationships with SimScale, ENGYS and Rescale are documented from the NAVASTO side, and their post-acquisition status is not public, which is an unusual source of ambiguity for a product now sold by the acquirer. The announced NVIDIA dependency cuts both ways: PhysicsNeMo and Warp extend what the model architectures can do, while the same ecosystem contains tools that could absorb part of the function NavPack provides.

Funding

NAVASTO was bootstrapped, and the founder said so in plain language. In the acquisition announcement, Matthias Bauer wrote: "Second, because we bootstrapped our way here. Getting paid for what you do, more than it costs, is one of the best signs you're on the right track." That is the most direct evidence available on the capital structure, and it is consistent with the absence of any record of institutional venture investment.

The only money that can be identified is public grant funding and commercial revenue. Under Horizon 2020, the SME-Inst project InSoMa recorded a total cost of EUR 71,429 with an EU contribution of EUR 50,000, with NAVASTO as coordinator. DECOROUS recorded EUR 388,750 fully contributed by the EU, allocated to the TU Berlin side and therefore not NAVASTO revenue. Commercial revenue was never quantified publicly, although CORDIS states the intent that licence sales and services would bring significant revenue growth, and the InSoMa work delivered a proxy-based optimisation application validated on a real automotive OEM use case.

What the databases say is different. PitchBook lists financing rounds of 4 for NAVASTO, with 18 employees and an acquisition date of October 17, 2024. The most consistent reading is that the figure counts non-equity money, such as the two EU grants and possibly other grant or secondary inflows, and not equity venture rounds. Crunchbase lists the European Union among organisations featuring the company, which is a public grant and not a venture investor, and its Recent Events field is an unfilled placeholder containing "Raised Funding Round / Jan 1, 1970 / Lorem ipsum dolor sit amet", so that section cannot be used as financing evidence at all; the same page hides the founding year. Searches of the published portfolios of BMW i Ventures, Porsche Ventures, High-Tech Gruenderfonds and HTGF return no NAVASTO position. No valuation, total funding figure or debt financing has been disclosed. The single revenue number in circulation, USD 1 million to 5 million, comes from a contact database that labels it an estimate based on industry analysis rather than a disclosed figure. For the transaction itself, the price, the form of consideration and any earn-out are undisclosed.

DISCREPANCIES

Six source conflicts were identified and adjudicated item by item.

1. The acquisition date. Autodesk's newsroom carries October 15, 2024 for the signed definitive agreement. PitchBook records the company as acquired on October 17, 2024. Stefan Blume's Autodesk University 2025 course page states the acquisition by Autodesk happened in December. Some secondary sources compress this into "acquired in October 2024", and at least one autogenerated page claims 2025. Adjudicated: announced October 15, 2024, completed in December 2024. The PitchBook date is two days off and reflects the practice of filling the field with the announcement date; the 2025 claim contradicts first-party sources and is not used.

2. The founding team. The consistent public record names Stefan Blume as managing director and Matthias Bauer as chief executive and co-founder. A third name, Paul Buskirk, appears in no source at all and is most likely a confusion with a similarly named person elsewhere. Julien Weiss, Chair of Aerodynamics at TU Berlin, is connected through co-authorship and staff movement but is not a founder. Adjudicated: Blume and Bauer only. The unsourced name is excluded, and the TU Berlin chair is described as an academic collaborator rather than a founder.

3. The accuracy numbers. Six formulations are in circulation: an error of 1.1 percent from a third-party rendering of the NAVASTO site, accuracy above 98 percent from an Autodesk employee's post, accuracy of 99 percent from a SimScale partner announcement, the official qualitative claim of agreement within the inherent noise of simulations, the fully specified CDFAM 2024 result, and the 2026 official blog repeating that qualitative claim. Adjudicated: the three numeric claims are marketing figures with no disclosed test conditions, retained only with that labelling; the CDFAM 2024 result and the official qualitative claim are the usable evidence. The three numbers are mutually consistent in magnitude, since 1.1 percent error corresponds to about 98.9 percent accuracy, which makes them corroborative of one another rather than independently verifiable.

4. The customer roster. The NAVASTO-era site listed GM, Lamborghini, Audi, Volkswagen, Airbus, Seat and an entity transliterated as "Damco". The LinkedIn page names Airbus, AUDI, Volkswagen and MAN. The December 2025 Autodesk announcement names GM, Audi, Volkswagen, Seat and Lamborghini; the 2024 Autodesk post named VW, Audi, GM and Airbus. Adjudicated: Airbus and MAN are early-period customers from 2014 to 2019 and no longer appear in the current roster; the "Damco" entry has no first-party source and is excluded, with MAN the most plausible reading of the original; the current roster is GM, Audi, Volkswagen, Seat and Lamborghini, with GM Motorsports, Audi, Bentley, INEOS Britannia and multiple Formula 1 teams as named project cases.

5. The funding record. The founder's statement that the company bootstrapped its way here and the absence of any venture investor stand against a PitchBook count of four financing rounds and a Crunchbase investor field naming the European Union. Adjudicated: bootstrapped, with no external venture capital. The PitchBook count treats public and grant money as financing rounds, the Crunchbase investor entry is a public grant, and the Crunchbase Recent Events block is an unfilled placeholder unusable as evidence.

6. The brand. Three statements coexist: the product and the company page are branded Autodesk NavPack; the Alias 2027 feature is named "Navasto Plugin for Alias"; and the legal entity NAVASTO GmbH continues to exist with its original Berlin and Hamburg locations. Adjudicated: all three are true at once. The external product brand is Autodesk NavPack, while the NAVASTO legal entity and part of the plugin naming survive.

Risks and Diligence Gaps

The transaction and the financials are entirely invisible. No source discloses the price, and Autodesk's published results contain no separate NAVASTO consideration or goodwill line that could be used to infer it. Pre-acquisition revenue, gross margin, loss and cash position were never disclosed, and the only revenue figure in circulation is explicitly an estimate from a contact database. Whether earn-out or retention terms exist is unknown, and whether the entity still operates as a separate business unit or was folded into the automotive design organisation has no first-party statement.

Every accuracy claim in the public record lacks test conditions. The 1.1 percent, 98 percent and 99 percent figures have no denominator, no geometry set, no indication of averages versus maxima, and no statement about strict separation of training and evaluation data. The official formulation that predictions agree within the inherent noise of the simulations carries no RMSE, no correlation coefficient and no uncertainty band. For a product sold on accuracy, this is the largest gap in the public record.

The uncertainty quantification is a feature description rather than a method. Whether the ensembles are deep ensembles, multi-head networks or MC dropout, how they are calibrated, what distance measure drives the interpolation-versus-extrapolation judgement, and what the confidence threshold means are all undisclosed. The founder has acknowledged that extrapolation involves luck, but no quantitative failure boundary has been published.

NavPack Core has no public price. The "Contact Autodesk for pricing" listing means unit economics and gross margin cannot be assessed from public information, and the comparison between the designer-facing SKU and the engineer-facing one cannot be made.

The patent moat is thin. The known patent family describes a flow guide body for an aircraft with a fluidic exciter, first filed in September 2018, and one database records only a single document. The machine-learning and graph-network core the business runs on shows no evident patent protection. What protects the business is data, workflow embedding and brand, and all three are replicable in principle by a competitor with the same distribution reach.

The status of third-party channels after the acquisition is unknown. SimScale still described itself as working in collaboration with NAVASTO as of June 2026, while the ENGYS agreement from 2023 has no public update. If NavPack Core is now a first-party Autodesk product, the relationship with these resellers and platform partners carries a potential channel conflict, which is a hypothesis worth testing rather than a conclusion supported by sources.

The February 2025 Autodesk restructuring is an open question for this team. Autodesk reported a restructuring in February 2025 with roughly 9 percent of its workforce, around 1,350 people, affected. Whether it touched the former NAVASTO organisation is not disclosed. Team retention after a small acquisition is a standard diligence question, and here the headcount after December 2024 has not been updated publicly.

Dependency on the NVIDIA ecosystem both accelerates and dilutes. The October 21, 2026 announcement of PhysicsNeMo and Warp integration extends the model architectures and GPU workflow, and also places part of the roadmap in a third party's stack, where native tools in the same ecosystem can absorb some of the same function.

Independent platforms are squeezing the CAD-embedded position. PhysicsX, with USD 300 million raised and a reported USD 2.4 billion valuation, and Neural Concept, with USD 100 million and more than 50 enterprise customers, both sell without depending on a CAD vendor and both target the same engineering budget with far larger disclosed resources.

The documentation is thinner than the marketing. Detailed release notes for NavPack 2026, Alias 2026.1 and Alias 2027 are largely absent, main NavPack documentation beyond the Alias plugin introduction is not public, the graph-network architecture has no implementation-level public description, and the older NAVPACK API and command-line tooling show no clear maintenance status. The announced NVIDIA integration had not shipped at the collection date and carries no timetable.

Overall Assessment

• Technology 7.5. Direct geometry learning with graph neural networks and point cloud transformers, support for NURBS, SubD, mesh and mixed representations, a design-side requirement of no watertight input, a complete three-stage pipeline with a documented pre-check, two distinct uncertainty mechanisms including field-level ensemble estimates, an official and clearly stated limit on what the model will do, a fully specified CFD validation with 500 runs whose error is of the same order as the reference solver's own error, and a deployment surface spanning Alias, ParaView, Blender, Python and Ansa. Deducted for the absence of implementation-level detail on the graph architecture, the lack of any published calibration protocol for the uncertainty maps, and the fact that every headline accuracy figure lacks test conditions.
• Commercialisation 6.5. A fully integrated product with two SKUs, public list pricing for the design SKU in both the US and Chinese stores, an onboarding service, three additional distribution routes, named production deployments at GM Motorsports, Audi, Bentley, INEOS Britannia and multiple Formula 1 teams, and an official statement that OEM adoption doubled in twelve months. Deducted for the absence of any absolute customer count, revenue figure, retention metric or service usage figure, the closed pricing on the core SKU, and customer relationships that appear only in image captions rather than narrative text.
• Funding 4.5. The strongest point is not a strength but a fact of structure: the company reached acquisition without external venture capital, on revenue from licence sales and services, supported by two EU Horizon 2020 grants representing a small public contribution, with the founder's own statement confirming the bootstrap. That is a genuine achievement for a German deep-tech company in simulation. Deducted for the absence of any disclosed valuation or transaction price, no disclosed total funding, no disclosed pre-acquisition revenue, and the resulting situation in which an outside reader cannot mark this asset.
• Product maturity 7.0. A coherent two-SKU matrix with a designer-facing plugin and an engineer-facing Python toolkit, generative and predictive capabilities composed into a single loop with a named cadence, sensitivity maps and surface uncertainty promoted to first-class product concepts, broad integration across Alias, VRED, ParaView, Blender and Python, and a 2026 general availability release. Deducted for mostly marketing-level feature documentation, the beta status of some tool integrations, unconfirmed continuity of the NAVPACK API and command-line tooling, and the announced-but-unshipped NVIDIA integration.
• Market position 6.0. A defensible niche defined by design-side geometry handling, embedding inside the dominant automotive styling environment, a designer-facing uncertainty vocabulary and distribution through a global channel, with real industrial evidence from racing, premium automotive and America's Cup programmes. Deducted for a small, roughly 20-person technical organisation at acquisition, concentration in automotive with limited visible penetration in the other domains the product claims, the absence of any published market share, and the presence of well-capitalised independent platforms addressing the same buyer.

References

1. Autodesk News, "Autodesk AI strengthens Fusion and Alias", October 15, 2024: acquisition announcement, drag-coefficient description, pre-completion risk language.
2. Autodesk Design Studio blog, "Welcome NAVASTO", December 10, 2024: early customer roster and product positioning.
3. Autodesk Design Studio blog, "AIF 2025: How NavPack Brings Engineering Intelligence into Alias", November 19, 2025: two-session summary, the "This is not a solver" line, pipeline, sensitivity maps, official highlights.
4. Autodesk Design Studio blog, "Left-shifting in automotive design", November 27, 2025: three conditions for left-shifting, confidence indicators, uncertainty visualisations.
5. Autodesk Design Studio blog, "Physics AI for collaboration: Autodesk NavPack 2026", December 8, 2025: general availability, OEM adoption doubling over twelve months, current customer roster.
6. Autodesk Design Studio blog, "Design smarter cars faster", January 29, 2026: surface uncertainty terminology, FAQ on replacing CFD and FEA, data-volume guidance.
7. Autodesk News, "Autodesk and NVIDIA accelerate engineering feedback loop", October 21, 2026: PhysicsNeMo and Warp integration plan, Bentley customer caption.
8. Autodesk NavPack product overview, purchase and system-requirement pages, plus Autodesk Store and Chinese storefront listings: Core and Design definitions, licence model, pre-acquisition download route, USD 3,155 and CNY 6,773 pricing.
9. Autodesk Alias feature list and Alias 2026.1 help documentation: the "Navasto Plugin for Alias" feature name, NavPack Design plugin introduction.
10. CDFAM 2024 conference record of the NAVASTO talk: DLR research-car protocol with 500 runs, 350 training and 150 validation samples, 8-hour ground truth, 70 millisecond and 20 millisecond latencies, the two uncertainty mechanisms, data-volume figures and transfer learning.
11. NAFEMS AIML 2023 abstract: a graph neural network trained on 350 car-shape samples, cross-confirming the sample count.
12. EU CORDIS, InSoMa grant agreement 876160 factsheet and periodic report: EUR 50,000 EU contribution against a EUR 71,429 total cost, the ESI PAM-CRASH interface, the OEM validation case, licensing and pricing decisions.
13. Matthias Bauer on LinkedIn, acquisition announcement post containing the statement that the company bootstrapped its way here.
14. ENGYS news release of October 30, 2023 and the INEOS Britannia announcement of October 19, 2023: distribution agreement and customer testimonial.

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of October 22, 2026; please refer to official company announcements and filings.