Legal name: NP Company
NP Co. (NP Company) Deep Dive: A 2025-founded Paris Inria-born AI-native simulation startup that raised a EUR 6M pre-seed led by Partech with Mistral AI co-founders as angels, claiming a pre-trained physics foundation model at 1,000x speedup — most claims unverified
A Paris Inria-born AI-native simulation startup (founded 2025) whose model family is officially named GEMs, claiming a pre-trained physics foundation model at 1,000x speedup and a 50,000x full-assembly target; a EUR 6M pre-seed led by Partech; selected into the French Tech 2030 third cohort.
Executive Summary
NP Company (NP Co.) is an AI-native simulation company founded in Paris in 2025 by Emmanuel Menier as CEO and Matthieu Nastorg as CTO, both PhDs in AI for simulation from Paris-Saclay University and both from Inria, the French national research institute for digital science. Its claim is to replace legacy solvers with transformer models pre-trained on industrial physics, compressing a single design iteration from days or weeks to seconds while holding the same fidelity across a full assembly. The model family carries the official name GEMs (Generalized Engineering Models).
On June 2, 2026 the company announced a 6 million euro pre-seed led by Partech, with the Peugeot family office and angels Guillaume Lample and Cedric O, co-founders of Mistral AI, Florian Douetteau, founder of Dataiku, and Vincent Luciani, founder and CEO of Artefact. Partech manages close to 3 billion euros of assets across 220 portfolio companies in 40 countries and is a leading Paris deep tech investor.
The disclosed performance claims come in two tiers. The first is 1,000 times speedups demonstrated on industrial benchmarks organised by the aerospace company Safran. The second is a clear path to 50,000 times on full-assembly problems, which the company explicitly frames as a path rather than a result. The two must be read separately: 1,000 times has been demonstrated, while 50,000 times is a target.
Two external signals arrived in October 2026. On 12 October the company was named to the Cherry Ventures Heartbeat 2026 shortlist, which picked 10 companies out of more than 10,000 screened across Europe, with NPco the only one selected. On 16 October it was selected into the third cohort of French Tech 2030, where the official page lists 80 deep tech companies including 11 in the AI category and NPco sits among the AI entries; selected companies receive one year of state support that is non-financial in nature. These are programme and ecosystem signals, which sit on a different layer from the product, customer and revenue evidence that is still absent.
One framing point matters. This is a company with a coherent thesis, a credible team and the thinnest evidence base in this series. It is pre-product and pre-revenue, with no public API, demo, model card, pricing or named customer, and no headcount figure published by the company, while third-party directories place the team at 11 to 50 people. Every speedup and fidelity claim is a company statement with no independent third-party reproduction. The key questions are three more specific ones: where the pre-training data comes from, what the 1,000 times figure was benchmarked against, and whether 6 million euros can fund pre-training of a physics foundation model.
Company Overview
Founder Backgrounds
The founders' academic records map closely onto the company's claims. Nastorg's research centers on machine learning acceleration of numerical simulation, particularly graph-neural-network Poisson solvers: DS-GPS, the Deep Statistical Graph Poisson Solver presented at NeurIPS 2022 ML4PS and posted as arXiv 2211.11763, learns the solution mapping of the Poisson equation; later work includes an implicit GNN solver for Poisson-like problems in 2023 and a multi-level GNN preconditioner in 2024. The technical meaning of that line is that it does not replace the whole CFD pipeline but the most expensive part, the linear solve, acting as an accelerator or preconditioner for a conventional solver.
Menier works on reduced-order models and hybrid machine learning and CFD methods, with a doctorate at Inria and IRT SystemX from 2020 to 2024 and earlier work as a research engineer at Air Liquide in 2019 and 2020 building surrogate models for hydrogen production plant simulation. That industrial experience matters for understanding customer needs: industrial buyers want tools that finish inside the design cycle, produce trustworthy results and survive review, with model appearance a secondary concern.
What Emerging From Inria Means, and What It Does Not
The company says it emerged from Inria, and the release describes Inria as the institute that shaped a generation of European AI companies, Mistral AI among them. The narrative carries real value in French deep tech: the Inria brand signals academic rigor, access to public research resources and a talent network.
The boundary needs stating. No public information discloses intellectual property transfer, licensing or formal spin-off terms. Strictly speaking, emerging from Inria is currently a lineage of people and culture rather than a legal technology transfer. For diligence this means one question must be confirmed: whether the core algorithms involve Inria employee inventions or third-party patent licenses.
Round Classification
The company calls the round a pre-seed, while some databases record Seed. The site's phase enumeration has no pre-seed value, so the entry is stored as seed while the body retains the original pre-seed wording. This is a classification difference rather than a factual conflict.
Government and Ecosystem Recognition
Two external signals from October 2026 fill gaps that were empty in the earlier version of this record.
First, NP Co. was selected into the third cohort of the French government's French Tech 2030 programme, with the list disclosed by media reporting on 16 October 2026. The official government page lists 80 deep tech companies in that cohort, 11 of them in the AI category, and NPco appears among the AI entries. Selected companies receive one year of state support that is non-financial in nature, and the working group's annual budget was reduced from EUR 6.8M to EUR 1.8M. This item reads literally: selection brings no cash, and public information does not disclose what specific support NPco received.
Second, on 12 October 2026 the company was named to the Cherry Ventures Heartbeat 2026 shortlist, which selected 10 companies out of more than 10,000 screened across Europe, with NPco the only one selected. The brand name used in that ecosystem is NPco, and third-party directory entries place the headcount band at 11 to 50 people; the company has published no headcount figure.
Technology
The Problem a Physics Foundation Model Must Solve
Understanding the claim requires understanding what it aims to replace. A conventional simulation workflow runs through geometry cleanup, meshing, equation discretization, iterative solution and post-processing. Solution cost grows superlinearly with mesh size, so a full-vehicle external aerodynamics model with hundreds of millions of cells takes hours to days for a single steady-state solve and several times more for transients or parameter sweeps.
The surrogate approach fits a mapping from input parameters to output fields with machine learning and skips the iterative solve at inference. It has been studied for over a decade and has industrial deployments such as Ansys SimAI and Neural Concept, but two limits persist: every new geometry and operating condition needs retraining, which hurts generalization, and training data is expensive because a conventional solver must generate it.
The physics foundation model claim is to pre-train a general model on large volumes of simulation data so that it learns a general representation of physical behavior and delivers value immediately when deployed to a customer, without per-customer retraining. If true, the commercial implications are large: it converts project-based custom modeling into a standardized software product, changing both gross margin structure and scaling speed.
The Official Model Family Name: GEMs
The model family carries the official name GEMs (Generalized Engineering Models). Company-facing and ecosystem channels use the GEMs name, while the June release describes the same family as a pre-trained physics foundation model. Both designations point to one technical proposition: GEMs is the product-family name, and pre-trained physics foundation model is the functional description. This report uses both side by side and does not merge them into a single label.
Three Technical Routes for Transformers in PDEs
The company emphasizes a transformer architecture, and there are three possible concrete routes, with public information not saying which it uses.
The first treats the domain as patches, as vision transformers treat images. It reuses mature large model training techniques but loses continuity at patch boundaries, fits complex geometry poorly and scales attention quadratically with patch count.
The second uses physics-aware tokenization, slicing by physical structure rather than a regular grid. Transolver in 2024 is the reference, slicing discretized points by physical state and applying attention within slices to reduce complexity from quadratic in mesh size to approximately linear while retaining global coupling. This route performs best in academic benchmarks and is why this site indexes Transolver papers.
The third replaces a sub-step of the solver such as the linear solve or preconditioner rather than the whole pipeline end to end. This is closest to Nastorg's GNN Poisson solver work, is engineering-safer, and struggles to support the narrative of seconds-level full-assembly simulation.
The three differ greatly in technical substance, verifiability and commercial meaning. The company says transformer without disclosing architecture, so outside observers cannot tell which applies. This is the single most important evidence gap.
The Two Tiers, 1,000 Times and 50,000 Times
The 1,000 times figure is described as demonstrated on industrial benchmarks organised by Safran. The details that matter are missing: which baseline solver was used, whether in-house, open source or commercial; problem scale in cells and degrees of freedom; the definition and quantification of fidelity in relative error and key physical quantities; hardware configuration including GPU model, count and whether comparison used equivalent hardware; and whether the benchmark is reproducible. Without these, 1,000 times is directional evidence rather than a performance conclusion.
From engineering experience, 1,000 times is achievable under specific conditions. If the baseline is a complete high-fidelity transient CFD workflow including mesh generation and many iterations while the surrogate produces a result in one forward pass, a two order of magnitude gap is not implausible. But the comparison is partly unfair: a surrogate's accuracy ceiling is set by its training data, so it is fast inside the training distribution and may fail outside it, while a solver returns reproducible results for any input. The fair comparison is elapsed time at equal accuracy, which requires an explicit error metric.
The 50,000 times figure is explicitly a path rather than a measurement. Its logic can be inferred: full-assembly solution cost grows superlinearly with part count and coupling complexity, while surrogate inference cost grows roughly linearly and can be amortized through batching, so the ratio grows with scale. The argument is mathematically sound, contingent on the model holding accuracy at full-assembly scale, which is the hardest part.
The Pre-Training Data Problem
The biggest engineering obstacle to a physics foundation model is data. Language models scrape trillions of tokens from the internet, while physics models must generate data by simulation or measurement. Pre-training a general model across aerospace, defense, energy, electronics and automotive requires datasets spanning geometric, material, boundary condition and coupled-physics diversity, potentially millions to tens of millions of cases.
Building such a dataset is extremely expensive and involves a chicken-and-egg problem: generating data requires conventional solvers, the very thing being replaced, and high-fidelity solve cost caps dataset size. The industry response has three forms: many low-fidelity cases, which limits accuracy; few high-fidelity cases for calibration, which limits scale; or public datasets, which limit coverage.
Disclosure on data sources, scale and composition is absent. At 6 million euros, cloud GPU budgets would roughly cover training a mid-sized model plus several benchmark demonstrations, not a cross-industry general pre-training corpus. That tension between the pre-training narrative and the funding scale is central.
Technical Shortfalls
First, the architecture is undocumented. There is no white paper, technical blog, model card or paper describing the shipped architecture. The founders' published work is GNN and reduced-order modeling while the product narrative is transformer, with no published bridge.
Second, the definition of fidelity is missing. Same fidelity is the core claim, yet no error norm, tolerance on key physical quantities or applicable operating envelope is given.
Third, generalization is unvalidated. A general model across six industries and multiple physical domains including external flow, internal flow, heat transfer and structures requires cross-domain evidence that public information does not contain.
Fourth, the certification path is unclear. Aviation and defense simulation tools must satisfy airworthiness and military standards, typically requiring extensive verification and validation documentation. The black-box nature of AI models conflicts structurally with explainability requirements, a barrier shared by every AI simulation company.
Product and Commercial Model
Product Status
As of October 2026 there is no named public product. The website offers company background and hiring information with no API documentation, interactive demo, model card, pricing page or deployment specification.
The disclosed roadmap has two stages. The near term is the pre-trained physics foundation model itself deployed on customer infrastructure, including on-premises deployment, which is a key selling point for data-sovereignty-sensitive aerospace, defense and energy buyers. The longer term has two directions: automated design tools that generate optimized architectures, and real-time operational simulators for industrial infrastructure that extend simulation from design into operations as a digital twin.
The real-time operational simulator direction is notable because it moves beyond design tools into operations, using surrogates in place of online simulation to support real-time optimization and anomaly detection. Customer value is higher because operational decisions are time-sensitive, but the model must remain stable over long runs and handle alignment between sensor data and simulation.
The Unknown Business Model
No business model is disclosed. Four possibilities exist: software license by seat or compute, SaaS by call volume, an on-premises appliance suited to defense customers, or project-based services.
Their margin structures and scaling speeds differ greatly. Licenses and SaaS carry high margin and replicate, appliances bring high unit price with heavy delivery, and services bring low margin and poor scalability. The claim of pre-training with no per-customer training logically points to license or SaaS, since no engineering per customer would be required, but that is inference.
Boundaries of the Customer Relationship
Safran is the only named industrial partner and its role is benchmark organiser. The release does not say Safran is a paying customer or that procurement exists. Treating Safran as a customer is a common misreading; at present it is a technology validation partner.
The distinction matters in diligence. A benchmark organiser supplies scenarios and data while a customer supplies money. The former proves someone is willing to look, the latter that someone is willing to buy.
Pricing and Go-to-Market
Pricing is undisclosed. Six target industries are listed: aerospace, defense, energy, electronics, data centers and automotive. The breadth is itself a risk signal, since a pre-seed company addressing six industries simultaneously suggests no minimum viable market has been found.
Peers are more focused: Emmi AI on aerospace and automotive, Jua on energy, PhysicsX on aerospace and automotive. Focus is normal at an early stage, and breadth usually appears when a company can deliver across several industries at once. NP Co.'s breadth is more likely a statement of technical generality than a go-to-market strategy.
Competitive Landscape
Tiers of AI-Native Simulation
NP Co. is the earliest-stage company in this tier. Comparables fall into three groups.
The first is funded peers with shipped or demonstrable products: PhysicsX with a 300 million dollar Series C, Vinci AI with a 250 million dollar Series B, Luminary Cloud with a 72 million dollar Series B, Neural Concept with about 138 million dollars cumulative on a CAD-native graph neural network route, plus BeyondMath, SimScale and Rescale. These have products, customers and in some cases revenue, while NP Co. has none of the three.
The second is direct narrative overlap: Emmi AI in Austria, serving the same aerospace, automotive and semiconductor verticals with physical AI models, acquired by Mistral AI on May 19, 2026. NP Co. is effectively entering the field Mistral had just absorbed.
The third is plug-in AI from platform giants: Ansys SimAI, Siemens Simcenter PhysicsAI, Dassault SIMULIA, Cadence and COMSOL. Their advantage is installed base and validation data with AI sold as an add-on, and their disadvantage is architectural baggage from desktop-era products now adapting to cloud and GPUs.
Head-to-Head With Emmi AI
Emmi AI is the most direct comparison, also building physical AI models, also targeting aerospace and automotive, also emphasizing AI acceleration of simulation, and now resourced by Mistral. Differences may lie in technical route, where Emmi's details are not public, and in commercial stage, since Emmi has customers.
NP Co.'s differentiation is the claim of pre-training with no per-customer training. If that holds, it delivers more efficiently than peers requiring custom training. If it does not, the difference reduces to brand and team.
The Paradox of Mistral Founders Backing a Competitor
The timeline deserves recording: Mistral AI acquired Emmi AI on May 19, 2026, and 13 days later on June 2 the two Mistral co-founders invested personally in NP Co., a competitor in the same field.
Three readings coexist. The market is large enough for multiple routes and a personal investment creates no strategic conflict. Personal angel investing is separate from corporate M and A and acts as a hedge on a frontier direction. And the Inria lineage creates trust, since Lample, Menier and Nastorg belong to the same French AI research network.
Menier offered a pragmatic account: Emmi was exploring options while NP Co. was not on the market, having just closed its round. The explanation is plausible without resolving the structural question: if Mistral endorsed Emmi's technical route, is NP Co.'s route complementary or substitutive? Public information cannot answer.
Europe Versus China
European AI simulation startups typically start from academic methods and deep tech funds such as Partech, Eurazeo and HV Capital, while Chinese peers more often start from domestic substitution and industrial capital, as with Supreium, ShonCloud and Tenfong. The paths differ in how validation is established: European companies build benchmarks and papers before finding customers, while Chinese companies land in customer projects before adding papers.
NP Co. follows the European pattern: NeurIPS papers and an Inria background first, then a Partech investment, with product and customers after. The advantage is fast technical credibility and the disadvantage is slow commercial validation.
Financing and Capital Structure
Round and Investors
The 6 million euro pre-seed was announced June 2, 2026 and led by Partech. Partech is headquartered in Paris with offices in Berlin, Dakar, Dubai, Nairobi and San Francisco, manages close to 3 billion euros, holds 220 portfolio companies across 40 countries, was founded 40 years ago in San Francisco and is a leading European deep tech investor.
Participants include the Peugeot family office and four angels: Guillaume Lample and Cedric O of Mistral AI, Florian Douetteau of Dataiku and Vincent Luciani of Artefact. The cap table carries strong signal value, assembling four representative figures from the French AI and data ecosystem, which materially helps later fundraising and recruiting.
Use of Proceeds
Stated use is expanding the research team and continuing foundation model development, with longer-term work on automated design tools and real-time operational simulators, while hiring AI researchers and engineers with physical engineering backgrounds.
At 6 million euros and Paris research salaries of 80,000 to 150,000 euros per senior researcher before employer charges and equipment, the round covers roughly 15 to 25 people for 18 to 24 months or about ten people for three years. Third-party directories place the team at 11 to 50 people, a band consistent with that estimate. Accounting for GPU training cost, the effective runway is shorter.
Tension Between Round Size and Foundation Models
This is the point the report most wants to stress. The core cost of training a physics foundation model is compute plus data generation: compute for a mid-sized model may run from hundreds of thousands to millions of GPU hours, and generating millions of high-fidelity cases may cost more than the compute itself.
Six million euros looks tight for either. Three explanations are possible. The foundation model may be far smaller than a language model, since physical fields have far fewer degrees of freedom. Training data may come from public datasets and partners such as benchmark data from Safran, keeping self-built cost low. Or this round is genuinely a seed and actual model training depends on the next round.
All three may hold, and each implies a different risk profile. Under the first, the technical barrier may be underestimated. Under the second, data dependence becomes the weak point. Under the third, the company must prove pre-training is feasible within 12 to 18 months to raise again.
Valuation and Exit
Valuation is undisclosed. At a typical European pre-seed post-money of 15 to 30 million euros against a 6 million euro round, dilution would run from 20 to 40 percent, which is normal.
Exit paths are acquisition by Ansys, Siemens, Dassault, Cadence or Synopsys, or by a cloud provider or an AI company such as Mistral, or independent growth. Given industry consolidation, with Siemens acquiring Altair, Synopsys acquiring Ansys and Keysight acquiring ESI, acquisition is the more realistic path.
DISCREPANCIES
1. Round classification and platform enumeration gap. The company and Partech call the round pre-seed while some databases record Seed, and this site's phase enumeration has no pre-seed value so the entry is stored as seed. This is a classification difference rather than a factual conflict. This report uses the 6 million euro pre-seed as the factual basis while retaining the pre-seed wording in the body. Source tier: OFFICIAL (Partech release) over database entry conventions.
2. Currency and amount of the financing. The official figure is 6 million euros per the Partech release, while secondary sources write 6 million dollars as an exchange-rate approximation or 7 million dollars through rounding or conflation with other rounds. The dollar figures carry no stated exchange-rate date and cannot be merged with the euro figure. This report uses 6 million euros. Source tier: OFFICIAL over secondary sources.
3. The 1,000 times claim lacks methodology. The company says 1,000 times speedups were demonstrated on industrial benchmarks organised by Safran, without disclosing the baseline solver, problem scale, fidelity metric, hardware configuration or reproducibility. This report labels it a vendor claim rather than a confirmed performance figure. Source tier: company release, OFFICIAL but self-assessed.
4. The 50,000 times figure is a target, not a result. The company frames a clear path to 50,000 times on full-assembly problems, an aspirational statement at a different level from the demonstrated 1,000 times. Coverage often presents the two side by side as achieved. This report separates the two tiers in the summary and technology sections. Source tier: OFFICIAL (company wording).
5. Safran is a benchmark organiser, not a customer. The release says the benchmark was organised by Safran without confirming a paying customer relationship or procurement. Some readings treat Safran as a customer. This report limits Safran to technology validation partner. Source tier: OFFICIAL (release wording).
6. Emerging from Inria lacks legal documentation. The company says it emerged from Inria and the release describes Inria as the institute that incubated companies such as Mistral. Public information contains no IP transfer, license agreement or formal spin-off terms. This report distinguishes a lineage of people and culture from a legal technology transfer. Source tier: OFFICIAL (narrative) with documents absent.
7. The paradox of Mistral founders backing a competitor. Mistral acquired Emmi AI on May 19, 2026, and 13 days later Lample and Cedric O invested personally in NP Co., a competitor in the same field. The timeline is OFFICIAL and the investment is a personal angel position separate from corporate strategy. Both facts stand without being smoothed over, and no inference is drawn that Mistral directed the investment. Source tier: OFFICIAL (timeline) and MEDIA (interpretation).
8. The Augur former name is unverified. Only one Chinese secondary source says the company previously operated quietly under the name Augur, with no support in English or official sources. This report does not treat it as fact. Source tier: UNVERIFIED.
9. Model family naming. The company release and the company's own wording describe a pre-trained physics foundation model, while company-facing and ecosystem channels carry the official family name GEMs (Generalized Engineering Models). Both designations refer to one model family, one as a product-family name and one as a functional description. This report keeps the two side by side in the technology and product sections and does not collapse them into a single label. Source tier: OFFICIAL (company wording) with an ecosystem naming record.
10. The nature of the French Tech 2030 selection. The third cohort was disclosed by media reporting on 16 October 2026; the official page lists 80 deep tech companies, 11 of them in the AI category, with NPco among the AI entries. Selected companies receive one year of state support that is non-financial in nature, and the working group's annual budget was reduced from EUR 6.8M to EUR 1.8M. Selection does not equal funding, and public information does not disclose which specific support NPco received. Source tier: OFFICIAL (government page) plus MEDIA (relayed reporting).
11. The Heartbeat shortlist and the headcount band. The Heartbeat 2026 shortlist was published on 12 October 2026 and selected 10 companies from more than 10,000 screened across Europe, with NPco the only one selected. The brand name used in that ecosystem is NPco, and third-party directory entries give a headcount band of 11 to 50 people while the company publishes no headcount figure. The word only applies to that one shortlist event and is not read as a company-level uniqueness claim. Source tier: MEDIA and third-party directory entries.
Risks and Diligence Gaps
First, pre-product and pre-revenue. No customers, pricing, API or model card exist, and all speedup and fidelity claims are company statements. That is normal for a company founded in 2025 but means no commercial metric can cross-check the technical claims.
Second, the gap between the architecture and published work. The founders' peer-reviewed work is graph neural networks, reduced-order models and Poisson solver acceleration while the product narrative is end-to-end transformer simulation. No public bridge exists in the form of a white paper, architecture diagram or training data description, making this the largest technical uncertainty.
Third, training data sources and scale are undisclosed. Data is the core asset of a physics foundation model and its acquisition cost determines feasibility. The tension between a 6 million euro round and a claim of a general model across six industries needs explanation before the next round.
Fourth, fidelity is undefined. Same fidelity is the central selling point without an error metric, applicable envelope or failure boundary. The first question industrial buyers ask is whether results can be trusted, and the answer is not in public information.
Fifth, certification and validation barriers. Aviation and defense simulation tools require airworthiness or military standard verification, and AI explainability, traceability and certification conflict structurally. No certification path is disclosed.
Sixth, an over-broad market positioning. Six industries at once for a pre-seed company means dispersed resources. Early companies normally validate a minimum viable market, and NP Co.'s framing reads more as technical generality.
Seventh, team size and talent competition. The company says it is hiring and has published no headcount figure; third-party directories list the team at 11 to 50 people, which is a directory figure and not a company disclosure. AI talent competition in Paris is intense across Mistral, Dataiku, Artefact and academic labs, and startups must attract senior researchers with equity and mission at high cost.
Eighth, single lead investor dependence. Partech is the only institutional investor alongside a family office and angels, so next-round pricing and endorsement largely depend on Partech following on.
Composite Assessment
Technology scores 6.5. The founders' academic grounding is solid, with Nastorg's GNN Poisson solvers and Menier's reduced-order modeling and industrial surrogate experience, the pre-trained physics foundation model thesis is coherent with clear industrial value, and transformers for PDEs have academic support in the Transolver line. Deductions for an undocumented architecture, the gap between published work and product narrative, an undefined fidelity metric, and unvalidated cross-domain generalization.
Business scores 3.5. The target market is real, since simulation bottlenecks exist across six industries, the pre-training claim would deliver a better margin structure if true, and real-time operational simulators offer long-term imagination. Deductions for pre-product and pre-revenue status, no customer, no pricing, no disclosed model and an over-broad positioning.
Finance scores 6.0. The 6 million euro pre-seed led by Partech with a cap table of four representative French AI and data figures carries high signal value, and the use of proceeds is clear. Deductions for a small absolute amount, tension with the capital needs of foundation model training, and dependence on a single institutional investor.
Product scores 3.0. There is no public product form: no API, demo, model card, pricing or deployment package. The roadmap of automated design tools and real-time operational simulators is directionally sound but distant from delivery, and product evidence is entirely absent.
Market scores 7.0. Engineering simulation is a large market estimated at about 30 billion dollars by 2030, AI-native simulation is moving from concept to product as peer funding and M and A show capital endorsement, and the European deep tech ecosystem plus French AI talent supply provide soil. Deductions for the head start of incumbents and well-funded peers, certification barriers, and the long validation required by the replace-the-solver narrative.
Overall 5.3. This is a high-conviction, low-evidence bet. The upside rests on credible founders, a coherent thesis, a strong-signal cap table and a large market, while the downside is an undocumented architecture, speedup claims without methodology and a pre-product company facing incumbents such as Ansys, Siemens and Dassault plus funded peers such as PhysicsX, Neural Concept and Emmi AI that already ship. The judgment on physical AI is that the direction being bet on, pre-trained physics foundation models replacing per-customer training, is the most valuable and hardest step in this field and that success would mean far more than a faster solver, while at the current evidence level this is an investment in team and thesis rather than in technology and product.
References
- Partech release, NP CO RAISES 6M PRE-SEED TO BRING FRONTIER AI TO INDUSTRIAL ENGINEERING, BACKED BY MISTRAL AI FOUNDERS, June 2, 2026: 6 million euro pre-seed led by Partech with the Peugeot family office and angels Guillaume Lample, Cedric O, Florian Douetteau and Vincent Luciani; founded in 2025 by Emmanuel Menier and Matthieu Nastorg, PhD researchers in AI for simulation from Paris-Saclay University emerging from Inria; core technology built on a transformer architecture pre-trained on industrial physics; 1,000 times speedups demonstrated on industrial benchmarks such as those organised by Safran with a clear path to 50,000 times on full-assembly problems; proceeds for research team expansion and foundation model development with longer-term work on automated design tools and real-time operational simulators; target industries aerospace, defense, energy, electronics, data centers and automotive
- Founder quotes in the same release: Menier describing engineering as the next frontier for AI rather than the next chatbot and simulation as the decades-long bottleneck in industrial design; Nastorg stating the aim was not a faster version of existing tools and that a pre-trained foundation model allows exploring thousands of designs in the time previously needed for one simulation
- Partech institutional information in the same release: headquartered in Paris with offices in Berlin, Dakar, Dubai, Nairobi and San Francisco; close to 3 billion euros under management; 220 portfolio companies across 40 countries on four continents; founded 40 years ago in San Francisco
- Financing record, The SaaS News, June 2026: NP Company raised a 6 million euro pre-seed led by Partech with the Peugeot family office and four angels; proceeds to scale development of transformer-based physics simulation models and grow the AI research and engineering team; founded in 2025 by Menier and Nastorg as researchers from Inria and Paris-Saclay University
- Financing record, Startup.eu: Paris-based company described as spun out of Inria; 6 million euro pre-seed led by Partech with the Peugeot family office and angels Lample, Cedric O, Douetteau and Luciani; proceeds for research team expansion, foundation model development, automated design tools and real-time operational simulators
- Financing record, EU Information Service: confirming the 6 million euro pre-seed and investor list, the six target industries, and the pre-trained transformer models on industrial physics data with claims of up to 1,000 times speedups on industrial benchmarks and continued work on larger-scale assembly simulation
- Financing record, InforCapital, June 2, 2026: round type pre-seed with a 6 million dollar equivalent value; associated companies listed as NP Company, Mistral AI, Safran, Dataiku, Artefact, Inria and Paris-Saclay University; early demonstrations showing up to 1,000 times speedups on benchmarks set by Safran with a trajectory toward 50,000 times on full-assembly challenges
- Founder academic work: M. Nastorg et al, DS-GPS, A Deep Statistical Graph Poisson Solver for faster CFD simulations, NeurIPS 2022 ML4PS and arXiv 2211.11763; An Implicit GNN Solver for Poisson-like problems, 2023; Multi-Level GNN Preconditioner for Solving Large Scale Problems, 2024
- Founder background: Emmanuel Menier personal site, doctorate at Inria and IRT SystemX from 2020 to 2024, earlier research engineer at Air Liquide from 2019 to 2020 building surrogate models for hydrogen plant simulation, research on reduced-order models and hybrid machine learning and CFD methods
- Industry background from public sources: engineering simulation market estimated at about 30 billion dollars by 2030 according to Dealroom; AI-native simulation peers including PhysicsX with a 300 million dollar Series C, Vinci AI with a 250 million dollar Series B, Luminary Cloud with a 72 million dollar Series B, Neural Concept with about 138 million dollars cumulative, plus BeyondMath, SimScale and Rescale
- Consolidation background from public sources: Mistral AI acquired Emmi AI of Austria, which serves aerospace, automotive and semiconductor customers with physical AI models, on May 19, 2026; Siemens acquired Altair, Synopsys acquired Ansys, Keysight acquired ESI, Autodesk acquired NAVASTO and CoreWeave acquired Monolith AI, reflecting consolidation across simulation
- Technical background from public sources: three main transformer routes for PDEs comprising patch-based processing, physics-aware tokenization such as Transolver slicing attention that reduces complexity from quadratic in mesh size to approximately linear, and replacement of solver sub-steps; main obstacles to physics foundation models include training data that must be generated by simulation, non-standardized tokenization of geometry and mesh, and hard cross-domain generalization; certification of AI simulation in aviation and defense conflicts with explainability and verification documentation requirements
- French government official page, lafrenchtech.gouv.fr, French Tech 2030 third cohort list: 80 deep tech companies listed, 11 in the AI category, with NPco among the AI entries; selected companies receive one year of non-financial state support, and the working group's annual budget was reduced from EUR 6.8M to EUR 1.8M
- SeedOps coverage dated 12 October 2026: NPco selected for the Cherry Ventures Heartbeat 2026 shortlist, 10 companies chosen from more than 10,000 screened across Europe with NPco the only one selected; the same entry records the GEMs naming of the model family
- Third-party company directory entry for NPco / NP Company: brand name NPco and a headcount band of 11 to 50 people, which is a directory figure and not a company disclosure
- Relayed reporting of the French Tech 2030 third cohort list, 16 October 2026: the cohort disclosure and the AI category placement of NPco
This report is compiled from public sources for informational purposes only and is not investment advice. The company is private and pre-product, all performance claims are company statements without independent verification, and funding and team data come from releases and third-party databases with conflicts adjudicated in the text. Data is current as of 2 November 2026.
Related papers
- Fourier Neural Operator (Original Paper)
- DeepONet (Original Paper)
- Transolver: A Transformer Solver for Large-Scale Industrial Geometries
- Transolver is a Linear Transformer: Revisiting Physics-Attention through the Lens of Linear Attention
- PhysicsNeMo: Neural-Operator-Based Physics Simulation Framework
- Physics-Informed Learning Annual Review 2025-2026
Related tools
Related cases
- AirfRANS High-Fidelity Airfoil Surrogate Benchmark Case
- DrivAerML High-Fidelity Road-Car Aerodynamics Dataset and Surrogates
- Physics-Informed CNN for Battery Pack Temperature Fields
- Wind Turbine Aeroelastic Simulation and Wake Control: A Verifiable Chain from OpenFAST to Reinforcement Learning
- TopoDiff: Diffusion-Model Topology Optimization and Its Evaluation Metrics
Related reports66
- 51World Deep Dive: A Listed Digital Twin Company Where 51Sim Becomes the First Growth Engine While 51Earth Nearly Vanishes, and Two Irreconcilable Financing Naming Conventions Must Both Be Kept
- AMI Labs: A USD 1.03B Seed Bet on World Models, LeCun's Non-Generative Path and Its Blank Scorecard
- Anshi Pacific (PERA) Deep Dive: A Beijing Domestic-CAE Leader Founded in 2003 Whose PERA SIM General Simulation, AIROM Reduced-Order Models, AgentX Multi-Agent and Jingzhi iGPT/Whale Span from Industrial Simulation to Physics AI
- AspenTech Deep Dive: The Process-Simulation Incumbent Acquired by Emerson and Its Hybrid Industrial-AI Route
- Autodesk Deep Dive: Neural CAD Foundation Models and a $200M World Labs Bet on Physical-World AI, with NAVASTO Adding CFD Surrogate Capability
- Bentley Systems Deep Dive: iTwin Digital Twin and Generative AI Empower Full-Lifecycle Infrastructure Engineering
- BeyondMath Deep Dive: A Foundation Physics AI on the Neural-Operator Route, a $19M Consortium Budget, and Three Misread Facts
- Cadence Deep Dive: System-Level Multiphysics EDA and the Millennium M2000 Supercomputer Anchor Physical AI, with AuraStack Agentic AI and Fidelity GPU CFD as Key Levers
- COMSOL Deep Dive: Four Decades of Multiphysics Solvers, Entering Physical AI via DNN Surrogate-Model Training and All-Physics GPU Acceleration
- Dassault Systèmes Deep Dive: 3DEXPERIENCE and SIMULIA Virtual Twin Physics Behavior Anchor Industrial AI, with LEO Virtual Companion and GPU Explicit Dynamics as Physical-AI Levers
- DP Technology: From the Gordon Bell Prize to National-Level Unicorn Status and the Door of a Hong Kong Listing, the AI-for-Science Leader's Technical Delivery and Commercial Blank
- Emmi AI: From Linz Spin-out to a EUR 330M Mistral Acquisition in 17 Months
- Fainite Deep Dive: A Zurich Physics-AI Simulation Startup Whose Caltech, ETH Zurich, Cambridge and Google Team Uses Physics-Aware Deep Learning to Cut Simulation from Tens of Hours to Seconds, With Only Venture Kick Grants and No Priced Round
- Fysics AI: The PhysX Founder's Second Act, a Domestic Differentiable Physics Engine and Its Ecosystem Play
- Frontop (301313) Deep Dive: From Digital Exhibition Halls to a Physical-AI Data Foundation - FTE Engine, FunSim Simulation and the Gap Between Positioning and Revenue
- Godela Deep Dive: A YC X25 Physics-AI Simulation Startup Whose Ex-Apple, Google and Intel Team Uses Physics-Constrained Surrogate Models to Cut Simulation from Days to Seconds, at Seed Stage with Pilot-Only Access
- HollyWave (688682) Deep Dive: RDSim AI Electromagnetic Simulation and Engineering Agents Close the Simulation-Measurement-Product Loop
- IAT (Alt) Deep Dive: China's First Listed Independent Automotive Design Company Pivots to AI-native R&D with the Yuanfang, Yufeng and Taiyi 3.0 Product System and an NVIDIA Omniverse Partnership, Reporting RMB 975M Revenue and a Widened RMB 250M Net Loss in 2025
- Inductiva Deep Dive: A Closed Case in Open-Solver Orchestration, Eight Premises That Must Be Corrected, and Where Co-Founder and CTO Hugo Penedones Went Next
- Intesim (INTESIM) Deep Dive: A Home-grown Chinese CAE Leader Founded in 2009 with Self-developed Solvers for Structure, Fluid, Electromagnetics, Thermal and Acoustics and Multiphysics Coupling, a ~RMB 100M B Round from CNII and Shenzhen Capital, and 2025 Revenue and Bookings Both above RMB 100M
- iSoftStone (301236) Deep Dive: A ChiNext-Listed IT Services Giant Whose TianShu Platform Builds a Three-Layer Physical-AI Stack with Self-Developed TwinHub Digital Twin, FSim Industrial Simulation, CAE Multiphysics and Phyxis Embodied Control
- Jua Deep Dive: The Earth Physics Transformer as a Large Physics Model for Energy Trading
- Keysight (with ESI Group) Deep Dive: Virtual Prototyping and Hybrid AI Bridge Measurement and Simulation, with Natural-Language Assistants in EDA
- Lightwheel Deep Dive: A Physical AI Data and Evaluation Infrastructure Startup Founded by a Former NVIDIA Simulation Lead, Seven Rounds Totalling Roughly CNY 2.65 Billion at a Valuation Above USD 2 Billion, Seats on Both the Newton and EgoVerse International Standards Committees, With Two Incompatible Accounts of Q1 Orders
- Luminary Cloud: The Physics AI Model Factory Backed by NVentures
- MathWorks Deep Dive: MATLAB/Simulink and Simscape Multiphysics Power Physical AI, with R2026a Copilots and an Agentic Toolkit for Edge Intelligence
- Maniformer (Mifeng) Deep Dive: An Agibot-Spun Physical-AI Data Infrastructure Firm Whose Bodyless MEgo Capture Hardware Is in Mass Production and Which Closed Three Rounds in Half a Year Led by China Telecom
- Mistral AI: How Physics AI Moved From Language Models to EUV Lithography and Nuclear Power
- Monolith AI: The British Engineering-AI Firm That Cut Physical Test Counts Instead of Accelerating Simulation, Now Inside CoreWeave
- Monumo Deep Dive: A Cambridge Deeptech Company Redesigning Electric Motors with the Anser AI Engine, Tens of Millions of Multiphysics Simulations per Day, GBP 10.5M Raised in Total, and a Series A Being Raised
- NAVASTO: Moving Solver Insight Onto the Designer's Desk, and How Autodesk Bought That Entry Point
- Navier: Handing CFD Workflows to Agents, a Seed-Stage Bet on AI for Hardware Engineering
- Neural Concept: An EPFL Spin-out Betting on Neural Surrogates, with $138M Raised
- NVIDIA Deep Dive: Building the Full-Stack Physical-AI Infrastructure with PhysicsNeMo, Warp, Cosmos World Foundation Models and Omniverse
- Physical Intelligence Deep Dive: A $2.1B Bet on the GPT Moment for Robotics, With pi0.7 Self-Reporting 60-80 Percent Success on Unseen Tasks and Neither Pricing nor Revenue Disclosed for the Pi API
- Physical Superintelligence Deep Dive: A $58M Seed Bet on Virtual Physicists, With Emmy Targeting AI Data Center Multiphysics Optimization as Its First Commercial Proof Point
- PhysicsX: Pretrained Geometry and Physics for Industrial Engineering at a $2.4B Mark
- Prometheus: A $41B Mark, $18.2B Raised, and Not Yet a Product
- PTC Deep Dive: Creo/Onshape and the ThingWorx Digital Thread Power Physical AI, with NVIDIA Isaac Sim Bridging Robotics Design to Simulation
- Qianjue Technology Deep Dive: How a Body-Agnostic Embodied Brain Pairing the Predictive World Model Polibrain with Brain-Inspired Partitioning and Polynomial Representation Reached a Hundred Thousand Cleaning and Service Robots
- Quanscient Deep Dive: Cloud-Native and Quantum-Powered Multiphysics Simulation as the Data Engine for AI-Native Hardware Engineering
- Rescale Deep Dive: An HPC Operating System and Desktop Engineering Workbench in the Cloud, a $115M Series D, and Five Misread Facts
- Schrödinger Deep Dive: Physics-Based FEP+ Simulation and Machine Learning Power Molecular Discovery, with the 2026 Bunsen Agentic Co-Scientist
- Secondmind Deep Dive: A Cambridge Engineering-AI Firm Whose Bayesian Active Learning Cuts CAE Simulation and Test-Bench Runs by up to 90%, with Roughly USD 73M Raised and Mazda Both Customer and Investor
- Shichuang Technology Deep Dive: SupreCAST Cloud-Native Material-Processing Simulation and Industrial Agents Advance the CAE+AI Strategy
- ShonCloud (Shunyun) Deep Dive: Meshless Particle-Method CFD for EV and Nuclear Fluid Multiphysics
- Shuzhi Maijin Deep Dive: A Shenzhen Fully Self-Developed SNN-plus-LBM CFD Startup Claiming Second-Level Results on Hundreds-of-Millions of Grids on Consumer Hardware, With a Tens-of-Millions-RMB Angel Round From a Ministry-of-Finance-Affiliated Fund and an FPGA Migration Plan
- Siemens (incl. Altair) Deep Dive: The 2025 Altair Acquisition Unites Simcenter and Altair, with PhysicsAI Geometric Deep Learning and PhysicsAI Generate as the Physical-AI Core
- Silurian Deep Dive: A Generative Forecasting Transformer Earth-System Foundation Model for Weather Risk on Energy and Infrastructure Assets
- SimScale Deep Dive: A Cloud CAE Platform Built on Core-Hour Subscriptions, EUR 52M of Equity Kept Separate from EUR 25M of Debt, and Five Misread Facts
- Skild AI Deep Dive: A General Robot Foundation Model Incubated at CMU's Robotics Institute, With a USD 1.4B Series C at a USD 14B Valuation, a Conflicting Series B Size, and an ARR Figure That Is a Run-Rate Rather Than Revenue
- Songying Deep Dive: A Huawei-Cloud-Alumni Startup Building a Domestic Physical AI Operating System, With ORCA OS Positioning Against NVIDIA Omniverse, Three Incompatible Financing Narratives, and a Verifiable Third-Party Anchor in a CNY 20 Million Minority Investment
- Sudo Tech Deep Dive: Zero-Real-Data Simulation Training, the Fully In-House Sudo R1 Stack, and an Atomic-Skill Commercialisation Strategy from a Startup Valued Above CNY 20 Billion in 17 Months
- Suochen Technology: The Pain of Becoming a Physics-AI Platform from a CAE Vendor
- Synopsys (Ansys) Deep Dive: The 2025 Ansys Acquisition Turns the EDA Leader into a Silicon-to-Systems Engineering Powerhouse, with Ansys SimAI and the NVIDIA Physical-AI Partnership as the Core Physics-AI Levers
- TenFong Deep Dive: An Academician-Led Chinese Industrial Simulation Software Vendor Whose Official Timeline and Database Records Use Incompatible Round Names, With an Official Financing PDF and an Investment Database Both on File
- Tianfu (Nanjing Tianfu Software) Deep Dive: A 2011-founded Chinese Intelligent-CAE Veteran That Closed a Several-Hundred-Million-RMB Series C, Pivoted to an Industrial World Model in 2024, With Unreconcilable Financing and Valuation Records and No Disclosed Revenue
- Tongyuan Soft Control Deep Dive: MWORKS (Syslab/Sysplorer) as a Domestic MATLAB/Simulink Alternative, with MWORKS AI and the MOGICK Engineering Agent Advancing Home-Grown Physical AI
- Vinci AI: Turning Physics Simulation from a Checkpoint into a Continuous Constraint
- Westlake Robotics Deep Dive: Westlake University's First Robotics Spin-out Builds a Dual-Pretrain Brain-and-Cerebellum Architecture with WR1 Whole-Body Foundation Model and GAE Behaviour Expert
- World Labs: Fei-Fei Li's Spatial-Intelligence Unicorn, Acquired by AMD for $8.2B
- XVERSE Deep Dive: A Simulation-World Route to Physical AI Built on the XVERSE Model Family, an Edge-Cloud 3D Engine and the Physics-Based MotionGen Model
- SimWE: From In-House Solvers to a Physical-AI Engine, the Domestic CAE Leader and What Comes Next
- Yunjing Zhifang Deep Dive: A Peking-University-Incubated Cloud-Native GPU CAE Company with the Simversus Platform and the YJ-DORGP Small-Sample Intelligent Solver, Reporting 275 Percent Order Growth, 160 Percent Revenue Growth and Breakeven in 2025
- Zhongke Diwiji (Five Ages) Deep Dive: A CAS-Institute-of-Automation and Tsinghua Spin-out Whose FAM Embodied Foundation Model Uses Action Optical Flow for Cross-Embodiment Generalisation, Reaching 97% Success from 3 to 5 Demos, with EraBrain and Jiling Robots Serving Sinopec, Leapmotor and Midea
- ZWSOFT (688083) Deep Dive: Overdrive Kernel and ZW3D Wukong 2027 Underpin an Integrated CAD/CAE/CAM Stack, with a CAx Agent Platform as the Practical Route to Engineering AI
This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.