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

Executive Summary

NP Company (NP Co.) is a Paris-based, AI-native simulation software startup founded in 2025 by Emmanuel Menier (CEO) and Matthieu Nastorg (CTO), both former Inria / Paris-Saclay researchers in machine learning for simulation. The company is building pre-trained transformer-based "foundation models for physical engineering" that it claims deliver high-fidelity simulation results in seconds rather than days or weeks, across full assemblies. On 2 June 2026 it announced a EUR 6M pre-seed round led by Partech, with participation from the Peugeot family office and angel investors including Guillaume Lample and Cedric O (co-founders of Mistral AI), Florian Douetteau (founder of Dataiku), and Vincent Luciani (founder and CEO of Artefact).

The thesis is coherent and the founder pedigree is strong: Nastorg's peer-reviewed work centers on graph-neural-network Poisson solvers for faster CFD (DS-GPS, NeurIPS 2022 ML4PS; Psi-GNN, 2023; a multi-level GNN preconditioner, 2024), and Menier's on reduced-order models and ML-CFD hybridization at Air Liquide. However, the company's product narrative emphasizes transformers, a direction not yet reflected in any published architecture, white paper, model card, or benchmark methodology. The headline 1,000x speedup is a vendor claim demonstrated on Safran-organized industrial benchmarks with no disclosed comparison baseline, fidelity metric, or hardware; the 50,000x full-assembly figure is an explicitly stated future target, not a measured result. As of October 2026 the company is pre-product and pre-revenue, with no named customers, no pricing, and no disclosed headcount.

Company Overview

• Founded: 2025. Headquarters: Paris, France. Website: npco.ai.
• Founders: Emmanuel Menier (CEO) and Matthieu Nastorg (CTO). Both earned PhDs in AI for simulation at Paris-Saclay University and conducted research at Inria. Menier's doctorate (Inria / IRT SystemX, 2020-2024) focused on deep learning and dynamical systems, with an earlier stint as a research engineer at Air Liquide (2019-2020) building surrogate models for hydrogen-production plant simulation. Nastorg's research focused on ML acceleration of numerical simulation, especially Poisson solvers via graph neural networks.
• Origin story: the company "emerged from" / was "born out of" Inria, the French national research institute that also incubated Mistral AI. No IP-assignment, license, or formal spinout agreement has been disclosed.
• Pre-seed taxonomy note: the master list and press materials describe the round as "pre-seed." This site's phase taxonomy has no pre-seed value, so the record uses "seed" while the prose retains "pre-seed."

Technology: The Pre-trained Transformer Physics Foundation Model

NP Co.'s core claim is a new class of physics simulator powered by transformer models pre-trained on industrial physics. The company adapts the architecture behind modern large language models to physical simulation. Two design choices are stressed:

1. Pre-training, not per-customer fine-tuning. Unlike earlier AI simulators that required extensive customer-specific training before deployment, NP Co. says its models deliver value the moment they are deployed on a customer's infrastructure (including on-premise, which matters for data-sovereignty-sensitive aerospace / defense / energy buyers).
2. Full-assembly fidelity. Where legacy simulators resolve a single component in days or weeks, NP Co. claims results in seconds at the same fidelity across an entire assembly.

Crucial gap: the founders' published, peer-reviewed work is built on graph neural networks, reduced-order models, and Poisson-solver acceleration (Nastorg: DS-GPS, arXiv:2211.11763; Psi-GNN, 2023; Multi-Level GNN Preconditioner, 2024). The company's product story is transformer-centric. No architecture, training-data description, tokenizer, or benchmark protocol has been published that bridges the two. The "pre-trained industrial transformer physics foundation model" remains, as of this writing, an unverified thesis rather than a documented system.

Product and Commercial Model

No named product exists beyond the "NPco" brand. There is no public API, no demo, no model card, no pricing, and no stated deployment package. The stated longer-term roadmap includes (a) automated design tools that generate optimized architectures and (b) real-time operational simulators for industrial infrastructure. The commercial model (license, SaaS, on-premise appliance, or compute-hour) is undisclosed. Target sectors are aerospace, defense, energy, electronics, data centers, and automotive.

Competitive Landscape

NP Co. is the earliest-stage entrant in the AI-native simulation cohort covered by this site. For context:

• Well-funded peers with shipped or demonstrable products: PhysicsX (Series C, USD 300M), Vinci AI (Series B, USD 250M), Luminary Cloud (Series B, USD 72M), Neural Concept (~USD 138M, CAD-native GNN), BeyondMath, SimScale, Rescale, NAVASTO (acquired by Autodesk), Monolith AI (acquired by CoreWeave), Emmi AI (acquired by Mistral, May 2026).
• Direct narrative overlap: Emmi AI (Austrian, acquired by Mistral) pursued physics-AI models for the same aerospace / automotive / semiconductor verticals. NP Co. is, in effect, a fresh entrant into the exact space Mistral just absorbed.
• Platform giants: Ansys (SimAI), Siemens (Simcenter PhysicsAI), Dassault (SIMULIA), Cadence, COMSOL treat AI as a plug-in to an installed base.
• Strategic paradox: thirteen days after Mistral acquired Emmi (19 May 2026), two of Mistral's co-founders (Lample, Cedric O) personally wrote a cheque into NP Co. Three readings exist: (i) the market is large enough for several approaches; (ii) personal angel bets are distinct from corporate strategy and hedge against a fast-moving frontier; (iii) the Inria connection runs deep and the founders are trusted. Menier offered a pragmatic line: "Emmi was exploring options, and we were not on the market, having just come out of the round."

Financing

• Round: EUR 6M pre-seed, announced 2 June 2026.
• Lead: Partech (Paris-based global deep-tech investor; ~EUR 3B AUM).
• Participants: Peugeot family office; angel investors Guillaume Lample and Cedric O (Mistral AI co-founders), Florian Douetteau (Dataiku founder), Vincent Luciani (Artefact founder and CEO).
• Use of proceeds: expand the research team and continue developing foundational models; longer-term work on automated design tools and real-time operational simulators.
• Conflicts: secondary sources report the amount as USD 6.0M (FX approximation) and USD 7M (Gaebler, rounded / conflated); and one database labels the round "Seed" rather than "pre-seed." See DISCREPANCIES.

DISCREPANCIES

1. Phase taxonomy gap. The master list and press call the round "pre-seed"; this site's phase enum lacks a pre-seed value, so the record uses "seed" while prose retains "pre-seed."
2. Funding amount currency. Official EUR 6M (Partech release). Secondary: USD 6.0M (FX) and USD 7M (Gaebler, rounded). Not reconcilable to one figure without a stated FX date.
3. Round label. Partech / company: pre-seed. Gaebler database: Seed.
4. Mistral co-founders back a competitor. Mistral acquired Emmi AI on 19 May 2026; 13 days later Lample and Cedric O personally invested in NP Co., a same-sector competitor.
5. "Augur" prior name unverified. A single Chinese-language secondary source states NP Co. previously operated as "Augur." No English or primary source corroborates this.
6. 1,000x "demonstrated" speedup has no methodology. Safran-organized industrial benchmarks; no disclosed legacy-solver baseline, fidelity metric, hardware, or problem scale.
7. 50,000x is a target, not a result. The full-assembly 50,000x figure is described as a "clear path," not a measured outcome.
8. Safran as benchmark organizer vs customer. Press says benchmarks were "organised by" Safran; Safran is not confirmed as a paying customer.

Risks and Due Diligence Gaps

• Pre-product / pre-revenue: no customers, pricing, API, or model card; all speedup and fidelity claims are vendor statements.
• Undisclosed team size: company says "currently hiring"; no headcount public.
• Undisclosed IP / spinout terms: "born out of Inria" is narrative; no license or assignment disclosed.
• Architecture gap: founders' published work is GNN / ROM / Poisson; product story is transformer. No published bridge (white paper, architecture, training data).
• Breadth vs evidence: a single pre-trained model serving aerospace / defense / energy / electronics / data centers / automotive at full-assembly fidelity is extraordinarily broad for a pre-seed with no published evidence; generalizability is unproven.
• Validation inertia: aerospace / defense adoption faces rigorous certification and integration hurdles against entrenched solvers.

Overall Assessment

NP Co. is a high-conviction, low-evidence bet. The upside case rests on a credible founder pedigree (genuine ML-for-simulation researchers), a coherent "pre-trained physics foundation model" thesis, a strong seed-signal cap table (Partech plus Mistral / Dataiku / Artefact founders and the Peugeot family office), and a large addressable market (engineering simulation projected at ~USD 30B by 2030 per Dealroom). The downside case is that the transformer physics-FM architecture is undocumented and unproven, the 1,000x / 50,000x claims are methodology-free vendor statements, and the company is pre-product in a field where incumbents (Ansys, Siemens, Dassault) and well-funded AI-native peers already ship. The Mistral co-founders' investment is a notable signal but also a strategic paradox that the company has not fully resolved in public. We rate technology 6.5, business 3.5, funding 6.0, product 3.0, market 7.0.

References

• Partech press release, "NP CO RAISES 6M PRE-SEED TO BRING FRONTIER AI TO INDUSTRIAL ENGINEERING, BACKED BY MISTRAL AI FOUNDERS" (2 June 2026).
• Partech portfolio page for NPco.
• Dealroom news note on NP Company (June 2026).
• TechFundingNews, "Why did Mistral's co-founders back a EUR 6M funding round in AI simulation startup days after acquiring its competitor" (June 2026).
• Studio Global AI (Chinese), note on the "Augur" prior name (unverified).
• Emmanuel Menier curriculum (emenier.com).
• M. Nastorg et al., DS-GPS: A Deep Statistical Graph Poisson Solver for faster CFD simulations (NeurIPS 2022 ML4PS; arXiv:2211.11763).
• M. Nastorg et al., An Implicit GNN Solver for Poisson-like problems (2023); Multi-Level GNN Preconditioner for Solving Large Scale Problems (2024).
• Gaebler Ventures and startup.eu funding records.

Disclaimer: This report is compiled from public information for research purposes only. It is not investment, legal, or engineering advice. Speedup and benchmark claims are vendor statements and have not been independently verified. Figures may change as the company discloses more.