Emmi AI: From Linz Spin-out to a EUR 330M Mistral Acquisition in 17 Months

Executive Summary

Emmi AI was an Austrian deep-tech company founded in December 2024 as a spin-off from Johannes Kepler University Linz and the NXAI research initiative around Sepp Hochreiter. On May 19, 2026, Mistral AI announced a definitive agreement to acquire it. Sifted, citing previously undisclosed internal documents, reported that Mistral valued the business at up to EUR 330 million in a combined cash-and-equity deal. That is roughly twenty-two times the EUR 15 million seed round the company closed seventeen months earlier.

Four facts frame the case.

First, the technical work is real and unusually strong for a company that never existed as a standalone for more than sixteen months. The lineage is traceable: Universal Physics Transformers (arXiv:2402.12365, NeurIPS 2024), NeuralDEM (arXiv:2411.09678, published in Nature Communications Physics in November 2025), and AB-UPT (arXiv:2502.09692, TMLR), extended by a technical report on automotive and aerospace applications (arXiv:2510.15808). AB-UPT reports state-of-the-art accuracy from 33 thousand up to 150 million mesh cells, training on a single GPU in under a day, and inference straight from CAD geometry without a mesh.

Second, the acquisition was unusually fast and unusually strategic for Europe. Mistral had raised close to EUR 2.8 billion at an EUR 11.7 billion valuation before the deal, and bought Koyeb, a cloud deployment platform, in February 2026. Emmi was its second acquisition inside three months. Linz became an official Mistral office alongside Paris, London, Amsterdam, Munich, San Francisco and Singapore. The entire team of more than 30 researchers and engineers moved to Mistral's Science and Applied AI units, and Johannes Brandstetter became Vice President, AI for Science. Those figures are point-in-time as at the May 2026 deal and predate Mistral's Series D of 8 September 2026, which took cumulative disclosed equity to roughly EUR 5.8 billion and the post-money to more than EUR 21 billion; the companion Mistral report carries the current totals. Third, the benchmark claims were overtaken within months. Emmi's own October 2025 technical report claimed top position on DrivAerML, DrivAerNet++, AhmedML, SHIFT-SUV and SHIFT-Wing. A third-party paper published later, comparing ten methods on the same datasets with all models configured at roughly 10 million parameters, places AB-UPT behind GeoTransolver on DrivAerML surface and volume, and behind the newer SMART architecture on both DrivAerML and SHIFT-Wing. The "state of the art" label did not survive independent re-evaluation.

Fourth, the commercial model was thin and the product was one partnership deep. The clearest evidence is SIMCON's Cadmould AI Solver, launched March 18, 2026 as the first large engineering model for injection moulding, co-developed with Emmi and running on hundreds of terabytes of data. A second product, NeuralWing, addresses transonic wing aerodynamics. Everything else was a framework, a demo, or a research model.

Our view: this is a well-priced exit on a strong but short-lived research position. The valuation of up to EUR 330 million on EUR 15 million of seed capital is the kind of outcome most European deep-tech founders cannot access. But the read-through for the category is more informative than the read-through for Mistral: industrial AI's economics rest on who can generate and license the training corpus, and the fastest way for a frontier lab to buy that capability was to buy the team that had already solved it.

Company Overview

Emmi AI GmbH was founded in Linz in December 2024 by three co-founders:

• Johannes Brandstetter, Co-founder and Chief Science Officer. Previously at CERN and Microsoft, where he co-architected Aurora, described as the world's first AI foundation model for weather forecasting. Associate Professor at the JKU Institute of Machine Learning. After the acquisition he became Mistral's Vice President, AI for Science.
• Dennis Just, Co-founder and CEO. A serial founder and company builder, responsible for strategy and go-to-market.
• Miks Mikelsons, Co-founder and COO. Responsible for operations, delivery and cross-functional execution.

The origin is unusually concrete. Brandstetter studied physics in Vienna, did a PhD in high-energy physics at CERN, then worked at the JKU and the University of Amsterdam with mentors Sepp Hochreiter and Max Welling. Hochreiter, co-author of the LSTM, had returned to the JKU from Microsoft Research in Amsterdam in 2023. The company grew out of the NXAI initiative, an Austrian AI research lab focused on industrial applications, and the JKU's Institute of Machine Learning. The name Emmi AI is a deliberate reference to Emmy Noether, whose work on symmetry and conservation laws supplied the framing for the company's physics-first framing.

The public origin story is unusually sharp. The company states that while manufacturing cycles had accelerated tenfold, simulation technology had been stuck in the 1970s. That observation, not a research breakthrough, is what the founders say motivated the company.

Technology

The architecture lineage: UPT, then NeuralDEM, then AB-UPT

Universal Physics Transformers (UPT), arXiv:2402.12365, NeurIPS 2024, is the foundation. The problem it solves is that neural operators have historically been problem-specific: graph networks for Lagrangian simulation, grid models for Eulerian simulation. UPT removes grid and particle latent structures entirely, encoding either into a fixed-size latent token set, rolling dynamics forward in that latent space with inverse encoding, and decoding at arbitrary query points in space-time. The scaling property that matters commercially is that the latent space stays fixed in size regardless of mesh resolution.

Authors: Benedikt Alkin, Andreas Fürst, Simon Schmid, Lukas Gruber, Markus Holzleitner, Johannes Brandstetter. Affiliations: ELLIS Unit Linz at the JKU, Emmi AI GmbH, Software Competence Center Hagenberg, and MaLGa Center at the University of Genoa.

NeuralDEM, arXiv:2411.09678, is the proof that the framework transfers across physics. It treats the Lagrangian discretisation of DEM as an underlying continuous field while modelling macroscopic quantities such as occupancy, transport, residence time and mixing as auxiliary fields, and a multi-branch transformer couples the particle and fluid phases while allowing conditioning on macroscopic parameters like internal friction angle. The scale claim: systems with 500 thousand particles and 160 thousand CFD cells, stable and mass-conserving over 28-second rollouts of 2,800 timesteps, in real time on a single GPU. The full dataset is 3.2 TB, released by request; one trajectory and the weights are public. It was published in Nature Communications Physics on November 18, 2025, and the repository has 22 stars and 2 forks.

AB-UPT is the industrial-scale breakthrough, and it is the reason the company was worth EUR 330 million. Three innovations:

1. Anchored attention reduces the quadratic complexity of self-attention, which is what allows inference on hundreds of millions of mesh cells rather than tens of millions.
2. A multi-branch Transformer separates geometry, surface and volume representations while maintaining cross-attention between them, giving linear inference complexity while preserving the expressivity of full self-attention.
3. A divergence-free vorticity formulation as a hard architectural constraint rather than a soft training loss. Predicted vorticity fields are free of non-physical sources and sinks by construction. Because the architecture is a conditional neural field, local differential operators can be applied and conservation properties enforced at prediction time.

The scale claim is the differentiator: end-to-end inference on 9 million surface cells and 140 million volumetric cells on a single NVIDIA GPU, two orders of magnitude beyond prior methods. DrivAerML alone runs 140 million cells with Hybrid RANS-LES, the highest-fidelity approach routinely deployed in the automotive industry. Models can train on a single GPU in less than a day and predict industry-standard surface and volume fields within seconds. Critically, AB-UPT can simulate from CAD geometry alone, removing the need for the expensive meshing step, and the mesh-free geometry input is described as a uniformly or isotropically tessellated representation.

The Noether framework

The most strategically interesting asset is not a model but the framework. Noether is Emmi's open-source deep learning framework for engineering AI, pip-installable as emmiai-noether, built on transformer building blocks, hardware-agnostic across CPU, Apple Silicon MPS and NVIDIA GPUs, with built-in multi-GPU and SLURM cluster support. Its recipe catalogue includes an AB-UPT showcase on DrivAerML, a multi-dataset external aerodynamics recipe covering ShapeNet-Car, AhmedML, DrivAerML, DrivAerNet++ and Emmi-Wing across AB-UPT, UPT, Transformer and Transolver, and a fluid heat transfer recipe on the SIMSHIFT heatsink benchmark.

A disclosed performance benchmark: training AB-UPT on ShapeNet-Car takes 5.7 minutes on an NVIDIA H100 against 135 minutes on a MacBook Pro M3 Max. Licensing is the commercially sharp edge: the Emmi Open Research License, described as based on Mistral's MNPL, is free for research and piloting but requires a separate commercial license for production use.

Noether lists endorsements from research groups at ETH Zurich, JKU Linz, the University of Washington, Sorbonne University, the University of Pennsylvania and TU Munich, and names Luminary, Rescale and Simscale as platform partners. That partner list is itself strategically significant: Luminary is a direct competitor.

The three named models

• NeuralWing — transonic aircraft wing aerodynamics, built on the largest 3D transonic wing dataset, 30,000 steady-state CFD simulations of a 3D wing across four geometry parameters and two inflow conditions. Inputs are an STL geometry mesh, speed and angle of attack; outputs are pressure, friction and velocity fields plus integral lift and drag. Geometry meshes are generated in real time from the four design parameters in a differentiable manner. A third-party listing site states an accuracy rate of 99.5%; that figure is not stated in Emmi's own material and should be treated as unverified.
• NeuralMould — the first digital engineer for injection moulding, released February 2026. Trained on over a million transient trajectories across thousands of materials, letting engineers choose geometries, select materials and place injection gates arbitrarily to test scenarios in real time and optimise process KPIs. Built in partnership with SIMCON.
• NeuralDEM — industrial particulate flows, open-sourced November 2024, as described above.

Performance Benchmarks

Company-reported results

The October 2025 technical report, arXiv:2510.15808, extends AB-UPT to two Luminary Cloud SHIFT datasets:

• SHIFT-SUV, automotive: mean relative error below 1 percent for drag force prediction, near-perfect correlation for lift, full 3D fields accurately reproduced. Training on a single NVIDIA H100 converges in 13.5 hours; inference completes in under 34 seconds for full-vehicle meshes of 45 million cells.
• SHIFT-Wing, aerospace: across Mach 0.5 and Mach 0.85, R-squared = 1.00 for both drag and lift forces, matching ground-truth CFD within 2 percent error. Forces inferred directly from isotropically tessellated CAD geometry at sub-second inference.

Emmi's own claim, stated in its newsroom, is that it holds top performance across all major publicly available CFD datasets: DrivAerML, DrivAerNet++, AhmedML, SHIFT-SUV and SHIFT-Wing.

The independent evaluation that undercuts the claim

A later third-party paper comparing ten methods on DrivAerML, SHIFT-Wing and SHIFT-Crash, with every model configured at approximately 10 million parameters and evaluated by relative L2 error, produces the following averaged surface and volume test errors:

• PTV3: DrivAerML 31.33 surface, 44.31 volume; SHIFT-Wing 20.87
• Erwin: 28.52 surface, 26.67 volume; SHIFT-Wing 11.52
• Transolver++: 11.32 surface, 13.04 volume; SHIFT-Wing 7.67
• GeoTransolver: 10.33 surface, 10.90 volume; SHIFT-Wing 7.50
• AB-UPT: 10.76 surface, 8.83 volume; SHIFT-Wing 10.07
• SMART: 7.93 surface, 7.01 volume; SHIFT-Wing 7.95

Read carefully, AB-UPT is second on DrivAerML surface behind both GeoTransolver and SMART, first on DrivAerML volume, and last of the six on SHIFT-Wing behind every competitor including PTV3. SMART is the paper's own proposed architecture and it wins on DrivAerML. A separate benchmark aggregator reports yet another model, DIM-GP, achieving the lowest errors on AhmedML and both DrivAerML surface channels among ten methods including AB-UPT.

Two caveats in the other direction. The AB-UPT numbers in that comparison come from a run configured at roughly 10 million parameters for fairness, not from the industrial-scale configuration Emmi ships, where the model handles 140 million cells. And a benchmark aggregator's own summary of that comparison lists AB-UPT with lower errors than the paper's table shows, which is a discrepancy the aggregator does not explain.

The adjudication: the "state of the art across all major CFD datasets" claim was true as stated in October 2025 and was overtaken by the time of the later evaluation. The company was the first to demonstrate neural surrogates at 100-million-cell industrial scale with a hard divergence-free constraint. It was not, and could not remain, permanently the most accurate.

Product and Commercial Model

The commercial product story is a single deep partnership rather than a customer roster.

SIMCON and the Cadmould AI Solver is the clearest evidence. On March 18, 2026, German injection moulding CAE vendor SIMCON announced the Cadmould AI Solver, described as the world's first large engineering model for injection moulding, co-developed with Emmi AI. Built on a transformer architecture and trained on hundreds of terabytes of high-fidelity data, it delivers filling pattern, pressure, shear rate and temperature in seconds rather than hours, up to 1000 times faster than traditional numerical solvers, and demonstrates geometric generalisation to unseen part topologies without retraining.

The workflow design is instructive. SIMCON positions the AI Solver as a fast compass for early iteration and Cadmould Flex, the classical numerical solver, as the definitive map for final sign-off. The company also released a free interactive research preview of the filling-stage model in the browser and opened an Early Access Partner Program with priority access to shrinkage and warpage prediction. SIMCON CEO Bastiaan Oud on the collaboration: "Working with the Emmi team has been fantastic. AI is changing how engineers iterate on designs, and Emmi AI pairs deep technical credibility with a rare ability to move fast."

Brandstetter's framing: "Pre-trained large engineering models are a step-change for engineering workflows. With high predictive accuracy, they enable a new default for simulation — one model, any geometry, any material, any gate configuration."

The customer list is where the record is weakest, and this is a discrepancy worth stating plainly. Secondary coverage names ASML, Stellantis, Veolia and Helsing as customers in the context of the acquisition, and separately names ASML, Airbus, Safran and Siemens Energy as partners in Mistral's May 27, 2026 Physics AI launch. The ASML case is instructive: the widely reported figure of EUV lithography diagnostic time cut from hours to eight minutes is a Mistral computer-vision model detecting engraving defects, as confirmed by ASML CFO Roger Dassen on the April AGM. It is not a physics simulation result and should not be attributed to Emmi's surrogate models. No source was found confirming a named Emmi end customer at scale in physics AI deployment. Emmi's own About page claims it is "powering Fortune 500 engineering teams" with enterprise contracts signed in early 2025, but names none.

The Noether commercial licence is the sleeper business model. A research framework that is free for research and piloting but licensed for production is the standard pattern for seeding an ecosystem and capturing value at deployment. It is also the reason the Noether repository lists Luminary, a competitor, as a platform partner.

Competitive Landscape

Emmi competed in the narrowest and most technical position in the category, and that position is what made it attractive to a frontier lab.

Its direct competitors were PhysicsX (London, 300 million dollar Series C in June 2026 at a 2.4 billion dollar valuation, with NVIDIA, Applied Materials and Siemens participating), Luminary Cloud (72 million dollar Series B in September 2025 led by N47 with NVentures participating, shipping the SHIFT pre-trained models that AB-UPT was then benchmarked against), Neural Concept (EPFL spin-out, 138.1 million dollars raised, a customer base including GM, Safran, Airbus and four Formula 1 teams) and Monolith AI (acquired by CoreWeave in 2025).

The deeper competition was the incumbents: Ansys inside Synopsys, Siemens Simcenter and BeyondMath, Dassault SIMULIA, Altair PhysicsAI. And the platform layer: NVIDIA PhysicsNeMo, NVIDIA Omniverse, Decart's Oasis.

Emmi's differentiation was open science plus a hard physics constraint plus raw scale. AB-UPT was the only model in the category published in a peer-reviewed venue (TMLR) with an OpenReview process, released with weights on Hugging Face under CC-BY-NC 4.0, and NeuralDEM was published in Nature Communications Physics. In a category where most claims are vendor blogs, that constituted a real and marketable credibility asset.

Its weakness was the mirror image: an open research framework is precisely what a large platform can absorb. Mistral did not need to buy a product or a distribution channel. It needed the architecture, the training recipes, the team, and the credibility. That is what it bought.

Funding and Acquisition

The funding history is short and entirely disclosed:

• Seed, announced April 2025, EUR 15 million, co-led by 3VC, Speedinvest, Serena and PUSH VC. The JKU official release states it was the largest seed round ever for an Austrian start-up, and that it was the first time these Austrian venture firms had invested together. The stated use of funds was expanding engineering and research teams across Europe, scaling product development, and onboarding enterprise customers in global industrial markets.
• Acquisition by Mistral AI, announced May 19, 2026, definitive agreement. Terms undisclosed officially. Sifted, citing previously undisclosed internal documents, reported Mistral valued Emmi at up to EUR 330 million in a combined cash-and-equity transaction. Trending Topics, when asked, neither confirmed nor denied.

The EUR 330 million figure is the single most valuable data point in this report and it is an independent journalistic finding, not a company disclosure. It implies roughly a 22 times return on the seed round in seventeen months, and it makes this one of the largest exits in the history of the Austrian start-up ecosystem, as the law firms advising the transaction describe it.

The legal advisory structure also confirms the participants. Schoenherr advised venture investors Speedinvest, 3VC, Serena Ventures and PUSH Ventures; DORDA advised the founders and the company; Puttinger Vogl and Oberhammer advised NXAI; Jeantet advised on French law; Wolf Theiss, Binder Grösswang and Squadra Avocats advised Mistral.

Personalities: Brandstetter became Vice President, AI for Science at Mistral. Dennis Just and Miks Mikelsons took responsibility for client relations. The workforce was expected to grow from 30 to 50. Austria's start-up state secretary Elisabeth Zehetner framed it as the creation of a European industrial AI hub rather than a simple sale, and the economic affairs minister Wolfgang Hattmannsdorfer pointed to European sovereignty in AI.

Discrepancies

Twelve conflicts were identified and adjudicated.

1. Location, Linz or Vienna. Reuters and Mistral's own release say Linz. Several secondary outlets say Vienna, and one in fact is titled "Vienna-based Emmi AI" while describing the same company. Adjudicated as Linz. The JKU, the Austrian government, Schoenherr, the original 2024 founding and Emmi's own registered address at Peter-Behrens-Platz 2, Linz 4020, all agree. The Vienna references appear to be contamination from a separate AI company.

2. Founding date, December 2024 versus 2024 generally. All sources agree on 2024; the precise month is December. The company was 17 months old at acquisition.

3. Team size, 30 versus 35. Mistral's release says "more than 30 researchers and engineers"; several outlets and databases say approximately 35. The difference is rounding plus the recruitment that occurred between the two dates. Both are used, with 30+ as the floor and 35 as the higher estimate.

4. Acquisition value, undisclosed versus up to EUR 330 million. Both are true. Mistral did not disclose terms. Sifted obtained internal documents giving a valuation of up to EUR 330 million, and a separate independent audit reportedly produced the same figure. Treat EUR 330 million as a well-sourced media report, not an official disclosure, and note the "up to" qualifier.

5. Whether this was an acquihire. Serena partner Guillaume Decugis said it was "definitely not" an acquihire, and the entire team plus the founder leadership was retained with a new office and a new VP-level title for Brandstetter. Sifted's analysis of the pattern: "A young company like Mistral, very highly-valued, which makes a bet and accepts to acquire a startup at a premium price and very rapidly, just one year after its seed — this is the kind of cycle you see in San Francisco rather than in Europe." Adjudicated as a genuine capability acquisition, not a acquihire.

6. ASML's eight-minute lithography diagnostic. This is a Mistral computer-vision result for defect detection on EUV machines, confirmed by ASML's CFO, not a physics-simulation result from Emmi's surrogate models. Several secondary outlets attached the number to the Emmi acquisition narrative. Rejected as an Emmi physics AI outcome.

7. Customer attribution. ASML, Stellantis, Veolia and Helsing appear in coverage as Mistral customers. Airbus, Safran and Siemens Energy are named by Mistral as Physics AI partners. No source was found naming a confirmed Emmi physics-AI end customer. Emmi's own commercial traction record, independent of its partner and framework, is thin and largely undisclosed.

8. Funding total: EUR 15 million (≈ USD 17 million; some databases list USD 17.1 million). This is not a contradiction but a currency and conversion-date artefact. EUR 15 million is the official figure from JKU and Emmi; the USD 17 million and USD 17.1 million figures are database conversions, and their USD 0.1 million difference comes from a different conversion date.

9. Seed round date, April 1 versus April 2025 versus May 2025. Databases variously record April 1, 2025, April 2025 and May 2025. The JKU announcement and Emmi's own page say April 2025; the Austrian start-up agency says April 2025. Adjudicated as April 2025.

10. Investors. The four venture investors — 3VC, Speedinvest, Serena and PUSH VC — are confirmed by the JKU release, by the law firms and by Emmi's own page. Some database listings add a long tail of funds including 645 Ventures, NFX, QED, Playfair, Primo and Sterling Equity. Only the four are verifiable; the long tail is unverified and is likely a syndicate-aggregation artefact.

11. "State of the art across all major CFD datasets." True as published in October 2025, overtaken by the time of independent later evaluation, where AB-UPT placed second on DrivAerML surface behind GeoTransolver and SMART, and last on SHIFT-Wing among the six compared models. The most defensible formulation is: Emmi was the first to reach industrial mesh scale with a hard physics constraint, and its lead was measured in months.

12. The 99.5 percent NeuralWing accuracy figure. Appears on third-party software listing sites but not in Emmi's own material, which does not state a single accuracy percentage for NeuralWing. Treated as unverified.

Risks

The company no longer exists as an independent entity. Every question about Emmi's standalone strategy, pricing, revenue or product roadmap is now a question about Mistral's industrial division. Any future assessment of the technology must be made through Mistral's disclosures, which are materially less granular than a venture-backed company's would be.

The benchmark lead was not durable. Independent evaluation placed AB-UPT behind two newer architectures within months of publication, and a second independent benchmark aggregator reports a further model ahead on two of the same datasets. The technical moat was an execution lead, not a structural one.

The "up to EUR 330 million" qualifier. The reported figure is a ceiling, the cash-equity split is undisclosed, and no regulatory filing substantiates it. The actual proceeds to shareholders were lower.

The commercial model was one partnership deep. SIMCON's Cadmould is the only documented revenue-generating deployment. A framework-plus-demo-plus-partner business is not a business, and it is why the company was acquirable rather than valuable on its own terms.

Team concentration risk. The core capability sat in a 30-person group around Brandstetter and Alkin at the JKU's ELLIS Unit. Alkin, the first author of UPT, NeuralDEM and AB-UPT, left Emmi in 2025 for Project Prometheus, a Zurich-based company, where he is listed as Founding Member of Technical Staff. A substantial portion of the research authorship did not come to Mistral. The architecture, weights and recipes did.

European AI sovereignty is a policy tailwind, not a business moat. The acquisition is repeatedly framed in terms of European sovereignty and compliance-friendly deployment for defence-adjacent data. That framing helps in public-sector and regulated tenders, and does not resolve accuracy or adoption.

Scoring

• Technology, 8.0. Three papers in strong venues including NeurIPS 2024, TMLR and Nature Communications Physics, a demonstrated jump to 140-million-cell inference on one GPU, a hard divergence-free architectural constraint rather than a soft loss, mesh-free CAD inference, a transfer learning recipe across a Lagrangian-Eulerian boundary, and an open framework that beats a MacBook Pro by more than twenty times on training time. Deducted for the lead not being durable, the industrial-scale configuration being outside the fair-comparison regime, and the departure of the lead author to another company.
• Commercialization, 6.0. One deep partnership that produced a genuine industry-first product with a named CAE vendor, a documented 1000x speedup claim, a real licensing model for the framework, and a reference customer list in the parent's hands. Deducted for a single documented commercial deployment before acquisition, no disclosed revenue, no independently confirmed end customer for the core physics AI, and a strategy that pointed at the framework rather than a repeatable sales motion.
• Funding, 7.5. EUR 15 million seed at a record for Austria, four credible European deep-tech funds, and an exit at up to EUR 330 million, roughly 22 times the seed in 17 months, validated by independent internal documents and by two law firms describing it as one of Austria's largest exits. Deducted only because the figure is a reported ceiling rather than a disclosed number and the actual shareholder proceeds are unknown.
• Product maturity, 6.5. A coherent ladder from research framework to named production models to a partner-shipped product: Noether, then NeuralWing and NeuralDEM, then NeuralMould, then the Cadmould AI Solver shipped by a CAE vendor with a free preview and an early access program. Deducted for the 16-month operating window, the absence of a mature platform layer, the Open Research License that blocks production use without a separate negotiation, and the fact that the flagship product shipped through a partner rather than directly.
• Market position, 6.5. The strongest research credibility in the category with peer-reviewed publication and open weights, and a technical position that a frontier lab judged worth up to EUR 330 million. Deducted for a business that was a research group rather than a competitor with revenue, the fastest possible exit rather than a durable franchise, and the risk that a category of this type consolidates into two or three platform owners rather than remaining a field of independent vendors.

References

1. Mistral AI, "Emmi joins Mistral to accelerate the AI-native industry" (2026-05-19)
2. Sifted, "Exclusive: Internal docs reveal Mistral valued M&A target Emmi at up to EUR 330m"
3. Sifted, "Mistral strikes second M&A deal in months with Austrian AI startup Emmi" (Daphné Leprince-Ringuet)
4. Johannes Kepler University Linz, "Emmi AI, a JKU Spin-Off Company, Receives Austria's Largest Seed Funding of EUR 15 Million" (2025-04)
5. Johannes Kepler University Linz, "Strategic Cooperation: Mistral AI Acquires JKU Spin-Off Emmi AI" (2026-05)
6. Reuters, "Mistral AI buys Austrian physics AI startup in industrial push" (2026-05-19)
7. Schoenherr, "Austria: Schoenherr advises VC investors in Emmi AI on landmark exit to Mistral AI"
8. Alkin et al., "Universal Physics Transformers: A Framework For Efficiently Scaling Neural Operators", arXiv:2402.12365, NeurIPS 2024
9. Alkin et al., "NeuralDEM - Real-time Simulation of Industrial Particulate Flows", arXiv:2411.09678, Nature Communications Physics (2025-11-18)
10. Alkin et al., "AB-UPT: Scaling Neural CFD Surrogates for High-Fidelity Automotive Aerodynamics Simulations", arXiv:2502.09692, TMLR
11. Alkin et al., "AB-UPT for Automotive and Aerospace Applications", arXiv:2510.15808
12. Emmi AI, "AB-UPT Sets New Benchmarks Across Major CFD Datasets" (2025-10)
13. Emmi AI, "Noether Framework: Industry-Grade Deep Learning for Engineering AI, Now Open Source"
14. Emmi AI, "NeuralMould: Our First Digital Engineer for Injection Moulding"
15. SIMCON / Business Wire, "SIMCON Unveils World's First Large Engineering Model for Plastic Injection Moulding" (2026-03-18)
16. Plastics Today, "New Software Slashes Injection Molding Simulation Time" (2026-03-19)
17. Independent third-party evaluation comparing ten methods on DrivAerML, SHIFT-Wing and SHIFT-Crash, arXiv:2610.03363
18. Invest in Austria, "Mistral AI acquires Austrian AI startup Emmi AI"
19. Leaders League, "Mistral AI acquires Emmi AI in landmark Austrian deep-tech exit"
20. Emmi AI GitHub organisation: Emmi-AI/noether, Emmi-AI/AB-UPT, Emmi-AI/NeuralDEM; Hugging Face EmmiAI

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