Inductiva Deep Dive: A Closed Case in Open-Solver Orchestration, and Eight Premises That Must Be Corrected

Executive Summary

This report covers a company that no longer operates. The legal entity Inductiva Research Labs, S.A. was registered in Porto, Portugal under NIPC 516243799, incorporated on 6 January 2021 as a spin-off of the Faculty of Engineering of the University of Porto (FEUP), and stated on its official contact page that it would "permanently cease operations" with effect from 31 January 2026. On 22 October 2026, the research cut-off date of this report, the company website inductiva.ai, the documentation site docs.inductiva.ai and the web console all returned HTTP 404. A Portuguese trade report published on 7 July 2026 described a workforce reduction of roughly 80 percent, phrased as "Word on the street" with no named source; that is a media formulation rather than a company announcement. No acquisition, listing, continuation or re-formation event appears anywhere in the public record.

What follows is a record of a closed case rather than a live coverage, and it is written in that register. The value of studying a company like this is not that it was successful. It is that its positioning, its funding, its technical choices and its exit mode are all fully documented, and each of them maps onto a live question in the current market for physics AI infrastructure. This report therefore reads the case for what it was, states the evidence for every claim, and marks clearly what remains unknown.

Eight premises about this company circulate widely, and each one has to be corrected before the rest of the record can be read correctly.

Premise one is the country. The company was widely placed in Zurich, Switzerland. The official press releases from InvestPorto and say-u describe it as founded in January 2021 and based in Porto; the Portuguese commercial register, compiled by North Data, gives NIPC 516243799 at Rua das Andresas, number 88, 3.5, 4100 Porto; the official GitHub organisation lists its location as Portugal. Three independent sources agree. One honest limitation belongs here: the Swiss federal commercial register could not be queried during this research, so the absence of a Swiss entity is established by three-way consistency rather than by a direct negative search in the Swiss registry.

Premise two is the product category. The company was widely described as a GPU-native simulation cloud. Its own wording is different: "Inductiva is a hybrid High-Performance Computing (HPC) platform built to simplify workflows and scale simulations", with the four selling points Simulation-Optimized, Programmatic Control, Hybrid Ready and Smart Scaling. The platform orchestrates upstream open-source solvers and does not develop a solver of its own. GPU acceleration comes from GPU ports that already exist in upstream projects such as CaNS, DualSPHysics, CP2K, GROMACS, AMR-Wind and Quantum ESPRESSO, together with cloud GPU machine types. There is no public source for hand-written CUDA or HIP kernels, for a self-developed matrix-free linear algebra stack, or for any proprietary GPU backend.

Premise three is the AI component and the ownership of the solver. The AI component is consistently overestimated in secondary write-ups. Four specific capabilities, namely a geometry surrogate model, Bayesian optimisation, active learning and inverse design, were checked one by one against the official documentation and no public source was found for any of them. What was actually shipped is large-scale synthetic data generation and inference with user-supplied models, plus the publication of datasets and research repositories. Alongside it, the assumption that the platform wrote a CFD solver of its own is also routinely carried forward. The official value proposition is "The power of open-source" and "tens of OS simulation packages", and the proprietary work sits in the orchestration and data layer.

Premise four is the founders. Three founders are documented: Luis (Morais) Sarmento as chief executive, Hugo A. Penedones as chief technology officer, and Clara Goncalves as chief operating officer. No public source links this company to any founder named Thomas H. or Kyriakos Antoniadis. The Thomas figures that circulate in the neighbouring Swiss scene belong to other companies, namely Thomas Leurent at Akselos in Lausanne and Thomas von Tschammer at Neural Concept, also in Lausanne.

Premise five is the exit. The company was reported as having been acquired by a large technology company. The source of that claim is a different company with a nearly identical name, Inductiv without the final letter a, which was acquired by Apple in May 2020 for machine learning data correction work. For Inductiva itself no acquisition source exists. What the record shows is a liquidation-style termination of operations.

Premise six is the solver and GPU support matrix. The official documentation contains no occurrence of JAX, SU2, PETSc or FEniCS. Elmer is supported and has an official guide. OpenFOAM is supported in both the ESI and Foundation distributions, and has no official GPU support in the Inductiva documentation.

Premise seven is the customer list. Heineken Group, NOS Telecommunications and Coca-Cola appear in a press release by the Portuguese partner DareData and are DareData's clients. The public Inductiva case list consists of academic researchers, a joint offshore wind project and ecosystem partners.

Premise eight is the set of borrowed figures. The USD 1B valuation belongs to Rescale in 2022. The USD 300M raise and the USD 2.4B valuation belong to PhysicsX in 2026. The statement that the company was still trading and taking orders in 2026 is not supportable, since the official position is that operations ceased on 31 January 2026 and the website, the documentation and the console all return 404 as of 22 October 2026.

Funding was thin. The only confirmed equity raise is a EUR 2.25M seed round in December 2023, led by Iberis Capital with Shilling and six named angels. A separate EUR 243,403 grant received in July 2023 through OX2 AB has an unclear character and is not added to the equity total. Post-money valuation, cap table, revenue and ARR are all without a public source. The database figure of USD 2.45M differs from the official EUR 2.25M in currency and timing.

The product was unusually transparent for a company of its size, and that transparency is the most useful material in the record. The pricing model is documented: prepaid credits, a one-time platform fee percentage that is itself undisclosed, pass-through of cloud compute, storage and data transfer charges, and a per-run orchestration fee of USD 0.02 for Individual, USD 0.01 for Enterprise and USD 0.005 for Academia, effective 1 December 2025. Quotas, per-hour spend caps and instance counts are published for three tiers, and the documentation itself uses two inconsistent tier naming systems, which is recorded as a conflict below.

The performance evidence is thin and entirely vendor-reported. The single checkable GPU benchmark is a stable atmospheric boundary layer case in CaNS, where a g2-standard-4 with one NVIDIA L4 took 25h03m and two NVIDIA H100 took 1h36m, a stated factor of 15.65. That is a comparison across different GPU models rather than a same-model scaling result, and it carries no grid size, no grid-independence check and no equivalence comparison against a CPU baseline. The separate headline of 10 to 100 times acceleration on the machine selection page has no case attached to it and is not usable as a performance statement.

On positioning within the market, Inductiva sits in the open-source solver orchestration quadrant, not in the GPU-native or AI-surrogate quadrant. Its two database-listed competitors are Rescale and SimScale, both cloud platforms, which is itself an informative signal about how the market classified it. Against them the difference is structural: Rescale and SimScale sell commercial CAE software licences delivered from the cloud, while Inductiva sold compute and orchestration for upstream open-source solvers, with no proprietary solver, no commercial licence revenue and no accuracy certification system.

This report's assessment, stated plainly and without editorialising, is that the technical and product work was real and the engineering transparency was above average for a seed-stage company, while the commercial base was never documented and the capital was too thin for a business whose cost base is cloud compute. The exit mode is the part that carries the most generalisable lesson: termination by liquidation rather than by sale leaves investors holding a residual claim rather than a continuing business, leaves customers without continuity of data and reproducibility, and leaves the industry with a data point on the open-source-orchestration route rather than with an acquired asset. The site schema's phase enumeration therefore tags this report as closed rather than as a funding stage, because the company no longer operates and the material fact is the termination, not the round it last closed.

Company Overview

The legal entity is Inductiva Research Labs, S.A., with the brand and word mark Inductiva.ai. It was incorporated on 6 January 2021 as Inductiva Research Labs Lda with share capital of EUR 5,000, and directors Luis Morais Sarmento and Hugo A. Penedones, and was converted to a sociedade anonyme on 11 August 2023, at which point the capital was EUR 6,375 and there were five shareholders. On 23 October 2023 the capital was increased to EUR 60,000 and Luis Miguel Araujo joined the board. The registered address is Rua das Andresas, number 88, 3.5, 4100 Porto, Portugal, the identifier is NIPC 516243799, and the EUID is PTIRNMJ.516243799. The registered activity codes are computer consultancy, computing infrastructure and data processing and hosting, research and experimental development in natural sciences and engineering, and specialised training in computer science consultancy. The auditor of record since 22 January 2024 is Pinto Leite, Machado Vaz and Associado - SROC Lda. The company carried the FEUP spin-off label from 2021. Financial statements for FY2021 through FY2024 were filed with the Portuguese registry but are not published in their text; only the filing is on record.

The sequence of events is short and can be reconstructed with reasonable confidence.

• 6 January 2021: incorporation as a limited liability company, capital EUR 5,000, two directors.
• 2021: FEUP spin-off status granted.
• December 2021: Clara Goncalves joins; she became chief operating officer in June 2023.
• 24 July 2023: a EUR 243,403 European Union grant received through OX2 AB, under the project "Atmospheric Physics and Turbulence for Wind Energy".
• 11 August 2023: conversion to a sociedade anonyme, capital EUR 6,375, five shareholders.
• 23 October 2023: capital increased to EUR 60,000; Luis Miguel Araujo joins the board.
• December 2023: EUR 2.25M seed round closed, led by Iberis Capital.
• July 2024: an accelerator or incubator transaction recorded by PitchBook, with amount and terms not disclosed, associated with the Unicorn Factory Lisboa 2024 scale-up list where the company appears seventh.
• 17 July 2024 and 17 July 2025: FY2023 and FY2024 annual accounts filed, text not published.
• 2 September 2025: the word marks "Inductiva.ai" and "Inductiva Research Labs" registered.
• 25 September 2025: official API release v0.18.0, with updated simulator combinations and credit purchases routed through Stripe.
• September 2025: the company publishes its own results post for the first OpenFOAM HPC Challenge, held on 1 and 2 July 2025 in Vienna at the twentieth OpenFOAM Workshop.
• November 2025: a public case in which Tiago Lucas Frutuoso ran more than 33,000 OpenFAST high-fidelity simulations in three weeks for a fatigue assessment at the Tocha wind farm.
• 31 January 2026: the official statement that operations cease permanently.
• 7 July 2026: a Portuguese trade report describing a workforce reduction of roughly 80 percent.

The three founders are documented through official releases, the Portuguese register and conference material. Luis (Morais) Sarmento, chief executive and co-founder, studied electrical and electronic engineering at the University of Porto in 1999, took an artificial intelligence master's in 2004 and a doctorate in informatics in 2010, and worked at Amazon, Google, PT and TonicApp. Hugo A. Penedones, chief technology officer and co-founder, studied informatics and computer engineering at the University of Porto from 2000 to 2005, was at EPFL from 2008 to 2011, and was at Microsoft Bing and DeepMind from 2015 to 2019, where he describes himself as one of the founding research engineers on AlphaFold, naming Marek Barwinski, Richard Evans, Laurent Sifre, James Kirkpatrick and Andrew Senior; he is a co-author on a 2024 Nature paper associated with that year's Nobel Prize in Chemistry. Clara Goncalves, chief operating officer and co-founder, studied agricultural engineering at the University of Porto and worked at Fraunhofer Portugal, at the innovation centre of the University of Minho medical school and as an executive director at UPTEC, the science and technology park operator.

A point of record concerns the academic affiliations. EPFL is confirmed through Penedones. No public source links any Inductiva person to ETH Zurich or to Oxford. The real academic ties are the University of Porto in electrical engineering, informatics and agronomy, EPFL, Fraunhofer Portugal, UPTEC and the FEUP incubation ecosystem.

Other named roles include Head of Product Execution for Claudia Da Silva, software engineer Paulo Barbosa, Rodrigo Valerio as a co-author of official OpenFOAM and cloud HPC reporting, and Eunice Oliveira, who gave an official CFD webinar on 11 July 2025. Sergio Santos represented Inductiva Labs at an INESC TEC ATTRACT event in 2024 on the path to European Union funding.

Headcount has five distinct published readings and they are all pre-closure snapshots. RocketReach shows 15 with a count of 13 on its own profile. PitchBook records 17 as of 2025. BounceWatch shows 17 and StartupSeeker 16. The LinkedIn company page gives the band 11 to 50. A database note refers to the remaining team as a small fraction of the previous size. The sequence of headcount over time has no public source.

Common Misreading Warning

The eight items below are the claims that circulate about this company and are wrong, or wrong in a way that changes the analysis. The correct reading is given in each case.

Misreading one: the company was Swiss, based in Zurich. The correct reading is that the entity is Inductiva Research Labs, S.A. in Porto, Portugal, a FEUP spin-off incorporated on 6 January 2021, with three independent sources agreeing: the official InvestPorto and say-u releases, the Portuguese commercial register compiled by North Data under NIPC 516243799, and the official GitHub organisation whose location is Portugal. The genuine Swiss companies in the physical-AI scene are Akselos in Lausanne and Neural Concept, an EPFL spin-out also in Lausanne; both should be filed separately, and neither has any equity or personnel link to this company. The limitation stated above applies: the Swiss register could not be queried directly.

Misreading two: it was a GPU-native simulation cloud. The correct reading is the official positioning, a hybrid HPC platform with a Python API that orchestrates open-source solvers. GPU capability is a property of the selected upstream solver and the selected cloud machine type, not of proprietary code written by this company.

Misreading three: it wrote its own CFD solver. The correct reading is that it packaged upstream official releases of OpenFOAM, CaNS, DualSPHysics, CP2K, GROMACS, AMR-Wind, Octopus and GX in its own container image repository, and its own value proposition was "The power of open-source" and "tens of OS simulation packages". The proprietary work sits in the orchestration and data layer: the job queue, automatic resubmission on preemption, spot and on-demand scheduling, elastic machine groups and MPI clusters, the Python API, the templating workflow for synthetic data, and the console.

Misreading four: it was an AI simulation platform or an AI physics engine. The correct reading is that the AI content in the product is synthetic data generation and inference with user-supplied models, with no public source for any built-in geometry surrogate, Bayesian optimisation, active learning or inverse design. The narrative of AI for science infrastructure survives, but it describes a data layer rather than a model layer.

Misreading five: it was acquired by a large technology company. The correct reading is that no acquisition source exists for this company, while the official statement says operations cease permanently from 31 January 2026 and a July 2026 report describes a reduction of roughly 80 percent of the workforce. The acquisition claim traces to Inductiv without the final letter a, acquired by Apple in May 2020, a machine learning data correction business founded at the University of Waterloo, the University of Wisconsin-Madison and Stanford University. The two are distinguishable on four points: the final letter of the name, the business, the acquirer, and the date.

Misreading six: it supported JAX, SU2, PETSc and FEniCS. The correct reading is that none of these appears anywhere in the official documentation. Elmer is supported and has an official guide covering ElmerSolver, ElmerGUI and ElmerGrid.

Misreading seven: it ran OpenFOAM on GPUs. The correct reading is that OpenFOAM has no official GPU support in the Inductiva documentation, while the GPU-capable set is CaNS, DualSPHysics, SPlisHSPlasH, CP2K, GROMACS, AMR-Wind, Quantum ESPRESSO, NWChem and GX, which carry a GPU image suffix in the container repository.

Misreading eight: Heineken, NOS and Coca-Cola were its customers. The correct reading is that those are clients of the partner DareData. The public Inductiva case list is academic researchers, the Voltron joint project and ecosystem partners.

Three further figures circulate under this company's name but belong to other companies, and one further statement is unsupportable. The USD 1B valuation belongs to Rescale in 2022. The USD 300M raise and the USD 2.4B valuation belong to PhysicsX in 2026. The statement that the company was still trading and taking orders in 2026 is not supportable, since the official position is that operations ceased on 31 January 2026, and the website, the documentation and the console all return 404 as of 22 October 2026.

Exit Mode: Liquidation Rather Than Sale

The termination deserves its own treatment because it is the part of this case that generalises best.

For the investors, a liquidation and a sale are entirely different outcomes. A sale at any valuation would have produced a realised return on a EUR 2.25M seed and a continuing claim on a business with a customer base. A termination of operations produces a residual claim on the balance sheet of a company whose financial statements for FY2021 through FY2024 are filed but not published, and whose liquidation procedure, creditors, employee severance compliance position and outstanding cloud provider balances are all without a public source. The asset base itself is unclear: the ownership, maintenance status and inheritability of the GitHub repositories inductiva/kutu, inductiva/inductiva, inductiva/task-runner, inductiva/datasets-generation and inductiva/meshnets after the closure were not confirmed during this research. Whether those repositories are maintained, archived or taken over by third parties is a question that applies to every open-source company that closes, and here it has no documented answer.

For customers, the practical consequence is discontinuity of reproducibility. The company's specific proposition was that a research team could script a workflow, run it on a managed machine group, generate a templated dataset, and reproduce it later. A closure breaks that loop at two points: the console and the stored simulation data are gone with the 404s, and the orchestration layer that made a run reproducible is not something a customer can rebuild from the container images alone. For the documented users, an academic researcher running 33,000 OpenFAST simulations at the Tocha wind farm and a joint project running hundreds of hull CFD cases for the Voltron CTV-26 vessel, the artefacts of that work are the durable output, and the tooling is not.

For the industry, the case is a data point rather than a verdict. It is a complete, documented instance of the open-source orchestration route: aggregate many upstream solvers, expose them through one API and one console, run them on rented cloud machines, and sell the orchestration and the compute. It shows the route working technically, with a published pricing model, published quotas and a reproducible benchmark; it also shows the route's central difficulty, which is that the differentiated surface is the orchestration layer while the cost base is the cloud bill, and that the revenue has to come from an orchestration fee or a platform fee percentage rather than from software licences. Whether the difficulty is fatal for the model in general, or fatal for this particular execution, cannot be settled from the record, and this report does not attempt to settle it.

Technology

The technical characterisation that governs everything else is this: Inductiva orchestrates open-source solvers, it does not write them. The official home page states: "Inductiva is a hybrid High-Performance Computing (HPC) platform built to simplify workflows and scale simulations. Whether on the cloud, on-prem, or anywhere in between, Inductiva brings everything you need to innovate, into one place." The four selling points are Simulation-Optimized, which means running pre-installed solvers or a user's own tools; Programmatic Control, which means scripting the workflow from Python; Hybrid Ready, which covers cloud, own hardware or a mix; and Smart Scaling, which means using benchmarking to choose a cost-performance point. A 2025 conference paper describes the company as a worldwide deep-tech company offering an integrated computational platform, states the "One language for everything: Python" and "Access to tens of OS simulation packages", and gives the value proposition of using AI-based methods to speed up simulation times and to explore design space efficiently.

The solver inventory is broad and was verified entry by entry. The official documentation and the container image repository inductiva/kutu, which holds 45 upstream solver image directories, together with the client repository inductiva/inductiva, which holds 39 simulator modules, cover OpenFOAM in the ESI 25.06, 24.12, 24.06 and 22.06 lines and the Foundation 13, 12 and 8 lines, OpenFAST, CaNS, AMR-Wind, DualSPHysics, SPlisHSPlasH, REEF3D, SWAN, SWASH, XBeach, SCHISM, GROMACS, CP2K, Quantum ESPRESSO, NWChem, FDS, Elmer, CM1, OpenSees, FVCOM, COAWST, GX, Octopus and custom_image for arbitrary containers. The earlier API 0.6 overview page listed fourteen of these. MPI-enabled solvers are OpenFOAM, REEF3D, SWASH, XBeach and OpenSees. Users commonly assume that JAX, SU2, PETSc or FEniCS are available; the full documentation search returns no occurrence of any of the four, while Elmer is documented with a full workflow guide.

The GPU story is the one most often overstated. Inductiva provides no evidence of writing CUDA or HIP kernels, of porting a solver to a GPU backend, of a matrix-free or self-developed high-performance linear algebra layer, or of any proprietary GPU backend at all. What it does provide is the packaging of GPU ports that upstream projects already ship, combined with the scheduling of cloud GPU machine types. The API-level GPU support flag covers CaNS, DualSPHysics, SPlisHSPlasH, CP2K, GROMACS, AMR-Wind, Quantum ESPRESSO, NWChem and GX. The accelerator-optimised machine families offered are the G2, A2 and A3 families, and the GPU instances in the documentation examples include a g2-standard-16 with one NVIDIA L4. OpenFOAM sits outside this set, because it has no official GPU support in the Inductiva documentation.

The performance record has to be read with the test conditions attached to every figure.

The one checkable GPU benchmark is a CaNS stable atmospheric boundary layer case. The baseline is a g2-standard-4 with one NVIDIA L4, which took 25h03m. The accelerated configuration is two NVIDIA H100, which took 1h36m. The stated factor is 15.65, with 15.7 appearing in a second place. The conditions caveat is decisive: this is a comparison across different GPU models rather than a comparison of instance counts on the same model, it is a vendor self-report, and no grid size, no grid-independence verification and no equivalence comparison against a CPU baseline is provided. It cannot be used as a general CFD acceleration factor.

Two further vendor-reported items exist. A Quantum ESPRESSO benchmark against Fugaku is published under the heading "Can Inductiva Beat Fugaku", with the detailed conditions held in the official benchmark directory and no third-party reproduction. An AMR-Wind stable atmospheric boundary layer case on a 512-cubed grid is described as requiring roughly 36,000 seconds and roughly 150 GB of memory in its CPU scaling benchmark.

The separate headline is the machine selection page, which describes GPU-accelerated workloads as running 10 to 100 times faster. That is a generalised vendor description of upstream solvers with no case bound to it, and this report does not use it as a performance conclusion.

On precision, the record is empty, and this needs stating plainly because it constrains what can be claimed. No public source exists for any systematic statement about the equivalence of results produced on the platform and results produced by the original or certified solver, whether bit-wise or within a stated tolerance, and no cross-solver validation report appears in the documentation. In the OpenFOAM HPC Challenge context the official emphasis falls on hybrid readiness and on the use of upstream official releases, without any accuracy comparison table. Any assertion that results on the platform are equal or superior to those of the original solver is unsupported by a source and is therefore not made in this report.

The AI content was checked item by item and is narrower than the narrative. A built-in geometry surrogate model has no public source; the closest thing in the documentation is inference with a user-supplied model, with DiffDock used as the example. Parameterised optimisation and Bayesian optimisation have no public source for a built-in implementation, since the "Explore designs running script variations" workflow is a scripted batch sweep over variants rather than a Bayesian optimiser; the artificial-intelligence optimisation inside the DareData collaboration package belongs to the partner. Active learning has no public source. Inverse design has no public source. What is shipped is a three-step synthetic data workflow, a base case, then templating, then several thousand variants in parallel, and the release of datasets and repositories: the windtunnel-20k dataset with 19,812 OpenFOAM cases, meshes generated from Stanford Cars and InstantMesh, a 70/20/10 split and 300 iterations; SimuStruct on OpenReview; and research repositories including meshnets, molecules-binding, GAN_SR_wind_field, datasets-generation, task-runner and kutu.

Product and Commercial Model

The product surface was coherent and unusually well documented for a company of this size. There is a Python client installed with pip and organised into simulator and fluid-scenario namespaces, a command line interface, and a web console positioned as real-time oversight with dashboards. The resource model distinguishes a dedicated MachineGroup, an ElasticMachineGroup, and an MPICluster spanning machines. There is a bring-your-own-hardware path that runs a local task runner on the user's own machines, a bring-your-own-container path, and a bring-your-own-software path, the last of which includes running a DiffDock inference. There is a benchmarks toolkit with a benchmark page and specialised entries such as the Fugaku comparison, and a console release cadence that reached v0.18.0 on 25 September 2025.

The pricing model is documented in the official how-much-does-it-cost page and consists of three charges. First, credits are bought as needed and a one-time platform fee, expressed as a percentage, is applied on top-up; the official page is offline and the percentage itself has no public source. Second, credits are consumed pay-as-you-go, with a task orchestration fee applied per simulation run, and the cloud provider's resource usage for compute, storage and data transfer subtracted from the balance. Third, the orchestration fee itself is quantified and was effective from 1 December 2025, applying only to runs started on or after that date: USD 0.02 per run for Individual, USD 0.01 for Enterprise and USD 0.005 for Academia. The documentation states plainly: "No subscriptions or hidden charges!"

The published price points are specific. A USD 5 top-up corresponds to approximately six hours on a c2d-highcpu-112 spot instance, roughly 700 core-hours, which is about USD 0.68 per hour. A core-hour on spot runs at about USD 0.01 and on standard machine types at about USD 0.04 to 0.07. Storage is about USD 0.02 per GB per month. Data transfer out is about USD 0.12 per GB, and data transfer in is free. No registration gift or free trial allowance appears anywhere in the documentation; the only verifiable free elements are free data import and the Academia orchestration fee.

Quotas are published for three tiers. Individual is capped at 768 vCPU, USD 10 of spend per hour, 12 instances and 1000 GB of storage. Enterprise and Academia are each capped at 10000 vCPU, USD 1000 of spend per hour, 100 instances and 5000 GB of storage. The official documentation is internally inconsistent on tier naming: the API reference names Standard, Power-user and Enterprise, while the pricing and quota pages name Individual, Enterprise and Academia, and no official explanation of the mapping was found. This report uses the Individual, Enterprise and Academia naming, which is the one carrying published numbers, and records the inconsistency as a conflict below. Enterprise is also the only tier that may run containers outside the KUTU image set and use MPI clusters.

Customers and use cases divide cleanly into two tiers, and the distinction matters for anyone reconstructing the business.

Officially presented are Tiago Lucas Frutuoso as an academic researcher, the Voltron Future CTV-26 RON offshore wind vessel project with Compute Maritime, Istanbul Technical University and AirShaper, in which several hundred hull CFD cases were executed in parallel; DareData as a partner, in a strategic collaboration on industrial data science announced from the partner's side, with claims of up to ten times faster development cycles and 3 to 5 percent energy savings, directed at heavy industry, renewables, metalworking, aviation, machinery, moulds and plastics; INESC TEC and the ATTRACT ecosystem; and the OpenFOAM community through the HPC Challenge. Also public are the windtunnel-20k dataset of 19,812 OpenFOAM cases and the SimuStruct structural dataset.

The documented cases share a profile: an academic researcher, a joint project and ecosystem partners. No reproducible commercial paying customer case with a contract value appears anywhere in the public record, and no media leak of a large enterprise paying customer list was found. Revenue, ARR, customer count and growth rate have no public source either; on the database side PitchBook carries only the financing and headcount records, and neither Preqin nor Dealroom has a verifiable revenue entry for this company.

Two attribution traps sit in this material and both are recorded below. The DareData release lists Heineken Group, NOS Telecommunications and Coca-Cola among the clients DareData has served, which makes them DareData's clients and not this company's. The ten times faster and three to five percent saving figures are the partner's own claims with no independent verification.

Competitive Landscape

Inductiva belonged to the cloud simulation platform quadrant, and the classification is instructive in itself: the only two competitors listed in its PitchBook profile are Rescale and SimScale, both cloud platforms, with no AI-surrogate company among them.

SimScale is a Munich company founded in November 2012, a spin-out from the Technical University of Munich, running a cloud platform for engineering simulation with a commercial software catalogue. Its disclosed equity funding includes a Series C of EUR 27M in January 2020 led by Insight Partners and an extension of EUR 25M in October 2021 with Draper Esprit and Insight co-leading, which the official release states brings the Series C total to EUR 52M, with the amounts of earlier rounds undisclosed. In December 2024 it signed EUR 25M of debt financing under the European Investment Bank InvestEU framework. As of October 2026 it was still operating. Rescale is a San Francisco company, business start date 18 November 2011, a Y Combinator W12 alumnus, running a cloud HPC and digital engineering platform with pay-per-use commercial CAE licensing through Ansys Elastic Licensing and a partnership delivering Siemens Simcenter, and it reported 1,350-plus applications and 190-plus architectures. On 7 April 2025 it closed a USD 115M Series D, with NVIDIA among the participants, taking cumulative funding above USD 260M, and it had not been acquired as of the report date.

The structural difference from Inductiva is not a matter of quality. SimScale and Rescale sell commercial software licences delivered from cloud infrastructure, which means licence revenue sits alongside the compute bill. Inductiva sold compute and orchestration for upstream open-source solvers, so there was no licence revenue line at all, and its differentiators were a single programming interface over many upstream packages, a templating workflow for synthetic data, and the ability to bring your own container, software or hardware.

The other half of the competitive set is the GPU-native and AI-surrogate route, which this report covers in separate company reports and which Inductiva did not belong to. PhysicsX describes itself as deploying artificial intelligence to transform how physical systems are engineered, combining artificial-intelligence-driven multiphysics inference with numerical simulation, and reported a USD 300M raise at a USD 2.4B valuation in 2026. Neural Concept is an EPFL spin-out in Lausanne led by Thomas von Tschammer, which released a physics- and geometry-aware design copilot in January 2026 and argues that artificial intelligence accelerates early-stage design rather than replacing simulation outright. Emmi AI was acquired by Mistral AI in May 2026, Monolith AI was acquired by CoreWeave in October 2025, and NAVASTO was acquired by Autodesk in October 2024. BeyondMath takes the neural-operator route. The distinction that matters is the layer: PhysicsX and Neural Concept sell models and licences and train surrogates themselves, whereas Inductiva generates training data and runs inference for models that users bring. In the artificial-intelligence-for-science value chain, Inductiva sits in the data production layer rather than the model layer.

Funding

The financing record is short, small and almost entirely reconcilable.

• 24 July 2023: EUR 243,403 received through OX2 AB as a European Union grant, under the project "Atmospheric Physics and Turbulence for Wind Energy". Whether this is a pass-through of grant funding, an advance or an equity investment has no public source, and its accounting treatment and clawback conditions are unknown. It is therefore reported separately and not added to the equity total.
• December 2023: EUR 2.25M seed round, led by Iberis Capital, a Portuguese venture firm, with Shilling, also Portuguese, and six named angels: Maria Joao Souto, Joao Penedones, Terry Huang, Ryan MacDonald, Ralf Herbrich and Joao Barros. The official sources give the amount, the lead, the participants and the use of funds, and they state that the company was founded in January 2021 and based in Porto.
• July 2024: an accelerator or incubator transaction recorded by PitchBook, with amount, counterparty and whether it involved equity all undisclosed; its association with the Unicorn Factory Lisboa scale-up list rests on the timing and the ranking of the name.
• 2024 to 2026: no further round disclosed, and no Series A or above appears anywhere.

Cumulative funding has two published readings. The official figure is EUR 2.25M, and the database figure is USD 2.45M, which differ in currency and in timing. This report uses EUR 2.25M, because the currency is explicit, the lead and the participants are named, and the primary releases state it directly. A combined total including the OX2 grant has no public source and is not used.

What is missing is the part that matters most for pricing any outcome. Post-money valuation has no public source. The cap table has no public source, and the registry only shows the number of shareholders moving from two at incorporation to three in April and May 2023 and five from August 2023, with capital of EUR 6,375 and then EUR 60,000, without identifying the holders or their proportions. Revenue, ARR, customer count and retention have no public source. The company was not listed: no IPO filing, no special purpose acquisition vehicle and no secondary market record.

The company's own framing of the moment it raised was that it combined large-scale physical simulation with artificial intelligence techniques, and the stated use of funds was to support engineers and scientists in forecasting. The July 2026 trade report connected the two facts, the EUR 2.25M seed and the shutdown, through the economics of a GPU cluster whose cost is largely fixed while the revenue that has to cover it was never documented. The causal chain in that report is a media interpretation and is recorded as such; the underlying facts it rests on, the seed amount, the reported workforce reduction and the official closure date, are each documented separately.

DISCREPANCIES

Ten source conflicts were identified and adjudicated item by item.

1. Registered location, Zurich against Porto. The first reading is the premise in circulation, a Swiss company in Zurich, together with a vague association with Swiss artificial-intelligence and GPU physics simulation. The second is Porto, Portugal, at Rua das Andresas, number 88, 3.5, 4100 Porto. The first has no source; the second is supported by an official InvestPorto and say-u release stating "Founded in January 2021 and based in Porto", by the Portuguese commercial register under NIPC 516243799, and by the official GitHub organisation location field. Adjudicated: Porto. Three independent sources agree and the alternative has no source. The associated ruling is that the Swiss companies in this scene are Akselos, founded by Thomas Leurent in Lausanne, and Neural Concept, an EPFL spin-out also in Lausanne, and neither should be merged with this record. The Swiss register itself could not be queried, and that limitation is disclosed rather than papered over.

2. Product positioning, GPU-native simulation cloud against hybrid HPC platform. The first is the premise in circulation. The second is the official wording, a hybrid HPC platform with a Python API that orchestrates open-source solvers. The second is supported by the home page copy, the 2025 conference paper, the documentation and the client source code. Adjudicated: the official positioning, with the qualification that the company did not develop a solver, that GPU capability comes entirely from upstream GPU ports and cloud machine types, and that no hand-written kernel, matrix-free stack or proprietary GPU backend is evidenced anywhere.

3. Founders. The first reading names Thomas H. and Kyriakos Antoniadis. The second names Luis (Morais) Sarmento, Hugo A. Penedones and Clara Goncalves. The second is supported by the official InvestPorto and say-u releases, the director records in the Portuguese register, and the signed conference material. Adjudicated: Sarmento, Penedones and Goncalves. No public source links this company to either of the other names, and the Thomas figures in the neighbouring Swiss scene belong to Thomas Leurent at Akselos and Thomas von Tschammer at Neural Concept.

4. Acquisition. The first reading is that a large technology company acquired it. The second is that no acquisition evidence exists, that operations ceased permanently on 31 January 2026, and that a workforce reduction was reported in July 2026. The only support for the first is the acquisition of Inductiv without the final letter a by Apple in May 2020. Adjudicated: the second. The misreading source is named explicitly, and the distinguishing features are the final letter, the business, the acquirer and the date.

5. Headcount. The first reading is 13, 15, 16 or 17 people across RocketReach, BounceWatch, StartupSeeker and PitchBook, plus the LinkedIn band of 11 to 50. The second is a reduction of roughly 80 percent reported in 2026 alongside a database status of Closed. The first is composed of database snapshots from 2025 or earlier and a platform self-reported band; the second is a weak media formulation and a database status field. Adjudicated: both are stated in parallel with the note that the timings are not comparable. No averaging is performed, the LinkedIn band is not treated as a current headcount, and no authoritative figure for the team size during 2026 exists. If the database range and the reduction report are both right, the actual team during 2026 was in the order of three people, which is in tension with the LinkedIn band and with the Closed status field.

6. Cumulative funding. The first reading is EUR 2.25M from the official releases and the trade press. The second is USD 2.45M from the database profile, which also lists 17 people as of 2025 and only two competitors. The third reading is a combined total that would include the OX2 grant of EUR 243,403. The first is an official release with a named lead and named participants; the second is a database profile; the third has no source. Adjudicated: EUR 2.25M as the equity total. The USD 2.45M figure is recorded as a currency and timing difference. The OX2 grant is listed separately and never added to the equity total, because its nature is unknown.

7. Acceleration claims, 15.65 times against 10 to 100 times. Both are official vendor statements, but the first carries a named case, named hardware and named timings while the second carries no case. Adjudicated: the 15.65 times figure is used only with its full conditions attached, namely CaNS, a stable atmospheric boundary layer, a g2-standard-4 with one L4 at 25h03m against two H100 at 1h36m, across different GPU models, as a vendor self-report with no grid size and no CPU equivalence comparison. The 10 to 100 times figure is characterised as a generalised vendor description of upstream solvers and is not used as a performance conclusion.

8. API tier naming. The API reference names Standard, Power-user and Enterprise. The pricing and quota pages name Individual, Enterprise and Academia. Both are official. Adjudicated: Individual, Enterprise and Academia is used as the primary naming, because it is the set with published quota numbers, and the existence of a second naming system inside the official documentation is recorded as an open documentation-consistency question. No official explanation of the mapping was found.

9. OpenFOAM HPC Challenge participation. The official company post states participation, describing hardware configurations, hyperthreading, domain decomposition and a result of 768 partitions on an elastic MPI cluster, with better value than more expensive hardware in some cases. The academic side, the challenge paper and the public data repository, lists twelve submitting institutions and Inductiva is not among them; the listed institutions are Wikki GmbH, Huawei, CINECA, University College Dublin, the Bundeswehr University Munich, BAW, the University of Minho, the Technical University of Munich, UKAEA, Engys and E4 CFD FEA Service, among others. Adjudicated: the two statements are presented side by side and the finding is that the participation is a company self-report whose submission data could not be located in the community repository. No claim is made that the company won or ranked in the challenge, and none is supported.

10. Corporate existence, a database Closed status against the official closure date. The first is a database status field. The second is the official contact page statement that operations cease permanently from 31 January 2026. The second is cited by a database that in turn cites the official page. Adjudicated: the official date is used, and the fact that the original announcement could not be retrieved is disclosed. The Wayback Machine and archive.org were unreachable throughout the collection period, with CONNECT tunnel errors and failed fetches, and archive.today returned no result for the site or the pricing page. The two statements point in the same direction and differ in evidence level rather than in substance.

Risks and Diligence Gaps

For a company that has closed, the diligence questions change shape. Most of what a live-company diligence would ask can no longer be answered, and the honest position is to record them as questions that remained unanswered at the point of closure, or as unknowns that still matter to anyone reconstructing this route. The following fourteen items are drawn from the full list of twenty recorded gaps and are rephrased for that purpose.

The closure announcement was never retrieved in its original form. The date of 31 January 2026 comes from a database quoting the official contact page. The Wayback Machine and archive.org were unreachable throughout collection, and archive.today had no snapshot. This is the single most consequential methodological limitation in the report, and it sits on the central fact.

The 404s are not documented as an orderly shutdown. The website, the documentation site and the console all return 404, but no DNS or hosting archive and no official statement was obtained distinguishing an orderly wind-down from a domain or host failure. On the evidence available the official statement is the basis for reading it as an orderly closure, and that is a reading rather than a verified hosting fact.

The Swiss negative is a consistency argument rather than a registry result. The Swiss federal commercial register API returned 401 and its interface did not render results. The conclusion that no Swiss entity exists rests on three-way agreement between the Portuguese register, the official releases and the GitHub organisation location, and not on a direct search of the Swiss registry.

Headcount and the workforce reduction cannot be reconciled. The database range of 13 to 17, the LinkedIn band of 11 to 50, the July 2026 report of roughly 80 percent and the database Closed status cannot all hold at the same time, and no authoritative figure exists for the team during 2026. The original report phrases the figure as "Word on the street", which is explicitly not a sourced number.

The one GPU benchmark cannot be reproduced. The 15.65 times figure for the CaNS stable atmospheric boundary layer case carries no grid size, no grid-independence verification and no equivalence comparison against a CPU baseline, and it compares different GPU models rather than scaling one model. Nothing about it can be independently checked now that the platform is offline.

The OpenFOAM HPC Challenge submission is not in the community repository. The official post gives a 768-partition result, while the challenge paper and the public data repository list twelve institutions without this company among them. The submission data was never located. For a retrospective, the question worth asking of any comparable claim is simple: where is the data.

There is no accuracy statement of any kind. No cross-solver or cross-version equivalence or validation material exists, so no conclusion about result fidelity can be drawn, and none is drawn.

The artificial-intelligence narrative has no verifiable product behind it. Geometry surrogate, Bayesian optimisation, active learning and inverse design all lack public sources. The shipped capability is synthetic data generation and user-supplied inference. For anyone citing this company as evidence of a platform capability, that distinction has to be carried forward.

The platform fee percentage is unknown. The platform fee percentage was the only self-operated revenue line in the pricing model, and it was never obtained, which leaves gross margin and any price comparison with Rescale or SimScale uncomputable.

Valuation, cap table and shareholder identities are missing. Whether a Series A followed the EUR 2.25M seed, at what post-money valuation, and with what ownership split, have no public source. For a liquidated entity this determines what the seed investors actually hold, and it remains unknown.

The 2024 accelerator transaction is undocumented. Amount, counterparty and whether it involved equity are all undisclosed, and the link to the Unicorn Factory Lisboa scale-up list rests on a timing and a ranking correspondence.

The character of the OX2 AB grant of EUR 243,403 is unclear. Whether it is grant pass-through, an advance or equity is unknown, as are the accounting treatment and the clawback conditions. In a liquidation these items sit directly on the balance sheet.

No commercial metric was ever disclosed. Revenue, ARR, customer count and retention have no public source, so there is no way to establish whether the business ever covered its own costs, and the pricing model alone cannot settle it.

The customer base was research-level and collaboration-type. The public cases are an academic researcher, a joint project and ecosystem partners, with no reproducible commercial paying customer case carrying a contract value. This is relevant to the general question of whether orchestration-layer revenue can be large enough to carry a cloud cost base.

Attribution risk on the partner's customer list. Heineken, NOS and Coca-Cola are clients of DareData, and the ten times and three to five percent figures are the partner's own claims. Any later reuse of this material must preserve that attribution.

Two further items are worth carrying forward as method notes rather than as findings. The claim that a founder's later venture was founded has no verifiable official page or registration record and is therefore not repeated as fact. And the architecture documentation cannot be validated after the fact: the simulator list of 39 modules and 45 container images cannot be tested for runnability once the platform is offline, and there is no indication of which capabilities were already being deprecated before the closure.

Overall Assessment

• Technology 6.5. Assessed on the state of the platform before operations ceased. The orchestration layer was real and substantial: a task queue with automatic resubmission on preemption, spot and on-demand scheduling, elastic machine groups and MPI clusters, a Python API and console, a templating workflow that produced thousands of parallel variants, and support for more than thirty upstream solvers including GPU-capable builds of nine of them. The engineering transparency was above average for a seed-stage company, with a documented pricing model, published quotas, published per-hour spend caps and a reproducible benchmark. Deducted for the absence of any self-developed solver or GPU kernel, for the single GPU benchmark being a cross-model vendor comparison with no grid size and no CPU equivalence check, for the complete absence of accuracy or equivalence material, for the OpenFOAM HPC Challenge submission data not being locatable in the community repository, and for the headline 10 to 100 times figure having no case attached to it.
• Commercialisation 3.0. Reflecting an operating business that has already terminated. The product and pricing design was genuinely usable and transparent, with prepaid credits, pass-through cloud costs and a per-run orchestration fee that differed by tier, and with three published quota tiers and documented spot and on-demand rates. Deducted for the operating termination itself, for no disclosed revenue, ARR, customer count or retention, for a public case list that contains no commercial paying customer with a contract value, for an undisclosed platform fee percentage that leaves gross margin uncomputable, for no published free allowance, and for two official documentation inconsistencies in tier naming.
• Funding 3.5. One seed round of EUR 2.25M in December 2023, led by a Portuguese venture firm with a second Portuguese firm and six named angels, which is documented at the level of lead, participants, amount and date. Deducted for the small absolute size relative to a business whose principal cost is cloud compute, for no subsequent round of any kind, for no valuation, cap table or shareholder identities, for the unclear character of the EUR 243,403 OX2 channel grant, and for an accelerator transaction whose amount and counterparty were never disclosed.
• Product maturity 5.5. Assessed on the state of the platform before operations ceased. The product surface was coherent across client, command line, console, three resource models, three quota tiers, three bring-your-own paths, a synthetic data workflow and a benchmarking toolkit, with a release cadence that reached v0.18.0 in September 2025. Deducted because the product cannot be operated or validated now that the platform is offline, because the console and documentation no longer resolve, because the simulator inventory was never testable after the fact, because the documentation is internally inconsistent on tier naming, and because the platform's differentiation sat in the orchestration layer while the cost base sat in the cloud bill.
• Market position 4.0. A clear position in the open-source solver orchestration quadrant, confirmed by the only two competitors listed in its database profile being the two cloud platforms Rescale and SimScale, plus a documented dataset and research-repository footprint that gave the platform an academic visibility that a peer of its size would struggle to match, and a partner-led industrial collaboration. Deducted for the operating termination, for the absence of proprietary solvers, commercial licence revenue or an accuracy certification system in a market where Rescale and SimScale monetise licence revenue directly, for a customer base confined to research and collaboration relationships, and for a capital base roughly two orders of magnitude below the platform peers in the same cohort.

References

1. Official website copy for Inductiva, inductiva.ai, including the hybrid HPC platform statement and the Simulation-Optimized, Programmatic Control, Hybrid Ready and Smart Scaling selling points, retrieved from cached copy in October 2026 after the site began returning 404.
2. Official documentation repository, github.com/inductiva/docs, covering How much does it cost, the USD 5 and 700 core-hours computation note, quotas, the CaNS GPU benchmark, the Quantum ESPRESSO and Fugaku benchmarks, the AMR-Wind benchmark, OpenFOAM versions and containers, computational infrastructure, cloud machine selection, the Generate Dataset workflow, bring-your-own-software inference, the v0.18.0 release notes, and the API reference pages on tiers, simulator overview and FAQ.
3. Official GitHub organisation, github.com/inductiva, with the client repository inductiva/inductiva, the container image repository inductiva/kutu with 45 upstream solver images including GPU variants, and the research repositories meshnets, molecules-binding, GAN_SR_wind_field, datasets-generation and task-runner.
4. Official LinkedIn posts from the company, September 2025 for the OpenFOAM HPC Challenge results and the 768-partition result, and November 2025 for the Tocha wind farm case with more than 33,000 OpenFAST simulations; also the company page giving the 11 to 50 band, the Porto headquarters and the 2021 founding year.
5. Official conference paper, "Inductiva API: Simulation and optimization at scale made simple", hosted at files.messe.de, 2025, for the integrated computational platform wording, the Python-first statement, the open-source claim and the AI-based methods value proposition.
6. InvestPorto official news release, December 2023, investporto.pt, and the say-u official release of the same month, for the EUR 2.25M round, the lead and the participants, the three founders, the FEUP spin-off label and the "based in Porto" wording.
7. North Data compilation of the Portuguese commercial register for Inductiva Research Labs SA, PT 516243799, for the incorporation date, legal form changes, capital history, shareholder counts, directors, auditor, filed annual accounts and the EUR 243,403 OX2 AB European Union grant.
8. PitchBook company profile, profile 489865-96, for the 2021 founding, the Porto location, 17 employees as of 2025, the USD 2.45M figure, the two listed competitors and the July 2024 accelerator transaction.
9. aVenture company page recording the status Closed and quoting the official contact page on the permanent cessation of operations from 31 January 2026.
10. devs.com.pt, "The Crunch at Inductiva: when high-end science meets venture capital reality", 7 July 2026, for the workforce reduction of roughly 80 percent, phrased as "Word on the street", the seed amount and the cloud compute cost argument.
11. DareData official news release on the industrial data science partnership, 2024, for the development speed and energy saving claims and for the client list attribution, and INESC TEC official news on the ATTRACT activity, 2024.
12. arXiv:2603.27565 on the first OpenFOAM HPC Challenge and the OpenFOAM HPC technical committee data repository, for the 237 data points, the twelve submitting institutions, the occDrivAer case definition, the aggregate energy consumption and the strong-scaling limit.
13. Media reporting of 28 May 2020 on the acquisition of Inductiv, without the final letter a, by Apple, for the origin of the acquisition misreading, together with the Inductiva official LinkedIn page and the CFD-Online webinar announcement of 11 July 2025.
14. Research method note: the Wayback Machine and archive.org were unreachable throughout collection, with CONNECT tunnel errors and failed fetches, archive.today returned no result for the site or the pricing page, the Swiss federal commercial register API returned 401 and its interface did not render, and raw GitHub endpoints timed out, which is why the GitHub REST contents and tree interfaces were used instead. All figures in this report carry a source and a date, and vendor marketing figures are labelled as vendor self-report or partner self-report.

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