BeyondMath Deep Dive: A Foundation Physics AI on the Neural-Operator Route, a $19M Consortium Budget, and Three Misread Facts

Executive Summary

BeyondMath is a four-year-old London deeptech company that trains foundation models of physics on first principles and sells them into the part of the engineering workflow where computational simulation is the bottleneck. It was founded in 2022 in Cambridge by Alan Patterson and Darren Garvey and is now headquartered in London. On 25 February 2026 it closed a $10M seed extension led by Cambridge Innovation Capital, which took the full seed round to $18.5M and total capital raised to $20.35M. As of the report date it remains a private independent company; no acquisition approach, offer, or listing process appears on the public record.

Three things about this company are widely misreported, and each of them changes how an investor should read everything else in this report.

First, the $19M figure is not a Honeywell contract to BeyondMath. It is the total three-year budget of Project STRATA, a consortium research programme funded through the UK's ATI Programme and led by Honeywell from its Yeovil plant, which employs more than 500 engineers. BeyondMath is one of five members, alongside 3T Additive Manufacturing, Qdot Technology and the Oxford Thermofluids Institute. BeyondMath's own share of that GBP 14.1M is not disclosed anywhere. A number of aggregators rendered the programme as "a $19 million three-year commercial partnership" with Honeywell. That is a category error: it converts a government-funded consortium budget into a direct customer contract, and an investment committee that prices BeyondMath off that number will be wrong by an order of magnitude.

Second, the technology is a neural-operator surrogate, not a GPU-rewritten solver. The speedups quoted by the company are what becomes available once the solver iteration loop is removed and replaced by a learned model, not what becomes available when the same solver is made to run faster on better hardware. A "1000x" figure produced by a surrogate and the 10x-34x figures routinely quoted for GPU-accelerated traditional CFD are different physical quantities and are not comparable. This report keeps them in separate columns throughout.

Third, the founders are Google, eBay, HomeX and Evi operators, not Dyson, Formula One or aerospace engineers. Alan Patterson holds an Oxford DPhil in electronics and spent much of his career on search, knowledge graphs and applied machine learning: Nokia, Betfair, Evi Technologies, Barclays, Google, eBay, and HomeX, where he was VP Head of Applied Science. Darren Garvey was Director Principal Research Engineer at HomeX, which was subsequently acquired by Rolls-Royce. The two are working together for the fourth time. No public source links either founder to Dyson, to an F1 team, or to a space programme. The only near-miss in the record is a French business outlet describing team members as coming from DeepMind, Oxford, Cambridge or Imperial College, which refers to staff rather than founders and names nobody. Formula One, in BeyondMath's case, is a customer relationship and not a founder background.

With those three corrections in place, the evidence looks considerably better than the marketing suggests, and considerably narrower than the narrative.

The single strongest external data point is published by a third party. In a joint report by the Aerospace Technology Institute and Capgemini, reported by diginomica in August 2026, BeyondMath's automotive aerodynamics case completes a full-field simulation in 1.5 minutes against 24 hours on a 720-core HPC machine, roughly 960x, while holding accuracy within 2% of a high-fidelity detached eddy simulation. The case is automotive aerodynamics, not aerospace airworthiness certification, and it should be cited as such.

The second-strongest point is the company's own DrivAerML benchmark post of 3 August 2026, which is unusually complete by industry standards: 500 DrivAer body variants, each at roughly 160 million cells in scale-resolving hybrid RANS-LES, with roughly $1M of CFD compute behind the ground truth. Physics Factory models show relative L2 errors of 3.53% on surface pressure, 3.83% on volume pressure and 4.98% on velocity, against 3.71/5.72/4.14 for Transolver-3, 3.82/6.08/5.93 for AB-UPT (Emmi AI) and 4.81/7.74/6.78 for Transolver. Drag Cd reaches R-squared 0.985 and lift Cl 0.975. The post also openly states that Transolver-3 leads on volume velocity, and that the full table including the rows they do not win is published below. The model was assembled autonomously by AI agent crews with no AI researchers or consultants involved and under 30 minutes of human hands-on time.

The third-strongest point is the most disciplined piece of engineering disclosure anywhere in the company's material: the October 2026 internal-flow post covering four engine intake manifold designs, 20 cylinders and 751 predictions at roughly 15 seconds each on a laptop CPU, against 264 CFD cases of which 259 were in the training set and 5 were held out, giving flow-coefficient mean absolute error 0.011, swirl number 0.008 and average pressure-drop error 1.6%, with no meshing and no solver iteration required.

Two numbers from August 2023 deserve to be treated as anchors, because they are the company's own and are therefore the hardest to dispute. Sifted reported that the model was roughly 0.8% less accurate than traditional supercomputing methods, and that Darren Garvey considered that acceptable for an F1 team while noting that for aircraft manufacturers such as Boeing a higher level of accuracy would be needed and that this might take about five more years. Read against the 2026 material, the trajectory is real: five years from that statement runs to around 2028, and no certification-related public information has appeared in the meantime.

Funding is real and bounded. A GBP 1.5M pre-seed in 2022 was led by UP.Partners with The Creator Fund and Ex-Fi; a $8.5M (GBP 6.5M) seed on 20 August 2024 was again led by UP.Partners with Insight Partners and InMotion Ventures, the investment arm of Jaguar Land Rover; and the February 2026 $10M extension was led by Cambridge Innovation Capital with the same three firms following on. The full seed round closed at $18.5M and CEO Alan Patterson confirmed a cumulative $20.35M raised. Valuation was explicitly refused: the CEO said the company is not disclosing its current valuation. The only figure in circulation, $32.9M, is self-labelled low confidence by the platform that publishes it and is inconsistent with the equity structures reported elsewhere, so it is not used as a fact in this report.

Commercialisation is early. The only named counterparties are Honeywell, and then only through the STRATA consortium, and one unnamed Formula One team whose chief CFD officer is quoted on the company site with the team and driver left blank. BeyondMath is running an Early Access Program, promotes Generative Physics Studio for automotive in early access, and has pushed Physics Factory as its 2026 flagship, with five agent crews named Organise, Analyse, Train, Optimise and Predict, and Cadgen2 as its AI CAD generation module. GPStudio appears in the Siemens Xcelerator ecosystem catalogue as a pretrained foundation model that does not need to be trained on customer data. The company stresses on-premise deployment inside the customer's firewall. There is no public pricing, no public revenue, no ARR disclosure and no unit-economics disclosure.

Our view: this is a technically serious, unusually honest-about-its-limits early-stage company whose narrative is roughly two years ahead of its commercial evidence. The technology is real, the third-party validation is the best in its cohort, and the founders have shipped machine-learning systems at consumer scale before. What is missing is everything that turns a seed-stage deeptech into a company: a price list, a named customer beyond a consortium seat and an anonymous F1 team, revenue, an independent reproduction of the headline benchmarks, and a credible path to aerospace certification. On the evidence available, this is a company worth a small position in the category and a large amount of diligence.

Company Overview

BeyondMath Ltd is registered in England and Wales as company number 14110252, with a registered address at 71-75 Shelton Street, London WC2H 9JQ, active, and a stated business nature of business and domestic software development. Both founders were appointed directors on 16 May 2022, with Alan Patterson also acting as company secretary. The company holds a UK sponsor licence in the Worker category, skilled worker, London, verified in February 2026, which is consistent with a company that does intend to hire and which shows no public sign of contraction.

The founding location is recorded differently by different sources, and the difference is worth preserving rather than smoothing over. Cambridge Innovation Capital, the lead investor in the 2026 extension, describes the company as "founded in 2022 in Cambridge by AI industry veterans Alan Patterson and Darren Garvey, and now headquartered in London." The company's own seed-round release is datelined London and states only that it was founded in 2022, without naming Cambridge. TechCrunch's 2024 coverage used the phrase Cambridge-based. A French business outlet wrote that the company thought in Cambridge and quickly established its headquarters in London. The most defensible reading is that the company was founded in Cambridge in 2022 and relocated its headquarters to London, with the two founder-level and database sources that describe the founding city saying so.

Alan Patterson, co-founder and CEO, earned a DPhil in electronics from the University of York between 1992 and 1996 after a first degree in electrical and electronic engineering at the University of the West of Scotland. His career runs Nokia Telecommunications, Nexan, Imerge, Oxford BioSignals, Betfair, Evi Technologies and True Knowledge, Barclays corporate and investment banking, spider.io, Google, eBay as Senior Director and Distinguished Engineer, and HomeX as VP Distinguished Engineer and then VP Head of Applied Science. At eBay his work was in the knowledge-graph area. The stated origin story is personal and specific: in the 1990s a single electromagnetic-field simulation took a week to run, and by 2022 that remained a bottleneck, which made the company believe the conditions for AI-assisted simulation had arrived.

Darren Garvey, co-founder and CTO, has more than fifteen years in AI and software development, and was Director, Principal Research Engineer at HomeX, a company later acquired by Rolls-Royce. The two men previously worked together at Evi Technologies, the company behind an AI assistant that Amazon acquired and that became foundational to Alexa, and this is their fourth project together.

A point of record: no source in the public record links either founder to Dyson, to a Formula One team, to a space agency or to Imperial College. The single adjacent claim is a French business description of the team as drawn from DeepMind, Oxford, Cambridge or Imperial College, with roughly 15 people organised into AI, HPC and computational physics groups, which is a statement about staff, not about founders, and which names nobody.

Team size has seven distinct published readings. Approximately 15 people in 2024 from the French business outlet; 18 from StartupIntros; 21 from PrivateMarketView in October 2026; 25 from TipRanks on 14 September 2026; a LinkedIn band of 11 to 50; a CEO statement of 18 people who "nearly doubled over the past year" across 10 nationalities in February 2026; and an official plan to double headcount during 2026. The safest usable statement is a range of roughly 20 to 30 people, with the official doubling described as a plan rather than an accomplished fact. The 10-nationalities figure is the one hard diversity data point available, and it comes from the CEO directly.

On governance, Adam Grosser, chairman of UP.Partners, served as a director from 19 April 2024 until his resignation on 4 February 2026, which places him at the table across the seed and the extension but no longer in the structure.

Technology

The technical characterisation that matters most is this: BeyondMath does not rewrite solvers on GPUs. It learns the operator underlying the partial differential equation and predicts the physical field directly. InMotion Ventures, an investor, describes the company's focus as neural operators that directly learn and approximate the underlying operators of partial differential equations. The company draws the distinction from classical surrogates in its own words: surrogate models memorise a geometry set and break when the shape changes, whereas the aim is foundation models trained on the laws of physics that are not locked to a single design. Garvey frames the objective as approximating the real world rather than approximating the simulations, which is why real-world data enters the training mix.

Training runs on NVIDIA DGX H200, and BeyondMath describes itself as one of the first adopters of that platform, with Carlo Ruiz, VP of AI data centre solutions for NVIDIA EMEA, named in the company's own release. The consequences for how the speedup numbers are read are direct. A solver that has been ported and optimised for GPU hardware typically delivers 10x to 34x; industry reporting gives the example of Honda running Ansys Fluent on NVIDIA Grace Blackwell for a 34x gain. BeyondMath's headline figure is "up to 1000x", and the two are produced by different mechanisms. The first is the same physics solved faster; the second is a different model of the physics that avoids solving it. They should be read as separate columns in any comparison.

Within the company's own numbers, the gradation is meaningful. The most defensible acceleration figure is the one published by a third party: 1.5 minutes against 24 hours on 720 cores, about 960x, in the ATI and Capgemini report. The ATI and Capgemini report also characterises the approach as grounding the tool in fundamental physics, having learned laminar-to-turbulent transition behaviour, Reynolds number effects, curvature effects and separated-flow structure, and the company as remaining directionally accurate on configurations it has not seen. That same case is automotive aerodynamics, and it is not an aerospace airworthiness validation.

The DrivAerML benchmark post of 3 August 2026 is the most complete technical artefact the company has published. The dataset is the standard automotive aerodynamics benchmark: 500 DrivAer body geometry variants, each resolved at roughly 160 million cells in a scale-resolving hybrid RANS-LES formulation, with roughly $1M of CFD compute behind the ground truth. The Physics Factory model reaches relative L2 errors of 3.53% on surface pressure, 3.83% on volume pressure and 4.98% on velocity, with drag Cd at R-squared 0.985 and lift Cl at 0.975. The comparison set is academic and industry models: Transolver-3 at 3.71, 5.72 and 4.14 with R-squared of 0.972 and 0.985; AB-UPT, which is Emmi AI's model, at 3.82, 6.08 and 5.93 with 0.963 and 0.975; and Transolver at 4.81, 7.74 and 6.78. Two features of that post deserve credit. It states explicitly that Transolver-3 remains ahead on volume velocity, and it publishes the full table including the rows the company does not win, which is close to unheard of in vendor benchmarking. It also describes the build: AI agent crews carried out data analysis, architecture assembly, training and evaluation, with zero AI researchers or consultants involved and under 30 minutes of human hands-on time, the architecture being assembled from the company's own component library.

The October 2026 internal-flow post is the most disciplined precision disclosure in the company's material. It covers four engine intake manifold designs (BMR4, BMR6, BMT4, BMT6), 20 cylinders and 751 predictions, each taking roughly 15 seconds on a laptop CPU. Against CFD, across 264 cases of which 259 were in the training set and 5 held out, the flow-coefficient mean absolute error is 0.011, the swirl number 0.008 and the average pressure-drop error 1.6%. On the 5 held-out cases alone the flow-coefficient MAE is 0.022, the swirl number MAE 0.021 and the pressure-drop MAE 54 Pa. No meshing is required, no solver iteration is required, the model can be trained on the customer's own CFD data, and it runs inside the customer's firewall.

On full-vehicle transient simulation, the company reports under 100 seconds for a complete car where classical methods take hours to days, and shows a full-scale 3D transient simulation of the 2026-specification Formula One car, credited to a team and driver left blank. On drag prediction from raw STL geometry across 50 unseen shapes, the reported drag-coefficient MAE is roughly 0.0032 with second-scale inference, and the company states that the model still ranks design variants consistently and captures the direction of drag change across three orders of magnitude of error.

Two caveats belong beside every figure above. The 500 million cell 3D transient claim in the STRATA release of 28 August 2025 and the 200 million cell figure on the current home page are inconsistent, and are best read as a 2025 peak capability claim alongside the 2026 currently promoted specification, with both quoted rather than one discarded. And the "1000x" figure appears essentially everywhere as an upper bound with no stated benchmark configuration: no solver, no core count, no indication of whether pre- and post-processing are included.

The earliest and most useful calibration came from the company itself in August 2023. Sifted reported that results were approximately 0.8% less accurate than traditional supercomputing methods, and quoted Garvey saying this was acceptable for an F1 team but that aircraft manufacturers such as Boeing would need higher accuracy and that it might take about five more years. In 2023 the first use case was a fine-tuned aerodynamic model for a Formula One team, and Patterson described "Einstein AI" as a reasonable long-term ambition while framing physical equations as the ground truth.

Product and Commercial Model

The product surface is broad and, in places, still moving. Generative Physics is the umbrella brand. Digital Wind Tunnel was the first product, aimed at aerodynamics, and was the vehicle for the 2024 launch coverage. GPStudio is listed in the Siemens Xcelerator ecosystem catalogue as an official product name and is described there as a pretrained foundation model that runs CAE workflows without requiring training on customer data. Generative Physics Studio, the automotive version, is in early access. The Early Access Program is the main call to action on the company site.

Physics Factory is the platform being pushed hardest in 2026. It organises five agent crews, named Organise, Analyse, Train, Optimise and Predict, and turns a customer's own CFD and FEA data into a deployable physics AI model. The stated value proposition is a change of operating model rather than a modest speedup: platform plus consultants takes 8 to 10 weeks, platform plus agents takes hours. Cadgen2 is the AI CAD generation module inside Physics Factory, with a published tablet-variant study in which generated geometry was constraint-checked, exported to BOM and STEP, and completed in under an hour, and an F1 aerodynamic design-of-experiments post covering a 28-handle study with geometry validity and fold metrics and per-run latencies of 465 to 547 milliseconds.

Deployment posture is on-premise, and this is a deliberate differentiator rather than an incidental detail. Simulation data, trained models and experiment logs remain on the customer's own infrastructure, inside the customer's firewall. For aerospace customers in particular, data residency and intellectual property containment are procurement requirements, so this is a credible differentiator against cloud-first competitors.

On customers, precision matters more than volume. Named counterparties are Honeywell, and then only in its capacity as a member of the STRATA consortium alongside 3T Additive Manufacturing, Qdot Technology and the Oxford Thermofluids Institute, which are consortium partners rather than customers. Beyond that there is one unnamed Formula One team, quoted on the company website as "Head of CFD at an F1 team" calling the technology the most transformative the team has seen, with the team and driver deliberately left blank, and there are unnamed categories covering major automotive, aerospace and electronics manufacturers, data-centre design, and semiconductor manufacturing.

The customer-count question is worth a note, because the difference between the two published formulations is one of meaning rather than style. Sifted in 2023 wrote of a fine-tuned system for a Formula One racing team, in the singular. TechCrunch in 2024 wrote that several F1 teams are already testing the software. "Several teams are testing" is the more conservative formulation and is the one this report uses, since a testing relationship is not a customer relationship.

Ecosystem position is stronger than the customer list. BeyondMath was among the first adopters of NVIDIA DGX H200, was mentioned in an NVIDIA GTC keynote, and joined the NVIDIA Inception Program in September 2025. It has an AWS partnership, appears in the Siemens Xcelerator ecosystem catalogue, and is a case study in the joint ATI and Capgemini report. In 2024 it received an ATI HUB Breakthrough Award Highly Commended distinction in the inaugural ATI awards.

On revenue, the honest answer is that almost nothing is public. There is no pricing of any kind, no seat price, no per-computation price, no project quotation, no revenue and no ARR disclosure. The only figure anywhere is a Seedtable estimate of $5M to $12M in estimated revenue, produced by a public heuristic model rather than by the company, and it should be labelled as a third-party estimate whenever it appears. Caplight characterises the business model as subscription SaaS plus usage-based SaaS plus services and consulting, which is broadly consistent with what the public material describes but equally is a database's classification, not a company disclosure.

Funding

The financing history is short, coherent and, unusually for this cohort, largely reconcilable.

• 2022: GBP 1.5M pre-seed led by UP.Partners, with The Creator Fund, Ex-Fi and private investors. The reported runway was to Q4 of 2024. Exact date and the complete investor list are not on the public record.
• 20 August 2024: $8.5M (GBP 6.5M) seed led by UP.Partners, with Insight Partners and InMotion Ventures, the investment arm of Jaguar Land Rover.
• 25 February 2026: $10M (GBP 7.4M) seed extension led by Cambridge Innovation Capital, with UP.Partners, Insight Partners and InMotion Ventures following on. The full seed round closed at $18.5M.

Cumulative capital raised is $20.35M, confirmed by name by CEO Alan Patterson. There is no disclosed valuation: the CEO said that as a private company at this stage of growth the company is not disclosing its current valuation.

The lead investor's strategy statement is unusually specific. UP.Partners chairman Adam Grosser, who also served as a BeyondMath director from April 2024 to February 2026, said BeyondMath represents exactly the kind of paradigm shift the firm seeks to support. Cambridge Innovation Capital, which states that it manages over GBP 600M with a 12-year track record across AI, semiconductors, life sciences and quantum computing, led the extension with principal Edward Inns. Insight Partners was represented by managing director Ganesh Bell and InMotion Ventures by managing director Mike Smeed.

One relationship deserves flagging for diligence rather than conclusion. InMotion Ventures states that BeyondMath's technology has the potential to significantly accelerate Jaguar Land Rover's design and engineering processes. JLR is therefore both an investor through InMotion and a plausible industrial customer, and it has never been officially listed as a customer. Separately, one database lists Garvey Asset Management among investors; the name matches the CTO's surname and may indicate a founder-affiliated vehicle, which is a reasonable question to ask directly.

Use of funds as stated in February 2026: expand commercial deployment of generative physics technology, increase research capability, double headcount during 2026, and extend into European, US and Japanese customers.

Competitive Landscape

BeyondMath sits in the AI-native simulation cohort, and the cohort has consolidated around it faster than around most of its peers.

PhysicsX, in London and therefore a direct neighbour, has raised a $300M Series C as of June 2026 after a $135M Series B in 2025, with Siemens and Applied Materials among the investors, and positions itself as an AI-native physics engine across aerospace, Formula One and semiconductors. Neural Concept, in Lausanne, closed a $100M Series C led by Goldman Sachs Alternatives Growth Equity and reports more than 50 enterprise customers including General Motors, GE Vernova, Leonardo, Eaton, Safran, Renault and several Formula One teams. Luminary Cloud, in San Mateo, raised $72M in a Series B in September 2025 and builds a GPU-native CFD solver with pretrained SHIFT models. Vinci4D, in Palo Alto, raised $250M at a $1.5B valuation on 6 October 2026, and specialises in chip thermal-mechanical-fluid problems. SimScale, in Munich, has disclosed more than EUR 27M and pairs cloud CAE with AI surrogates and an engineering agent.

Three consolidation events in the cohort are directly relevant to BeyondMath's position. Emmi AI was acquired by Mistral AI in May 2026, bringing a 30-plus-person team into Mistral and making Linz a Mistral office. Monolith AI was acquired by CoreWeave in October 2025. NAVASTO was acquired by Autodesk in October 2024 and integrated with Alias. BeyondMath is one of the few UK targets in this category that has not been acquired, which cuts both ways: it preserves optionality for the founders and their investors, and it also means the strategic bid is live.

The traditional incumbents remain the other half of the competitive set. Ansys SimAI, announced January 2024, claims 10x to 100x speedups and Ansys was itself acquired by Synopsys in July 2025. Siemens Simcenter PhysicsAI, following the completed Altair acquisition in March 2025, claims roughly 1000x. Dassault Systemes offers SIMULIA Virtual Twin across 370,000 customers. Cadence offers Clarity and Celsius plus MSC Nastran. NVIDIA itself fields PhysicsNeMo and Warp and announced AI Enterprise pricing at roughly $4,500 per GPU per year.

The one place where BeyondMath's positioning is genuinely distinctive rather than merely aspirational is the benchmark it chose to publish against academic models rather than against vendor marketing claims, and the place where it is most exposed is that a benchmark leaderboard is not a substitute for a customer contract.

Common Misreading Warning: Unrelated to Quantum or Photonic Computing

BeyondMath's entire body of first-party material, comprising the company website, its news releases, its developer blog, its STRATA case page, its Siemens ecosystem entry, its GTC references and its investor statements, describes the technical route as neural operators and geometry-deep-learning surrogates trained on NVIDIA GPUs. The words quantum, qubit, photonic and optical computing do not appear.

The most likely source of the confusion is a specific sentence in the official seed-extension news release. In the "About CIC" paragraph describing Cambridge Innovation Capital, the release states that the firm's investment scope includes quantum computing. That sentence describes the investor's scope, not the portfolio company's technology, and secondary material that splices the paragraph onto BeyondMath creates the impression of a quantum-CFD company that does not exist.

The companies that genuinely pursue quantum or photonic computation for CFD are a different set. BQP, also known as BosonQ Psi, runs the BQPhy platform and a hybrid quantum-classical solver called HQCFM, and claims that 30 logical qubits are sufficient for large-scale jet-engine CFD where classical methods would need 19.2 million compute cores; it has worked with Classiq and NVIDIA CUDA-Q and was recognised with a Quantum World Congress National Security Startup Award in 2023. QubitSolve, based in Morgantown, West Virginia, was awarded an NSF SBIR Phase II grant of $1,197,002 in December 2025 to solve Navier-Stokes with variational quantum algorithms, with an MVP expected by the end of 2027, in partnership with Infleqtion. Classiq, in Israel, is quantum compilation software and has demonstrated alongside BQP and NVIDIA. These are separate organisations on a separate technical timeline, and BeyondMath has no technical relationship with any of them.

The relationship that does exist is one of simultaneous competition and complementarity. Both approaches attack the same wall, which is the compute limit of classical solution of partial differential equations, but they arrive from opposite directions. BeyondMath's route is available today, whereas the quantum route depends on utility-scale fault-tolerant machines, a dependency BQP itself states plainly. A reader who places BeyondMath in the quantum category will misread its engineering maturity and its competitive set; the correct placement is AI-native surrogate modelling, competing for the same engineering budget and the same design-cycle bottleneck.

DISCREPANCIES

Seven source conflicts were identified and adjudicated item by item.

1. Total capital raised. The official release says the seed round closed at $18.5M. CEO Alan Patterson, named in Tech Funding News in February 2026, confirmed $20.35M total raised. Databases variously report $19.9M across three rounds, $20.36M, $27M, $28.7M and GBP 21M, and TipRanks reports $9M across one round, which is plainly wrong since it omits even the 2024 $8.5M seed. Adjudicated: $18.5M is the size of the seed round and $20.35M is the cumulative total including the 2022 pre-seed. The two differ by exactly the pre-seed amount and are therefore internally consistent. The higher database figures have no primary source; the TipRanks figure is an error.

2. Currency conversion of the seed round. $18.5M against GBP 13.7M (StartupMag), GBP 13.8M (PE Deals) and GBP 14M (BusinessCloud) is consistent. EU-Startups and Prime World Media give GBP 15.7M and EUR 15.7M, which appears to be a conversion and addition error that treats the $10M extension as EUR 8.4M and adds it directly to $18.5M. Adjudicated: the GBP 13.7M to 13.8M range is used; GBP 15.7M and EUR 15.7M are excluded as European-edition conversion errors.

3. Round naming. The company calls the February 2026 event a seed extension and states that it closes the seed round at $18.5M. Seedtable and PE Deals both classify it as a Series A. Adjudicated: the official naming, seed extension, is used, with a note that some databases auto-classify it as a Series A. Nothing in the company's disclosure supports the Series A label.

4. The nature of the STRATA arrangement. The Honeywell press release of 28 August 2025 states a GBP 14.1M (approximately $19M) three-year programme funded through the ATI Programme, delivered by the Aerospace Technology Institute with the Department for Business and Trade and Innovate UK, led by Honeywell from Yeovil, with five members. BeyondMath's own release says it was "selected to partner with Honeywell" and that "the consortium comprises industry leaders". StartupIntros and Tech Funding News render it as "a $19 million three-year commercial partnership" and as "securing a $19 million three-year commercial partnership". Adjudicated: this is a consortium programme budget and a consortium membership, not a Honeywell contract to BeyondMath. BeyondMath's own share is not disclosed. The partnership framing in secondary media is a misreading.

5. Platform scale specification. The STRATA release of 28 August 2025 states the platform runs 500 million cell 3D transient simulations. The current company home page states 200 million cells. Adjudicated: both are retained, read as a 2025 peak capability claim and the 2026 currently promoted specification respectively. Neither is treated as a correction of the other without further evidence.

6. How many Formula One customers. Sifted in August 2023 refers to a fine-tuned system for a Formula One racing team, singular. TechCrunch in August 2024 refers to several F1 teams already testing the software. Several databases write it as Formula One teams, plural. Adjudicated: the TechCrunch formulation, that several F1 teams are testing, is used, because it is both the more recent and the more conservative statement, and testing is not the same as being a customer.

7. Team size. Approximately 15 people (BusinessLife, 2024), 18 (StartupIntros), 21 (PrivateMarketView, October 2026), 25 (TipRanks, 14 September 2026), an 11 to 50 band (LinkedIn), 18 with a near-doubling over the year and 10 nationalities (Tech Funding News quoting the CEO, February 2026), and an official plan to double headcount during 2026. Adjudicated: the report states a range of roughly 20 to 30 people, treats the official doubling as a plan rather than an accomplished fact, and cites the 10-nationalities figure from the CEO as the only firm diversity data point.

A related item is recorded here because it is a figure that circulates widely: the $32.9M valuation. It is published by PrivateMarketView with the label "last reported valuation as of April 2024", is self-flagged by that platform as LOW confidence, and is described by it as the only thing ever priced. Implied by it, an $18.5M seed round against a $32.9M valuation would leave seed investors with about 22% in aggregate, which conflicts sharply with the equity structures reported by other databases, including a Seedtable estimate of 15% for UP.Partners and 15% for Insight Partners, which is itself labelled an estimate rather than a filed cap table. The valuation is therefore not used as a fact in this report. The official position is that valuation has not been disclosed.

Risks and Diligence Gaps

Every performance claim lacks independent reproduction. The 1000x speedup, the 200 million and 500 million cell figures, and the "world's largest foundational physics model" framing are all company statements. The only externally published accuracy data is the ATI and Capgemini case at 1.5 minutes against 24 hours with accuracy within 2% of detached eddy simulation, and that case is automotive aerodynamics rather than aerospace certification. The DrivAerML comparison set is academic models, Transolver and Transolver-3, plus AB-UPT, not production-grade industrial CFD, and the company concedes one metric. The diligence ask is specific: the train, validation and test split for DrivAerML, whether any third party has reproduced the benchmark, and blind test results on customer-owned data.

The precision-to-certification gap is the company's own admission. In 2023 Garvey said the model was around 0.8% less accurate than classical supercomputing and that an aircraft manufacturer such as Boeing would need more accuracy and that this might take about five years. That window runs to around 2028, and nothing certification-related has appeared publicly in the meantime, including no recognised validation methodology, no verification-and-validation process and no recorded regulatory engagement. Safety-critical adoption in aerospace and automotive is the ceiling on this business, and the gap is acknowledged by the founder rather than disputed by a critic.

The generalisation claim contradicts itself. The official position is that a foundation model trained on the laws of physics generalises to unseen designs and does not need large quantities of customer data. In the same 2023 period Garvey said that if a model understands cars, or car-like objects, it is not necessarily going to understand a plane, or a blood vessel. In the intake manifold case, 259 of 264 cases were in the training set and 5 were held out, which is an extremely small hold-out. The diligence ask is quantified zero-shot transfer evidence across automotive, aerospace and electronics.

The business model is early and part of it is services. There is no public pricing, no revenue, no named customer beyond a consortium seat and an anonymous F1 team, and no ARR. The likely revenue mix is platform subscription plus model customisation through Physics Factory plus consortium participation, and the customisation component carries a labour-intensive services character that would weigh on gross margin and on any software multiple. The diligence ask is the split between recurring and services revenue, and the contract terms and exclusivity of the ten largest customers.

Strategic bid risk is live in both directions. Three acquisitions have already happened in this cohort: Emmi AI into Mistral, Monolith into CoreWeave, and NAVASTO into Autodesk. BeyondMath is one of the few UK targets that remains independent. In the other direction, Mistral opened a Physics AI centre in Munich on 28 September 2026 with work alongside BMW, Siemens Energy and the Technical University of Munich wind tunnel, which is direct competitive pressure on European soil and not merely a financial event. PhysicsX, in the same city, with $300M of Series C funding, is the principal source of capital and ecosystem squeeze.

Scale and capital efficiency. The company raised $10M as recently as February 2026, runs a 20 to 30-person team, plans to double headcount during 2026, carries DGX H200-class hardware costs, and is expanding into US and Japanese customers. Caplight records open roles rising from 9 in late March 2026 to 14 by early April, an unusually aggressive hiring rate against headcount. The CEO's own framing is "more compute, more GPUs". A $18.5M seed round with this cost structure can plausibly run into a financing gap or a forced slowdown inside 18 to 24 months.

Data compliance and model-weight ownership in aerospace. The on-premise deployment posture addresses intellectual property and data residency, which are genuine procurement requirements in aerospace. But the models also need customer CFD data to train. If a customer contributes data, the ownership of the resulting model weights and the rights to use them are the terms that decide whether a customer can ever take its models elsewhere. These terms need to be read.

Customer concentration and a narrative that leans on Formula One. The brand narrative leans heavily on F1. Every customer quotation, the STRATA release and the CEO's statements all invoke Formula One experience. If the relationship ends, or if the identity of the team turns out to be a competitor building its own AI CFD capability, the coupling between narrative and revenue becomes visible. The company anonymises the team as "Head of CFD at an F1 team", and markets in a way that implicitly pressures F1 teams that do not use the product. The anonymity may also indicate exclusivity or confidentiality clauses, which is speculation and should be asked directly.

Information quality in commercial databases. Seven different sets of financing, valuation and headcount numbers circulate across the databases, including a plainly incorrect $9M total raised at TipRanks, $27M at Parsers.vc and $28.7M at PrivateMarketView. Any competitive comparison table built from these sources has to be recomputed from primary documents, which is why this report grades every source.

Overall Assessment

• Technology 7.0. Neural operators trained on first principles rather than GPU-accelerated classical solvers, DGX H200-class training, a genuinely complete published benchmark with full comparison tables and a conceded losing row, an autonomous agent-crew build pipeline with under 30 minutes of human time, a 1.5-minute against 24-hour result in a third-party institute report, and a disciplined internal-flow disclosure with a stated hold-out. Deducted for the absence of any third-party reproduction, the "up to 1000x" claim that carries no benchmark configuration, and the fact that the strongest external case is automotive aerodynamics rather than aerospace certification.
• Commercialisation 5.5. A real product surface with a flagship platform, five agent crews, an AI CAD module, an early-access program and on-premise deployment that addresses a genuine aerospace procurement requirement, plus ecosystem validation through NVIDIA, AWS and Siemens Xcelerator. Deducted for no disclosed pricing, no disclosed revenue, no ARR, no named customer beyond a consortium seat and an anonymous Formula One team, and a model-customisation revenue line with a labour-intensive services character.
• Funding 7.0. A GBP 1.5M pre-seed, an $8.5M seed and a $10M extension for $18.5M across the seed round and $20.35M cumulative, with a credible investor roster spanning UP.Partners, Insight Partners, InMotion Ventures and Cambridge Innovation Capital, and a lead investor that states its thesis in public. Deducted for no valuation disclosure, a headquarters relocation from Cambridge to London mid-flight, and no evidence of a follow-on round beyond February 2026 against a rising compute bill.
• Product maturity 6.0. A coherent ladder from Digital Wind Tunnel through GPStudio and Generative Physics Studio to Physics Factory and Cadgen2, with published throughput numbers for each and a working on-premise deployment story. Deducted because the flagship platform is the newest layer, the automotive product remains in early access, the customer-facing catalogue is largely unpriced, and the platform's scale specification is inconsistent between 2025 and 2026 disclosures.
• Market position 5.5. A published benchmark leadership position against academic reference models, a place in the Siemens Xcelerator ecosystem catalogue, membership of a five-party ATI-funded aerospace consortium led by Honeywell, early NVIDIA Inception status, and a UK base in a category that has already attracted three acquisitions. Deducted for a single unnamed Formula One relationship behind the entire customer narrative, category-level funding that is one to two orders of magnitude larger at PhysicsX and Mistral, and the absence of any certification pathway that would unlock the safety-critical vertical.

References

1. BeyondMath seed-extension announcement, 25 February 2026, beyondmath.com/news/seed-round-announcement, including the $10M extension, the $18.5M seed close, the founder history and the use of funds.
2. BeyondMath developer blog, including "Built autonomously: 1st on DrivAerML", 3 August 2026, and "Internal flow engines", 21 October 2026.
3. BeyondMath case study, Project STRATA, beyondmath.com/case-studies/project-strata-revoluntionizing-aerospace-through-generative-physics.
4. Honeywell Aerospace press release, UK aerospace manufacturing innovation programme, 28 August 2025, aerospace.honeywell.com.
5. diginomica, "How Aerospace Technology Institute clears AI use cases", 24 August 2026, and the joint ATI and Capgemini report "AI for Aerospace", capgemini.com/gb-en/insights/research-library/ai-in-aerospace.
6. Sifted, "Move over Oppenheimer, AI laws of physics", August 2023, for the 0.8% accuracy gap and the five-year aerospace horizon.
7. TechCrunch, 20 August 2024, on the Digital Wind Tunnel launch and on Formula One teams testing the software.
8. Tech Funding News, February 2026, and Pathfounders, February 2026, for the cumulative $20.35M confirmed by the CEO, the refusal to disclose valuation, the pre-seed amount and the founders' fourth project together.
9. Cambridge Innovation Capital investment announcement, cic.vc, and InMotion Ventures investment note, inmotionventures.com, for the neural-operator characterisation and the HomeX and Jaguar Land Rover relationships.
10. Siemens Xcelerator ecosystem entry for BeyondMath, siemens.com, and the ATI awards listing, ati.org.uk.
11. Companies House-derived company record for BeyondMath Limited, company number 14110252, plus the UK Worker sponsor licence record, for incorporation dates, directors and licence status.
12. Aerospace Technology Institute, ATI Programme documentation, and the FlightGlobal and Digital Engineering 24/7 coverage of the STRATA consortium and its five members.

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.