PhysicsX: Pretrained Geometry and Physics for Industrial Engineering at a $2.4B Mark

Executive Summary

PhysicsX is a London-based company, founded in 2019 and headquartered at Victoria House on Leonard Circus, that sells what it calls an AI-native engineering platform: a stack that replaces much of the numerical simulation in an industrial design loop with pretrained models.

Three facts frame the investment case.

First, the training-data position is unusually concrete. The company's aerodynamic model, LGM-Aero, is trained on a corpus of more than 25 million geometries and associated physics simulation, containing tens of billions of mesh elements and tens of thousands of CFD and FEA runs produced with Siemens Simcenter STAR-CCM+ and Simcenter Nastran. Siemens has publicly confirmed the collaboration and publishes the corpus figures itself. This is a genuine structural advantage, because the expensive part of training a physics surrogate is not the model but the high-fidelity simulation data behind it, and this corpus came from a commercial solver vendor's paid industrial analyses.

Second, the capital position is strong. PhysicsX closed an oversubscribed $300M Series C at a valuation of roughly $2.4B on June 8, 2026, led by Temasek, with new investors M&G Investments and Intrepid Growth Partners alongside existing investors Applied Materials, Atomico, General Catalyst, July Fund, NGP, NVIDIA, Radius and Siemens. The previous round was a $135M Series B in June 2025 at a valuation just under $1B, so the mark has moved roughly 2.4x in twelve months. Headcount is above 300, roughly double the level of a year earlier, and recognised revenue has doubled year over year with booked revenue tripled.

Third, the generalisation evidence is thinner than the market narrative. The headline speed figures — 10,000x to nearly 1,000,000x in some vendor statements, and 10^4 to 10^6 in Aviation Week's reporting of the CEO's own framing — describe total solution time against a full high-fidelity simulation, which includes geometry preparation, meshing and solver setup. The physically meaningful comparison is against the solver's solve phase alone. No peer-reviewed accuracy benchmark for the platform has been published, and the company reports customer outcomes in relative terms rather than absolute error figures.

Our view: strong data and capital position, unproven accuracy envelope. This company is worth studying because it makes a specific bet: that the scarce asset in industrial AI is not the model architecture but access to a corpus of high-fidelity simulation results that only an incumbent solver vendor's customer base can generate. If that is right, the winning position in this category will sit with whoever controls the simulation data rather than whoever writes the best surrogate. The open question is whether a platform trained on one vendor's solver output can hold its accuracy when applied to another vendor's solver, because that is the condition under which AI-native simulation displaces legacy CAE at scale rather than in a single toolchain.

Company Overview

PhysicsX Limited is registered in England and Wales, incorporated on August 1, 2019. It is headquartered in London, with an office in New York, and an expanding presence in the Bay Area and Singapore. An Asia-Pacific hub opened on Robinson Road in Singapore on October 8, 2026, with about 20 staff and a stated target of more than 100 within three years, including forward-deployed engineers, physics AI researchers and simulation engineers.

The founding team is unusual for the category, and it explains the product's character:

• Jacomo Corbo, Co-Founder and CEO — former co-founder and chief scientist at QuantumBlack, and former chief race strategist for the Renault/Alpine Formula 1 team. A founding team of this shape brings a strategy and optimisation mindset rather than a CAE pedigree.
• Robin Tuluie, Founder and Chairman — previously in research and vehicle technology roles at Renault/Alpine F1, Mercedes F1 and Bentley.

The leadership pairing is worth reading carefully. Neither founder comes from a computational mechanics background, which explains both the company's distinctive position and its main structural risk. The upside is that the company is not trying to make a better solver; it is trying to change what an engineering tool is. The downside is that the credibility burden falls entirely on the modelling team and on external validation.

The company describes itself as the physics AI company for industrials, with a mission to accelerate hardware innovation by overhauling industrial engineering and manufacturing. Its product framing covers the full lifecycle: early-stage design, manufacturing process optimisation, and real-time digital twins in operation.

Technology Analysis

The model family

PhysicsX's core abstraction is the Large Physics Model (LPM), a deliberate analogy to large language models in which the training corpus is simulation results and physical equations rather than text. Depending on the problem, the models draw on transformer architectures, geometric deep learning, or neural operators approximating solutions to partial differential equations. The output is a model that has internalised how a class of physical systems behaves and generalises to design variants it has not seen.

Alongside the LPM sits the Large Geometry Model (LGM), which handles the geometry generation problem. This split is the architectural core of the company's approach and it deserves attention on its own: it treats geometry as a learned generative object rather than a fixed input, which is what makes the zero-shot claim possible for new shapes.

LGM-Aero: the corpus

LGM-Aero is the most documented artifact. Siemens Digital Industries Software has publicly described the collaboration and states the corpus figures directly:

• More than 25 million distinct geometries with associated physics simulation.
• Tens of billions of mesh elements.
• Tens of thousands of CFD and FEA simulations, generated with Simcenter STAR-CCM+ and Simcenter Nastran.
• Fine-tuning with as few as a few tens of simulations is sufficient for a customer to adapt the deep learning model to its own application.

The model itself has been reported at roughly 100M parameters, with geometry compressed to a 512-dimensional representation. Training compute ran on AWS Batch and EC2, with the technical stack built on Microsoft Azure.

Why the corpus provenance matters. This is the single most consequential detail in the company's public technical record, and it cuts in a specific direction. A simulation corpus generated for industrial customers by a commercial solver vendor is not a random sample of shape space. It is a record of the shapes that paying industrial customers asked to be analysed — with high density exactly where engineering attention and budget already flow. The model is therefore most confident in regions of design space that the incumbent industry has already paid to explore. That is an advantage over a generic surrogate trained on random geometry, and it is also a structural limit on how far the model generalises into territory the industry has not needed to model.

Ai.rplane: the public demonstrator

Ai.rplane is a free, public-access technology demonstrator built on LGM-Aero, described as the first LPM for flight. It generates aircraft geometry and infers aerodynamic performance, flight stability and structural stress for arbitrary flying shapes, delivering geometry and results in under a second against the several hours a traditional numerical simulation would require. It is hosted at airplane.physicsx.ai.

The value of Ai.rplane is not the demo itself; it is that it lets a third party test the zero-shot claim without a commercial relationship. That is a meaningful disclosure choice, and it is also the only route by which outside researchers can form an independent view of the accuracy envelope rather than reading vendor figures.

The accuracy question, and what the independent evidence shows

The publicly stated speed advantage ranges widely. The CEO has described the platform as delivering "10,000 to nearly 1 million times" faster, and Aviation Week reported the same framing as 10^4 to 10^6. These figures compare against the total elapsed time of a complete simulation workflow — geometry handling, meshing, setup, solve and post-processing — which is where most of the wall-clock time in industrial CFD actually sits.

The comparison that matters for an engineering decision is different: it is the model's prediction against the solver's solve phase on the same geometry. On that comparison, the independent evidence available so far is instructive rather than promotional.

The SIMSHIFT benchmark (Setinek et al., arXiv:2506.12007) is the most directly relevant published work on out-of-distribution behaviour for industrial neural surrogates. It evaluates four industrial tasks — hot rolling, sheet metal forming, electric motor design and heatsink design — with state-of-the-art surrogates including geometry-informed neural operators. The headline finding is that surrogate performance degrades by roughly an order of magnitude outside the training distribution, and that domain adaptation techniques recover only about a quarter to a third of the resulting gap.

Two caveats apply in fairness. SIMSHIFT is not a benchmark of PhysicsX's platform, and its authors are affiliated with Emmi AI, a separate company in this space. And a surrogate that has been trained on 25 million geometries is operating from a very different starting point than the models SIMSHIFT evaluates. The relevant conclusion is the mechanism rather than the number: the gap between in-distribution and out-of-distribution accuracy is a property of the surrogate formulation and the data distribution, not of the vendor.

A useful comparison point from adjacent work is Luminary Cloud, which reports that after training on thousands of simulations its models remain within a low-single-digit percentage of solver output, saving close to $10,000 of compute per iteration. That is a different architecture with different training data, and it is included here as a reference for what a credible accuracy claim looks like in this market, not as a comparison of the two companies.

The independent validation gap. The public record contains no peer-reviewed accuracy benchmark for the PhysicsX platform, and no absolute error figure from a customer-run evaluation. Most company metrics are relative growth rates rather than absolute numbers. This is a common pattern in the category and it does not imply the technology is weak — it means the accuracy envelope is not publicly characterised.

Siemens as partner, not merely as data source

The relationship with Siemens deserves separate treatment because it is easy to misread. Siemens is simultaneously a shareholder, a commercial partner, and the source of the training corpus.

What is publicly established: Siemens Digital Industries Software states the collaboration, publishes the corpus description, and quotes its own senior vice president for simulation and test solutions on AI's implications for the simulation community. Siemens' framing is about open collaboration and about the AI-enhanced simulation industry generally.

What this means commercially is that a major incumbent is publicly validating a startup's approach to its own market. For PhysicsX, that relationship supplies training data, credibility and a channel into industrial accounts. For the market, it is a signal about how incumbents are responding to this category.

Products and Commercialization

Customer base and named references

The company names customers across its target industries, and these are the ones referenced publicly: Applied Materials, Siemens, Stellantis, Leonardo Aerospace, Rio Tinto and Microsoft Surface (thermal management for cooling fans). Seatrium, an offshore energy engineering company, has publicly stated that the platform reduced the time required to optimise pipe-layout designs by approximately 40%.

That list is unusual in public disclosure terms. Naming customers in a regulated-industry context like aerospace and defence is a harder disclosure than it is in software, because an aerospace prime's use of a new analysis tool can imply a level of design authority that may not exist. The list should be read as reference relationships as well as production deployments, and the distinction between the two is not publicly resolved.

Revenue and backlog

The commercial metrics disclosed alongside the Series C are the most concrete numbers available:

• Recognised revenue doubled year over year.
• Booked revenue tripled.
• Customer count more than doubled over the past year.
• Revenue expected to be close to $50M for the year, with a stated plan to more than double that in 2027.
• Roughly six months of customer backlog, which the CEO described as "a little untenable for us" and cited as a reason for the capital raise.
• Semiconductors on track to become the largest business segment by the end of Q2, overtaking the aerospace and automotive base the company started from.

The backlog figure is the one that deserves the most attention, because it distinguishes this company from the standard vertical-AI pattern. A six-month queue of customer demand means the constraint on growth is delivery capacity and hiring rather than demand generation. It also implies a meaningful risk: a backlog is only an asset if it converts, and conversion depends on engineers being available to support deployments.

Deployment and integration

The described customer journey runs through technical evaluation and data scope definition, workflow integration with systems including ANSYS, CATIA, Siemens NX, OpenFOAM, AWS and Azure in hybrid and on-premise environments, project deployment, result validation, and expansion across accounts and regions.

The integration list is the strategically important detail. PhysicsX integrates with Ansys, CATIA, Siemens NX and OpenFOAM — that is, with the incumbent toolchain — while its training data came from Siemens. The company is positioning itself as a layer that sits alongside the installed base rather than as a replacement that requires the installed base to be retired. For an enterprise buyer, that is the difference between an incremental line item and a multi-year platform replacement programme.

It also creates the company's central strategic ambiguity, discussed further below.

Regional expansion and public-sector partnership

Singapore is the stated Asian hub, supported by Singapore's Economic Development Board, with the Asia-Pacific headcount target of more than 100 within three years. Separately, PhysicsX and A*STAR announced a five-year, S$40M partnership covering maritime engineering, advanced manufacturing, semiconductor manufacturing and biotechnology. Singapore's Deputy Prime Minister and Minister for Energy and Trade and Industry Gan Kim Yong made a public statement of support for the company as part of the National AI Missions framing.

The A*STAR arrangement is worth noting on its own terms. It is a multi-year, publicly-funded research partnership with a national research body, which provides institutional validation and a research pipeline, and which also commits the company to delivering against a public agenda over a five-year horizon.

Pricing

Pricing is not public. The deployment model described — integration with incumbent CAE stacks, hybrid and on-premise deployment, enterprise accounts — is consistent with a negotiated enterprise software arrangement rather than self-serve pricing.

Competitive Landscape

The incumbents

The engineering simulation market that PhysicsX targets is dominated by Ansys, Siemens (Simcenter), Cadence, Synopsys, and Dassault Systèmes across CAD and simulation. These vendors have decades of solver validation, established qualification records with regulators and customers, and deep embedding in engineering processes — a position that is measured in decades rather than quarters.

The relevant observation is not that these products are static. Each of them has been investing in AI-assisted simulation features. The relevant observation is that PhysicsX's training corpus came from one of them, which means the incumbents retain the ability to generate comparable or better corpora from their own customer bases. Siemens has done exactly this with PhysicsX. The strategic question is whether a solver vendor trains models on its own simulation output and internalises that capability, or whether it licenses the capability from a specialist partner. Both paths are open, and the second is what PhysicsX is currently on.

The AI-native cohort

PhysicsX is one of the better-capitalised companies in the physics-AI group, alongside:

• Vinci AI — $250M Series B at a $1.5B valuation in October 2026, building a physics foundation model with GPU-native kernels for semiconductor thermal and thermo-mechanical analysis.
• Luminary Cloud — $72M raised with NVIDIA venture backing, focused on simulation for aerospace.
• Neural Concept — $100M raised, focused on aerodynamic design.
• Emmi AI — based in Linz, Austria, working on AI-driven simulation for industrial applications.

The capital comparison against Vinci AI is instructive. Vinci raised at a $1.5B valuation four months after PhysicsX raised at $2.4B, into a narrower and more recently validated product surface — a single semiconductor thermal domain with a peer-reviewed paper, against PhysicsX's multi-domain platform across six industries. The gap is explained by breadth of deployment, capital scale, and the strategic shareholder base, not by evidence quality in either direction.

The newest entrant, and the reason the category is intensifying

Project Prometheus, a new AI lab backed by Jeff Bezos, closed a $10B funding round in April 2026, focused on physical AI for industrial and manufacturing applications. Several other companies are developing AI world models aimed at similar industrial use cases.

A $10B raise for a company in this category is a first-order market-structure event, and it deserves a plain reading rather than a hyperbolic one. It establishes that very large pools of capital now consider industrial physical simulation to be a strategic category. It also raises the bar for everyone in the field on two specific dimensions: capital availability and ambition about scope. PhysicsX at $300M and $2.4B is well capitalised relative to its cohort, but the cohort's capital ceiling moved by an order of magnitude within the same year.

Where the differentiation actually sits

Three claims separate PhysicsX from incumbent tools, and each has a different strength.

Control of high-fidelity simulation data is the strongest claim, and it is the one with a structural rather than a rhetorical basis. High-fidelity CFD and FEA runs are expensive to produce; producing them at industrial quality requires exactly the solver stack and engineering effort that a company like PhysicsX has spent years building relationships around. Access to a corpus is not reproducible by a competitor who lacks the relationships, and the corpus is also what improves with each industrial engagement.

Geometry as a learned generative object is the claim with the clearest architectural novelty. Treating geometry as generated rather than given is what makes zero-shot operation on unseen shapes possible, and it is what separates the LGM/LPM split from a conventional surrogate that maps a given mesh to a given field.

Speed is the claim most exposed to scrutiny, for the reason noted above: the headline multiples compare against total workflow time rather than against solver time, and the marginal benefit to an engineer depends on how much of an iteration cycle simulation actually occupies. In a concept-design phase, where PhysicsX explicitly targets the stage that conventional CAE generally does not reach, the benefit is real and structurally large. In a validation phase, where a certified solver run is required anyway, it is much smaller.

Funding and Team

The funding history is short, large and steep:

• 2019: founded in London.
• June 2025: $135M Series B at a valuation just under $1B, with Temasek investing for the first time.
• November 2025: NVIDIA's venture arm NVentures added $20M, bringing the total for that round to more than $155M.
• June 8, 2026: $300M Series C at a valuation of approximately $2.4B, oversubscribed, led by Temasek. New investors: M&G Investments, Intrepid Growth Partners. Existing investors increased participation: Applied Materials, Atomico, General Catalyst, July Fund, NGP, NVIDIA, Radius, Siemens.
• Cumulative disclosed funding is approximately $487M (a $32M Series A, a $135M Series B, a $20M Series B extension and the $300M Series C).

The investor list reads as a strategic map. NVIDIA supplies the compute that trains and runs the models. Siemens is both a data source, a commercial partner and a shareholder. Applied Materials sells the equipment that builds the semiconductor tooling PhysicsX customers design — and is simultaneously a customer. Temasek, leading both the B and C rounds, provides long-horizon capital and, as the Series B coverage noted, a path into Asian industrial customers that European startups rarely access. The presence of a semiconductor equipment vendor both as customer and investor is an unusually direct form of strategic alignment in this category.

The valuation step is the key financial fact. From just under $1B to approximately $2.4B in twelve months is close to a 2.4x increase. That move is significant in a private market, but the more telling detail is how little dilution it took: the $300M Series C represents roughly 12.5% of the $2.4B post-money, close to the proportion the $135M Series B took of its then ~$1B mark. Both rounds were lightly dilutive, which is how the mark rose 2.4x without a large cash injection forcing it. The valuation is therefore a bet on the platform's reach across the industrial lifecycle, not only on the shipped models: it prices a company with roughly $50M of annual revenue against a delivery backlog and a platform whose accuracy envelope is not publicly characterised in absolute terms.

Headcount is above 300, having doubled in twelve months, spread across London, New York, the Bay Area, Singapore and Bangalore. The hiring profile the company describes is demanding — it requires simulation engineering (FEA/CFD) together with 3D graph or point-cloud geometric deep learning and MLOps, which is a rare combination to assemble at speed. The Singapore target of more than 100 staff within three years, including forward-deployed engineers, is a meaningful commitment against that constraint.

Risks and Challenges

The accuracy envelope is not publicly characterised

This is the central technical risk. Every speed and accuracy figure in the public record originates from the company, and no peer-reviewed or customer-run absolute error measurement has been published. The SIMSHIFT results on neighbouring architectures indicate that out-of-distribution degradation of roughly an order of magnitude is a common property of industrial neural surrogates, and that domain adaptation recovers only a portion of that gap.

The mechanism that would mitigate this — a corpus of 25 million geometries with tens of billions of mesh elements — is genuinely large, and it plausibly makes PhysicsX's models more in-distribution-robust than typical surrogates. It is not a substitute for published error bars on specific applications. The single most informative future artifact would be a third-party benchmark on a specific industrial problem, with absolute error stated.

The corpus concentration creates a strategic ambiguity

PhysicsX's training data came from Siemens Simcenter. The company simultaneously integrates with Ansys, CATIA, Siemens NX and OpenFOAM, and Siemens is a shareholder. This is a coherent commercial position today, and it also creates a structural question: what happens to the accuracy of a model trained on one solver's output when it is applied to another solver's problem?

If PhysicsX's models hold accuracy across solver ecosystems, that is a strong position and a genuine displacement of incumbent CAE at scale. If they are tied to the Simcenter data distribution, the business is closer to a highly effective partner to one ecosystem than a platform that replaces CAE across the board. The public record does not yet distinguish these cases, and the distinction determines the size of the addressable market.

The scale of the capital response

Project Prometheus closing a $10B round in April 2026 changed the capital context for the entire category. PhysicsX's own raise was substantial, and its backlog is a real constraint on growth. But a competitor with an order of magnitude more capital can fund longer engineering cycles, absorb a longer qualification horizon, or acquire the position. In a category where the incumbent's moat is qualification history and the newcomer's moat is data access, the decisive contest may be over who can afford to wait.

Backlog conversion

A six-month backlog is a good problem to have and a real operational risk. Conversion depends on forward-deployed engineers, on customer data access, and on the qualification process each customer runs internally. If onboarding capacity cannot scale at the rate the backlog requires, the constraint reappears as deployment delay rather than as lost demand — which damages the delivery metric rather than the order book.

Concentration in a small number of large accounts

The named customers are a short list of large industrial organisations. This is normal for the category and it produces very high contract values, but it also means the revenue base is exposed to the capital-expenditure cycles of a handful of companies. The stated shift of the largest segment from aerospace and automotive to semiconductors by the end of Q2 concentrates the exposure further, into a sector whose own cycles are pronounced.

The founder-background question

Neither founder comes from computational mechanics. This is a deliberate choice rather than a gap the company has to apologise for — it is the source of the company's distinctive market framing — but it places the entire engineering-credibility burden on the technical organisation and on external validation. For an aerospace or semiconductor customer whose sign-off depends on traceable accuracy, the quality of the modelling team matters as much as the benchmark results.

Assessment

Technology: 8.5 — a corpus of more than 25 million geometries with tens of billions of mesh elements and tens of thousands of high-fidelity runs, a public zero-shot demonstrator in Ai.rplane, and an architectural separation of geometry generation from physics inference that is genuinely novel. Deducted points for the absence of any published absolute error figure, no peer-reviewed accuracy benchmark for the platform, and speed claims that compare against total workflow time rather than solver time.

Commercialization: 7.5 — named customers across six industries including aerospace primes and a semiconductor equipment vendor, a disclosed revenue figure near $50M, revenue doubling and booked revenue tripling year over year, a six-month backlog, and semiconductor overtaking aerospace as the largest segment. Deducted points because deployed-versus-referenced relationships are not distinguished publicly, deployment models sit alongside incumbent toolchains rather than replacing them, and the absolute contract economics are undisclosed.

Funding: 9.5 — $487M raised in total, an oversubscribed $300M Series C at a $2.4B valuation, and an unusually strategic shareholder register spanning compute, industrial software, semiconductor equipment and sovereign capital. Deducted points only for the absence of a revenue base that would let the valuation be tested against delivery rather than against plan.

Product maturity: 7.5 — a documented model family (LPM and LGM), a production platform spanning six industries with real-world customer outcomes reported, and a live public demonstrator. Deducted points because the platform's accuracy envelope is publicly uncharacterised, several lifecycle stages are described as aspirations rather than shipped capability, and a 300-person team covering six industries and multiple physics domains is a broad operating surface.

Market position: 8.0 — one of the two strongest positions in the physics-AI category by capitalisation and deployment breadth, with a strategic shareholder base that is effectively a distribution network, and with Temasek's Asian channel addressing a structural weakness of European deeptech. Deducted points because the valuation prices a far broader platform than the currently evidenced revenue, category capital has stepped up by an order of magnitude, and the accuracy claim that would constitute a durable moat remains unquantified.

References

1. PhysicsX, "PhysicsX Announces $300M Series C to Accelerate Physics AI for Industrial Engineering" (June 8, 2026)
2. Siemens Digital Industries Software, "Siemens & PhysicsX collaborate to build AI-based deep physics simulation"
3. The Straits Times, "Deeptech unicorn PhysicsX opens regional hub in S'pore to scale AI-powered industrial engineering" (October 8, 2026)
4. Bloomberg, coverage of the Series C round at a $2.4B valuation (June 8, 2026)
5. greyjournal, "PhysicsX Raises $300M Series C at $2.4B Valuation" (June 2026)
6. InforCapital, "PhysicsX Raises $300M, Valuation Hits $2.4 Billion"
7. Aviation Week, interview with CEO Jacomo Corbo on speed claims and model architecture
8. Setinek et al., "SIMSHIFT: A Benchmark for Adapting Neural Surrogates to Distribution Shifts," arXiv:2506.12007
9. PhysicsX and A*STAR, five-year S$40M partnership announcement (October 2026)

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