Thesis
Manufacturers use computer-aided engineering (CAE) software to test how aircraft, vehicles, and other products will perform before building them. The global CAE market was estimated at $11.6 billion in 2025, and aerospace was its fastest-growing manufacturing sector, expanding at a 14.9% CAGR from 2021 to 2025. Much of that software still relies on the finite element method, a numerical technique that dates to the 1940s. Running these calculations can slow product development. For instance, before General Motors began adding large physics models to its car design process around late 2025, the simulation phase used to calculate a new concept’s drag took about two weeks.
Physics AI models are designed to speed up that work. Trained on simulation and test data, they predict how a physical system will behave without solving the underlying equations for each new design. As of April 2026, General Motors had trained an in-house model on its aerodynamics simulation results that could return a car’s drag coefficient in minutes. Many engineering teams have started using similar tools, but few rely on them heavily. In a vendor-sponsored survey of 350 engineering leaders, 92% reported deploying surrogate models, which approximate full simulations, in 2025, but only 8% of respondents said those models were extensively deployed in their workflows. As of 2026, more than 120 startups and small companies were building AI for CAE, backed by about $2.3 billion of venture capital.
The AI infrastructure buildout is adding to the volume of hardware that has to be engineered. Five of the largest AI infrastructure spenders were forecast to put $803 billion to $828 billion into total CapEx in 2026, and global spending on 300mm chip fab equipment was projected to rise 31% from 2025 to $149 billion in 2026. Data center equipment has to be engineered against power density, cooling, and interconnect requirements, while advanced cooling equipment can take several years to arrive. Cambashi expects the nearest-term value of AI in CAE to come from making scarce simulation engineers more productive.
PhysicsX develops AI models that predict how physical systems behave, along with a platform for training those models and putting them to work in design, manufacturing, and operations. The company claims its models run 10K to 1 million times faster than traditional simulation, and it sends its own engineers to work alongside customers. Its goal is to let engineers test designs with AI predictions instead of waiting for a full simulation of each one, using pre-trained models and its own teams to shorten the time between a design idea and a tested result.
Weekly Newsletter
Subscribe to the Research Rundown
Founding Story

Source: Atomico
PhysicsX was founded in August 2019 in London by Robin Tuluie, PhD (Chairman, former CEO) and Nico Haag (Director of Simulation Engineering). It was incorporated as Motodynamics Ltd and took the PhysicsX name in July 2020. Jacomo Corbo, PhD (CEO), joined as co-CEO in February 2023, and the company has called him a co-founder since its November 2023 Series A.
Tuluie is a theoretical physicist who discovered a small-scale effect in the cosmic microwave background as a postdoctoral scholar at Penn State. He has said he moved from academic physics into motorsport out of impatience with cosmology's slow feedback loop, "waiting for decades for observations to validate your models," compared with Formula 1, where "you find out every weekend whether what you've developed beforehand was right or not." He led R&D at Renault F1 from 2004 to 2011, during a period that included back-to-back double world championships, then held the same role at Mercedes F1 before becoming Vehicle Technology Director at Bentley Motors in 2016.
At Bentley, Tuluie said his team built "an end-to-end numerical simulation capability." He then founded an engineering consultancy, and he said in March 2026 that “when we hit the limits of scale, I left to start PhysicsX with Nicolas Haag, a brilliant young engineer who became my co-founder.” Haag worked at Bentley until September 2019 after earlier roles at Audi Sport and Mercedes-Benz, and holds an MSc from the Karlsruhe Institute of Technology.
Tuluie has said the founding belief that led to the creation of PhysicsX was that the best numerical simulation and machine learning could let engineering teams develop products at unprecedented speed. A typical car takes five to six years to go from concept to production, he noted, while a Formula 1 car takes one. When geometric deep learning began showing real-world potential, Tuluie said, the team saw a way to take that work forward.
Corbo met Tuluie in Formula 1. While completing a Harvard PhD under David Parkes, who is dean of Harvard's engineering school as of November 2026, he joined Renault's F1 team and served as Chief Race Strategist from 2006 to 2008. He said in 2025 that the specialists he met there "only really understood how to model things and handle data in very traditional ways," an observation he called "very much the thesis for starting" QuantumBlack, the AI consultancy he co-founded. McKinsey acquired QuantumBlack in 2015, and PhysicsX has said the group had grown to more than 1.5K technical practitioners by February 2023. PhysicsX said that month that Corbo had "collaborated deeply" with the company since its founding, and he has described joining it as "a reunion of sorts."
PhysicsX worked privately with engineering and manufacturing organizations for about three years before announcing its Series A in November 2023. As of November 2026, Tuluie was Chairman and Corbo was CEO of PhysicsX. The company has since added former QuantumBlack COO Chris Wigley as COO in November 2025 and Dael Williamson, previously EMEA CTO at Databricks, as Chief Technology and Platform Officer in October 2026.
Product
PhysicsX provides software and models intended to predict how a physical design will perform without running a full numerical simulation each time. Engineers train models on simulation, experimental, and operational data, and the company says those models can then evaluate geometries and operating conditions in seconds. The company presents this as a way to explore "thousands or millions of possibilities rather than a handful of candidate designs."
Its platform connects to existing CAD and CAE tools, which still generate much of the data its models learn from. Corbo has explained the case for inference over simulation in three parts. He argues that AI inference lets engineers search larger design spaces than a numerical solver can practically cover, that models conditioned on experimental data can exceed the fidelity of first-principles simulation, and that high-fidelity prediction can reach compute-constrained settings such as manufacturing process control.
PhysicsX Platform
The PhysicsX platform has five layers, which are data foundation, AI development, engineering workflows, agentic intelligence, and enterprise deployment. In April 2026, the company described its platform in terms of a Simulation Workbench, an AI Workbench, and Engineering Applications.

Source: PhysicsX
Data Foundation: The data foundation ingests a customer's existing data and orchestrates CAE jobs across cloud and high-performance computing, bringing simulation, experimental, and operational data together for model training. Its capabilities include simulation orchestration, data transformation, lineage, labeling, and integrations with engineering tools.

Source: PhysicsX
AI Development: The AI development layer is where teams train, fine-tune, validate, and deploy models. It offers more than 12 architectures, third-party libraries, active learning, uncertainty quantification, and model evaluation. A 2026 company post says that environment includes access to NVIDIA PhysicsNeMo and NVIDIA Apollo alongside PhysicsX's own Opora modeling framework.

Source: PhysicsX
Engineering Workflows: The workflow layer embeds models in design exploration, inverse design, optimization, manufacturing, process control, and virtual sensing. Engineers reach those models through interfaces, APIs, or Model Context Protocol (MCP) connections to the tools they already use. PhysicsX groups the resulting uses into design and development, manufacturing, and operations.

Source: PhysicsX
Agentic Intelligence: PhysicsX's engineering agents are designed to set up experiments, run models, investigate failures, interpret results, and decide on next steps, with evaluation gates and human review. In one company account, an agent working on an industry benchmark reverse-engineered the closed-source solver behind part of the test from its data. PhysicsX suggests that episode points toward large-scale "distillation of engineering simulation software" as agents become more capable.

Source: PhysicsX
Enterprise Deployment: The platform runs in PhysicsX-hosted environments, a customer's own cloud or virtual private cloud, on-premises, or fully air-gapped, and it connects to product lifecycle management and manufacturing execution systems. PhysicsX went live on the NVIDIA-powered Industrial AI Cloud with Deutsche Telekom in February 2026.
Large Physics Models
PhysicsX's Large Physics Models are pre-trained models intended to generalize across engineering domains, so a customer can adapt one with less of its own data. The company says they require up to 80% less task-specific training data to reach comparable performance. In August 2026, it released PX-Car-Aero-v1, a pre-trained automotive aerodynamics model available on the platform.

Source: PhysicsX
PX-Car-Aero-v1 is trained on PXNetCar, a proprietary dataset that PhysicsX reported had grown to 33.2K simulations across 444 baseline vehicles as of August 2026. The company generates that data through its own Data Factory, which runs Siemens's Simcenter STAR-CCM+ solver. PhysicsX said it built PXNetCar after finding that public benchmarks such as DrivAerNet++ covered too narrow a range of designs to test how models generalize.

Source: PhysicsX
PhysicsX reported that fine-tuning PX-Car-Aero-v1 on one vehicle study needed about 80% less simulation data than training the same architecture from scratch, and 30-70% less on the third-party DrivAerML dataset. The savings varied by metric and disappeared for surface pressure at the tightest accuracy targets.

Source: PhysicsX
PhysicsX's first public demonstration of this approach was Ai.rplane, a free public demonstrator launched in December 2024 and since retired. It ran on LGM-Aero, which PhysicsX billed at launch as the first Large Geometry Model for aerospace. Users could explore aircraft geometries and see predicted performance without running their own simulations.

Source: PhysicsX
LGM-Aero was trained on more than 25 million meshes from commercially available datasets, plus tens of thousands of simulations generated with Siemens tools. PhysicsX reported that the model, used without fine-tuning on a component from a European automaker, produced a 7% increase in aerodynamic performance and a 10% reduction in the component's mass compared with the existing optimized design. Siemens, which worked with PhysicsX on the demonstrator, said users could fine-tune such models for their own applications with a few tens of simulations.
Forward Deployed Engineering
PhysicsX's Forward Deployed Engineering team works alongside customers to apply the platform to specific engineering problems. The company says each deployment leaves behind "reusable models, data, infrastructure, and engineering intelligence." A 2025 company post put its simulation engineering team at 35 people, many with backgrounds in aerospace, automotive, energy, and semiconductors.
Market
Customer
PhysicsX's customers are large industrial organizations whose products depend on complex physics. As of November 2026, it served five sectors: aerospace and defense, automotive and heavy machinery, semiconductors and electronics, process industries, and energy and infrastructure. In June 2026, Corbo expected semiconductors to become its largest segment by the end of the second quarter. In October 2026, he said aerospace and defense were its largest sectors, followed by semiconductors.
As of June 2026, customers included Applied Materials, Siemens, and Stellantis. Applied Materials and Siemens are also investors. PhysicsX has also worked with Leonardo Aerospace, Audi, and Rio Tinto. In October 2026, Fortescue, an Australian metals and energy company, expanded a relationship of more than two years into a multi-year agreement covering its mining and metals operations as well as R&D.

Source: PhysicsX
GB1, Britain's America's Cup team, is one of the few customers with published results. PhysicsX and GB1 reported in October 2026 that early results from their hull-design workflow included a 5x reduction in high-fidelity simulation, 100x greater accuracy than GB1's previous panel-method simulation, and 100x faster evaluation overall.
Expert-network interviews show a wide range of adoption. In an interview published in February 2026, one PhysicsX customer described choosing the company after a formal request for proposals that included incumbent simulation vendors, followed by a phased pilot. That customer reported a 93% reduction in the time each design iteration took and said it could test 36 structural variants a week instead of five. In another interview published that month, a second customer said it kept PhysicsX to a small, paid pilot for the problems conventional CAE handles worst and had not replaced its existing simulation vendors.
PhysicsX has expanded across sectors and geographies. In terms of sectors, Corbo has described starting in discrete manufacturing, including aerospace, defense, and automotive, before moving into energy production and then process manufacturing such as semiconductors and metals. With respect to geographies, the company started in the UK and Europe, followed tier-one customers into North America, and opened an Asia-Pacific hub in Singapore in October 2026.
Market Size
The total CAE market was estimated at $11.6 billion in 2025, comprising $10.1 billion in software and $1.6 billion in provider services. Aerospace is forecast by one estimate to remain the fastest-growing manufacturing segment, at an 11.2% CAGR from 2025 to 2029. Another estimate puts the market at $12.9 billion in 2025 and projects $32.9 billion by 2033, a 12.8% CAGR from 2026 to 2033.
PhysicsX's near-term opportunity is smaller than the whole CAE market. In an interview published in February 2026, a customer who piloted the product estimated that end-to-end AI simulation tools could capture at most 20% of a carmaker's CAD and CAE spend, and more realistically 5-10% of that spend, because of switching costs and incumbents' AI products. AI could also shrink parts of the core market, since Cambashi warns that customers often expect continuous feature upgrades within their packages and that workforce cuts could reduce license counts.
Competition
Competitive Landscape
PhysicsX competes with three groups. Incumbent CAE vendors are adding AI to products engineers already use. Synopsys completed its acquisition of Ansys in July 2025, and Siemens completed its acquisition of Altair in March 2025, giving both companies simulation software and AI tools to sell through established customer relationships.
Specialist physics AI startups are building models and platforms from scratch. More than 120 startups and small companies had attracted about $2.3 billion as of 2026, and Vinci alone raised $250 million in October 2026. A third group of well-funded AI companies is approaching engineering through general-purpose models or agents, with Prometheus and Mistral AI leading the funding.
PhysicsX differs from most competitors by combining pre-trained models with forward-deployed engineers across five industrial sectors. In an interview published in April 2026, an employee at a PhysicsX competitor said PhysicsX could reach usability in 30-60 days, compared with about six months for some alternatives. Luminary also offers pre-trained models, and Vinci claims to need no customer-specific training, so PhysicsX's models alone are unlikely to separate it. Its embedded teams also make the business more labor-intensive than software customers can adopt alone.
The category remains divided over how far AI can go. Neural Concept's Thomas von Tschammer, for example, told IEEE Spectrum in April 2026 that "we will never fully replace simulations," while Corbo said, "the whole idea is to take numerical simulation … out of the workflow." General Motors said accuracy matters most near launch, when a physical wind-tunnel test supplies the figure used for fuel-economy certification.
Physics AI Startups
Vinci: Vinci, founded in 2023, provides an AI-native physics platform for hardware engineers, starting with thermal and fluid behavior in semiconductors. In October 2026, it raised a $250 million Series B at a $1.5 billion valuation, co-led by Advent, Temasek, and Xora, and it had raised a total of $296 million as of November 2026. Temasek had led PhysicsX's Series C four months earlier. Vinci claims its product works on new designs without customer-specific training, which would remove the fine-tuning step PhysicsX is trying to shrink.
Rescale: Rescale, founded in 2011, provides a cloud platform for simulation computing and data and has added tools that turn customers' simulation data into surrogate models. It raised a $115 million Series D in April 2025 from investors including Applied Ventures and NVIDIA, had raised a total of $272.8 million as of November 2026, and has not disclosed its valuation. Applied Materials' venture arm and NVIDIA have backed both Rescale and PhysicsX, so those investor relationships do not give PhysicsX exclusive access to them.
Luminary: Luminary, founded in 2019, offers a physics AI platform with pre-trained SHIFT models. It raised a $72 million Series B led by N47 in September 2025, with participation from NVIDIA's NVentures. As of November 2026, it had raised a total of $187 million, including debt, and had not disclosed its valuation. Luminary's February 2026 Private Cloud product brings it into the secure defense and aerospace deployments that PhysicsX also targets.
Neural Concept: Neural Concept, founded in 2018 as an EPFL spinout, provides an engineering AI platform that connects to CAD and simulation tools. It raised a $100 million Series C led by Growth Equity at Goldman Sachs Alternatives in December 2025, had raised a total of $136.8 million as of November 2026, and has not disclosed its valuation. More than 50 companies were using its platform when it announced that round, and Jaguar Land Rover presented its aerodynamic work with Neural Concept at NVIDIA's GTC conference in March 2026. Its emphasis on speeding up early design gives buyers an alternative to PhysicsX's push to move simulation out of the workflow.
Flexcompute: Flexcompute, founded in 2015, builds GPU-native solvers and a physics model that combines neural networks with first-principles physics. It received a strategic investment of undisclosed size from Samsung Ventures and Applied Ventures in October 2026, had raised a total of $81.9 million as of November 2026, and has not disclosed its valuation. Because Flexcompute owns the solvers that generate its training data, it does not depend on a third-party simulator the way PhysicsX's Data Factory does.
BeyondMath: Founded in 2022 in London, BeyondMath is building a foundational physics model for engineering simulation. It raised a $10 million seed extension led by Cambridge Innovation Capital in February 2026, had raised a total of $20.3 million as of November 2026, and has not disclosed its valuation. It is pursuing a similar foundation-model approach from the same city with far less capital, so it could become a bigger threat if physics models prove cheap to replicate.
Monolith AI: Founded in 2016 in the UK, Monolith AI built an engineering AI platform that learned from test and simulation data, with customers including BMW and Siemens. It had raised £10.6 million prior to its acquisition. CoreWeave announced the acquisition in October 2025. PhysicsX partnered with CoreWeave in March 2026 and said it would train its pre-trained models there, so one of its compute partners owns a competing physics AI product.
ToffeeX: ToffeeX, founded in 2019 as a spinout from Imperial College London, builds physics-driven generative design software for thermal and fluid components. It raised a £5 million Series A co-led by Presidio Ventures and East Innovate in November 2022, had raised a total of £6 million as of November 2026, and has not disclosed its valuation. Its focus on one family of physics makes it a partial competitor to PhysicsX's broader platform.
Incumbent CAE Vendors
Synopsys: Ansys SimAI is a software-as-a-service application launched in January 2024, before Synopsys acquired Ansys, that trains models on customers' existing Ansys or non-Ansys simulation data. Ansys said it could speed model-performance prediction by 10-100x on computation-heavy projects. Its 2026 R1 release added a desktop Pro offering alongside the original product. Because customers can use data they already have, SimAI lowers the setup burden that PhysicsX addresses with its own pre-trained models.
Siemens: Siemens's Simcenter PhysicsAI lets engineers train and validate surrogate models inside STAR-CCM+, with high-fidelity simulation retained as the reference. Siemens gained the underlying PhysicsAI technology through Altair and, in August 2026, extended it across more products. Siemens is an investor in PhysicsX, as well as being a named customer of the company, the supplier of the solver PhysicsX's Data Factory runs, and a competitor.
Cadence: Cadence competes through its simulation portfolio, which expanded in February 2026 when it bought Hexagon's Design and Engineering business, including MSC Nastran and Adams, for about €2.7 billion. System Design and Analysis was 16% of Cadence's revenue in Q3 2026. Cadence has described physics simulation as a source of data to train and validate physical AI, which would make it both a supplier to and a rival of AI-native engineering platforms.
Dassault Systèmes: Dassault's February 2026 partnership with NVIDIA targets engineering design through its SIMULIA software, aiming for near-real-time performance predictions for designers. Dassault also controls the CAD and product lifecycle software where engineering data originates, so PhysicsX must integrate with a vendor that is building overlapping features.
General Engineering AI
Prometheus: Prometheus, formally established in November 2025 and co-led by Jeff Bezos and Vik Bajaj, is building what it calls an artificial general engineer. It announced a $12 billion Series B at a $41 billion valuation in June 2026, three days after PhysicsX announced its Series C, and had raised a total of $18.2 billion as of November 2026. That total was more than 37x PhysicsX's $487 million. Bezos said Prometheus creates much of its data itself and may buy parts of companies to improve their manufacturing, an approach that could erode the dataset advantage PhysicsX is building.
Mistral AI: Mistral AI, founded in 2023, entered physics AI when it acquired Emmi AI in May 2026. Emmi had spun out of NXAI's simulation group. Mistral raised a €3 billion Series D at a valuation of more than €21 billion in September 2026 and had raised $7.5 billion in total as of November 2026. The deal gives a European model developer a physics AI team, which could put it in contention for the industrial work PhysicsX is pursuing in Europe.
P-1 AI: Founded in 2024, P-1 AI is building Archie, an agentic AI engineer for mechanical, electrical, thermal, and systems work. It raised a $50 million Series A in July 2026, led by NEA, has raised $73 million total as of November 2026, and has not disclosed its valuation. Working above solvers and surrogates, it could use models like PhysicsX's or become the interface that makes them interchangeable.
Business Model
PhysicsX does not publish pricing or contract terms. Its customer relationships begin with jointly scoped pilots that typically last one to three months, with PhysicsX and customer engineers working side by side. Larger customer relationships can become multi-year agreements, as with Fortescue. The company's enterprise deployment options and embedded teams point toward custom-quoted contracts rather than self-serve software.
PhysicsX’s growth potential depends on the expansion of its delivery capacity. In June 2026, Corbo cited a backlog of customer demand of about six months and called it "a little untenable," and he described the company as "very supply-side limited." In an interview published in February 2026, one customer said that bringing on a new use case took almost six weeks of data curation and retraining.
PhysicsX's main costs are personnel, GPU compute, and simulation data. Its team doubled to more than 300 people in the 12 months to June 2026. It planned to train models on CoreWeave and Deutsche Telekom infrastructure, and its Data Factory runs hundreds of CFD simulations per day. The company presents pre-trained models as a way to reduce the cost and time of adapting physics AI to new problems. Deutsche Telekom also provides a European channel, with T-Systems and PhysicsX teams onboarding customers together.
Traction
PhysicsX's revenue has shown rapid growth. Corbo said in June 2026 that revenue would be close to $50 million for the year and that the company aimed to more than double it in 2027. The company reported more than quadrupling its revenue over the two years before its June 2025 Series B. By its June 2026 Series C, it reported that recognized revenue had doubled year over year, booked revenue had tripled, and its customer count had more than doubled.
PhysicsX expanded its team significantly in 2024. PhysicsX's 2024 group accounts reported a pre-tax loss of £8.2 million, up from £4.3 million in 2023. Average headcount across each year rose from 28 to 74. PhysicsX reported headcount of more than 50 at the November 2023 Series A, more than 150 at the June 2025 Series B, and more than 300 by June 2026. As of November 2026, 353 people were associated with its LinkedIn page.
Customer evidence ranges from broad deployment to limited pilots. In a December 2025 interview, one customer described adoption as cutting vehicle development cycles from four to six years to as little as three, reducing physical prototypes by more than 40%, and lowering development costs by 10-20%. Another customer said in a February 2026 interview that it kept PhysicsX at pilot scale for about one and a half to two years.
PhysicsX added several partnerships during 2026. In March 2026, it partnered with GB1 and collaborated with Siemens Smart Infrastructure on physics AI for data center power distribution. In October, it opened its Singapore hub, which it said had 20 employees and a goal of more than 100, alongside a five-year partnership with Singapore's A*STAR research agency, under which the two organizations will commit S$40 million.
Valuation
PhysicsX announced an oversubscribed $300 million Series C in June 2026 at a valuation of about $2.4 billion, led by Temasek with new investors M&G Investments and Intrepid Growth Partners. The British Business Bank later invested $25 million as part of the round.
Before the Series C, PhysicsX raised a $135 million Series B led by Atomico in June 2025 and extended it in November 2025 with a $20 million investment from NVIDIA's NVentures at a valuation of nearly $1 billion. Its $32 million Series A, led by General Catalyst in November 2023, was raised at an undisclosed valuation. As of November 2026, PhysicsX had raised a total of $487 million, and its valuation had more than doubled in the less than seven months between the Series B extension and the Series C.
At about $2.4 billion, PhysicsX was valued at about 48x the $50 million of revenue Corbo expected for 2026. As of November 2026, Cadence traded at a revenue multiple of 13.9x, Synopsys at 9.6x, Dassault Systèmes at 5x, and Siemens, a broader industrial group, at 2.9x. Cadence and Synopsys generated trailing revenue of $6.1 billion and $9.4 billion, respectively, so PhysicsX's premium depends on sustaining growth well above theirs from a far smaller base.
Two 2024 acquisitions show what strategic buyers were willing to pay for CAE software companies. Cadence agreed in March 2024 to buy BETA CAE at about 14x its annual revenue of about $90 million. Synopsys agreed in January 2024 to buy Ansys at an enterprise value of about $35 billion, or about 15x its FY2023 revenue of $2.3 billion. PhysicsX's multiple at the time of its Series C was more than three times either figure, even on revenue expected for the year of the round.
Key Opportunities
Pre-Trained Models Across Industrial Verticals
PhysicsX has applied its pre-training approach most notably in automotive aerodynamics, where PXNetCar supports PX-Car-Aero-v1. Its October 2026 GB1 work pairs Large Geometry Models that generate hull concepts with Large Physics Models that predict their performance. Early results suggest this could reduce the retraining one customer identified as the product's main limitation. In a company case study, an automaker's brake-cooling project needed 50 CFD runs on its first vehicle platform and 25 on its second, while PhysicsX estimated a traditional surrogate workflow might need more than 200. PhysicsX is working to repeat this approach in other sectors, each of which requires its own data.
Data Center Engineering
AI infrastructure gives PhysicsX an engineering problem outside its original markets that it can expand to service. In March 2026, PhysicsX and Siemens Smart Infrastructure developed a model for data center busway systems that they said turned multi-day analyses into seconds. Busways carry power to server racks, and power density and cooling requirements shape how that equipment is designed. This work gives PhysicsX a route into data center spending through equipment suppliers.
Sovereign Industrial AI Demand
PhysicsX can deploy on-premises or fully air-gapped, and its partnership with Deutsche Telekom places it on Europe's Industrial AI Cloud. Corbo said in June 2025 that "geopolitical currents and the questions of sovereignty and supply chain resilience are most strongly felt" in industrial manufacturing. Its Singapore hub and five-year A*STAR partnership give it a local engineering and research base in Asia. These capabilities could help it win defense and critical-industry work that requires local control of data, although Mistral AI and Luminary are pursuing similar customers.
Key Risks
Incumbent AI Inside Existing Tools
Ansys SimAI and Siemens's Simcenter PhysicsAI bring surrogate modeling into software engineers already use, and a feature inside an existing contract is harder to displace. Siemens presents a particular exposure. PhysicsX's Data Factory runs Siemens's Simcenter STAR-CCM+ to generate training data, while Siemens is an investor and, as of November 2026, offered its own surrogate tools in that solver. PhysicsX therefore depends on a competitor's software for part of its data pipeline, and it remains exposed to that vendor's licensing and roadmap even though it reports that its models can adapt across simulator settings.
Model Reliability in New Conditions
PhysicsX's pitch is strongest if its models remain accurate outside their training data, which is where engineers most want help. The company acknowledges that accuracy degrades across all metrics on new designs. In AeroZoo, a November 2026 aerodynamics benchmark that did not test PhysicsX's models, held-out geometry families barely changed the error of one tested model, but its error ratio rose 2.4x on unfamiliar enclosures and as much as 9.2x on unfamiliar flow layouts. University of Michigan researcher Karthik Duraisamy has also argued that many successful benchmarks sit on low-dimensional problems where any competent reduced model will interpolate well.
Regulated engineering raises the bar further. General Motors uses a physical wind-tunnel test for the figure that feeds fuel-economy certification, and Cambashi expects regulatory compliance to govern the pace of adoption in sectors including aerospace and automotive. Solvers and their vendors may therefore stay in the workflow at the stage where errors cost the most.
Services-Heavy Delivery
PhysicsX's growth depends on engineers working with customers, and its gross margin is undisclosed. Its team doubled in the year to June 2026 while its recognized revenue also doubled, which implies little change in revenue per employee, and Corbo cited a six-month customer backlog. In an interview published in February 2026, one customer said that the tool did not scale well across use cases because each new problem required substantial retraining. The customer rated it seven out of 10 overall but one out of 10 against any expectation of replacing incumbent CAE tools. If delivery effort does not fall per customer, a valuation of about 48x expected 2026 revenue would rest on a business with services-like economics. Vinci's claim that its product needs no customer-specific training is aimed at that weakness.
Weekly Newsletter
Subscribe to the Research Rundown
Summary
PhysicsX provides pre-trained physics AI models and a platform that moves engineering work from numerical simulation toward AI inference across design, manufacturing, and operations. It had raised a total of $487 million as of November 2026, including a $300 million Series C at a valuation of about $2.4 billion, and its CEO expected revenue close to $50 million in 2026. Its investors include Siemens, Applied Materials, and NVIDIA, and Siemens and Applied Materials are also customers.
PhysicsX’s opportunity is to repeat its pre-trained model approach across industrial verticals while lowering the data and engineering work each customer requires. The central questions are whether those models remain reliable on new designs, whether delivery becomes less labor-intensive as PhysicsX grows, and whether it can stay ahead of incumbents such as Siemens and well-funded entrants such as Vinci and Prometheus.



