Luminary Cloud: The Physics AI Model Factory Backed by NVentures

Executive Summary

Luminary Cloud began in 2019 as a cloud CAE platform: rent NVIDIA GPUs, run traditional physics solvers on them, and charge by the minute. By 2026 it had repositioned as what it calls the Physics AI Model Factory — a platform that generates simulation data, trains neural surrogates on it, and sells the resulting pretrained models under the SHIFT brand. The Arcadia of the company is unusual: it is simultaneously a CAE cloud, a data generator, a model developer, and a defense contractor.

Four facts frame the investment case.

First, the founding thesis was unusually well-timed and unusually well-backed. Luminary emerged from stealth on March 13, 2024 with $115 million led by Sutter Hill Ventures, one of the largest seed-to-Series-A raises in US enterprise software that year. At the time the pitch was straightforward: CAE was the last major application category to stay on desktop workstations, and GPU-native cloud simulation running 100 times faster than legacy would move it. NVIDIA's chief scientist Bill Dally endorsed it publicly in the launch release.

Second, the company pivoted from selling compute to selling models, and the pivot is what makes it a Physics AI company rather than a cloud business. The Physics AI Factory unifies GPU-native simulation, large-scale data generation, model training, and deployment into one pipeline. The SHIFT family now spans SHIFT-SUV (automotive, with Honda), SHIFT-Wing (aerospace, with Otto Aerospace), SHIFT-Crash (full-vehicle crash, announced April 2026), and three defense models announced January 12, 2026: SHIFT-CCA, SHIFT-Submarine and SHIFT-Pump.

Third, the accuracy position was surrendered in March 2026. Luminary announced on March 16, 2026 that it was the first Physics AI platform to adopt NVIDIA's new GeoTransolver architecture, retraining its entire SHIFT portfolio on it, claiming up to 10x greater accuracy and a 45% to 138% speedup on SHIFT-Wing. Reading the underlying NVIDIA benchmark makes the concession explicit: on SHIFT-SUV Estate volumetric velocity, GeoTransolver achieves 1.36% relative L1 error against AB-UPT's 2.25%; on SHIFT-Wing volumetric velocity, GeoTransolver reaches 1.92% and 2.00% at Mach 0.5 and 0.85 against AB-UPT's 9.56% and 9.51%. Luminary's own published SHIFT-SUV drag error was around 3.45% mean. The company won by adopting NVIDIA's architecture, not by out-inventing it.

Fourth, the business model and the customer base are real but thin, and revenue has never been disclosed. Luminary has been unusually good at publishing named customers: Northrop Grumman, Otto Aviation, Joby Aviation, Piper Aircraft, Trek Bikes, Cobra Golf, Sceye, Natilus, Rune Aero, Mueller, Welbilt, Honda, Northrop Grumman Space Systems. It has been unusually bad at publishing numbers: no revenue, no ARR, no contract values. LeadIQ's only point estimate is $50 million to $100 million, which is a database estimate and should be treated as such.

Our view: this is the best-capitalized and most distribution-ready independent in the Physics AI category, and it has already shown that its technical edge was rented rather than owned. The company is worth studying not because it will win the neural surrogate race but because it shows what a real business looks like when the physics AI product is a model portfolio plus a data pipeline plus a solver cloud, rather than a research demo. The risk is that NVIDIA is simultaneously its supplier, its architecture provider, and, through NVentures, its shareholder.

Company Overview

Luminary Cloud, Inc. is a Delaware corporation founded in 2019, headquartered at 101 South Ellsworth Avenue, Suite 250, San Mateo, California 94401. It was formerly known as Silicon Box. It was incorporated in 2019 and made its first SEC Form D filing in December 2020, with onboarding of customers beginning in 2023.

The founding team is unusually well matched to the thesis:

• Jason Lango, Co-founder and CEO until 2024. Previously founded and served as CTO of Bracket Computing, which raised $130 million and was acquired by VMware in 2018. Previously a venture partner at Primary Venture Partners. At the 2024 launch he described the origin: "While software engineering has become more agile thanks to advances in cloud technologies, physical engineering hasn't kept pace."
• Juan Alonso, CTO and Co-founder. Chair of the Department of Aeronautics and Astronautics at Stanford University and former director of NASA's Fundamental Aeronautics Program. This is the single most valuable asset in the company: a Stanford aero department chair who ran a NASA research program and now builds a commercial CAE company.
• Pete Schlampp, CEO from 2024, previously Managing Director at Sutter Hill Ventures, and before that chief strategy officer and EVP of product development at Workday, with earlier roles at Cisco, IronPort Systems and Solera Networks. He is an investor who became CEO, which is a distinctive governance pattern in this category.
• Sudhir "Suds" Menon, Chief Product Officer, 25+ years across Cloudera, Sema4.ai, SnappyData, VMware, Pivotal and TIBCO.
• Peter Lyu, Field CTO and VP of Solutions, PhD in Aerospace Engineering from the University of Michigan, previously led GTM teams in cloud HPC and CAE.
• Ethan Carr, Head of Sales, from Wiz, Apollo GraphQL and Salesforce. Edward Frye, VP of Security and Compliance, former CISO at Aryaka Networks and four-time CISO. Cong Ho, SVP of Finance, from AppOmni, Agari and RingCentral.

Jason Lango's exit from the CEO role in 2024 and Schlampp's arrival is a real observation worth recording. The founder who articulated the thesis left the top job within four months of the public launch, and the Sutter Hill managing director who led the $115 million took over. The board includes T.J. Rylander of N47, Victor Miller and Michael Speiser of Sutter Hill, and Juan Alonso as a director.

Headcount is measured three different ways: 84 (Crustdata, September 2026, up 64.7% year on year), 86 observed professional profiles (Unify, July 2026, up 28.4% since August 2025), and 73 (PitchBook, 2024) with LeadIQ giving approximately 77 as of August 2026. Functionally, 32 percent of classified profiles are in Information Technology, 22 percent in Engineering, 12 percent in Sales.

Technology

What actually changed: from cloud solvers to learned surrogates

The 2024 product was a GPU-native CAE platform. Its own claim was running high-fidelity simulations 100 times faster than legacy vendors by running NVIDIA GPU clusters in the cloud, with elasticity so a customer could spin up hundreds of GPUs for one job and scale them down after. CTO Juan Alonso described the specific claim: a 100-million-cell RANS simulation that took four to six hours on a cluster could run in a couple of minutes. The platform was built on Google Cloud, with stated intent to expand to other providers, and included Lumi AI, an engineering design copilot that automates mesh generation.

The pivot to Physics AI is not a pivot away from simulation. It is a claim that generating the training data is the real product. The Physics AI Factory is described as unifying GPU-native physics simulation, large-scale data generation, AI model training, and model deployment and inference into a single pipeline. Luminary's own framing: "Luminary has solved the biggest problem of generating the necessary synthetic data to build and train accurate Physics AI models."

The defensible part of the business is therefore a data generation loop, not a neural architecture. No Luminary-authored peer-reviewed paper establishing a new surrogate architecture was found. Its models are built on NVIDIA PhysicsNeMo and, since March 2026, on NVIDIA's GeoTransolver.

SHIFT-SUV: the most transparent dataset in the category

SHIFT-SUV is the best-documented artifact, and its transparency is unusual for this category. The geometries are parameterised variations of AeroSUV, the open-access SUV reference model from FKFS Stuttgart, whose open licence is non-commercial only. The dataset was curated with Honda's support and covers discrete variations (chassis type, cooling system, rear section) and continuous ones (windscreen angle, nose radius, rear taper).

The scale and cost of the data generation loop is where the real moat sits: over 1,000 simulations were run on the platform, each taking approximately 1.2 hours on an NVIDIA H100. The training used NVIDIA's open-source PhysicsNeMo framework. The company was publicly targeting 25,000 variations by end of 2025.

The accuracy results, and the generalisation caveat, are worth reading in full:

• Test set: approximately 3.45% mean drag prediction error, maximum around 6.5%. 75 percent of predictions fall within 5 percent error and every prediction within 10 percent.
• Validation against high-fidelity CFD in Blender showed high reliability on fastback-type vehicles with performance degrading on out-of-distribution shapes such as station wagons. Luminary's own conclusion: SHIFT-SUV suits early-stage aerodynamic study.
• A specialised-versus-general trade-off was documented: training on only 500 fastback cases improved accuracy to 2.10% mean relative error with 94 percent of predictions under 5 percent, but using the full 1,000-case dataset improved generalisation to other SUV variants while slightly degrading drag accuracy on fastbacks.

That last trade-off is the most useful published data point in this entire category, because it is the concrete demonstration of the generalisation-versus-specialisation tension that every physics AI company faces.

The defense turn

On January 12, 2026, at AIAA SciTech in Orlando, Luminary released three defense models: SHIFT-CCA (Collaborative Combat Aircraft, a USAF program for AI-driven autonomous drones flying alongside crewed fighters), SHIFT-Submarine (hydrodynamic efficiency under internal hardware constraints such as battery capacity) and SHIFT-Pump (compressible rotating turbomachinery, optimizing pressure, flow and efficiency). Each was built from scratch on synthetic data generated by Luminary's GPU-native solvers, exercising high-speed external aerodynamics, fully-submerged incompressible hydrodynamics, and compressible rotating turbomachinery respectively.

The strategic framing is explicit: "If the DoW wants to outpace adversaries, physics AI is not optional — it is a mission-critical necessity." The models are positioned as delivering "secure, U.S. controlled Physics AI capabilities" for the Department of War across air, maritime and ground domains, inside "an end-to-end, secure, U.S.-based platform purpose-built to design, train, and deploy Physics AI models inside trusted environments."

At AIAA SciTech in March 2026, CTO Juan Alonso said Luminary expected to make at least two of the models open-source on Hugging Face, explaining the reasoning: "When you generate 20,000 different variants of an aircraft, that takes a lot of time and effort and expertise, so many people who could contribute to improving the physics AI models are unable to do so because they cannot generate realistic data sets of high quality. We are putting them out there to help push the field forward." That is a genuine strategic commitment to open data, and it is also a hedge against the possibility that a platform vendor captures the value.

SHIFT-Crash and the transfer learning claim

Announced April 14, 2026, SHIFT-Crash predicts time-dependent fields for a full-vehicle frontal crash in seconds, trained on 5,000 crash simulations based on a 2010 Toyota Yaris reference vehicle. Traditional FEM takes 10 to 12 hours per crash test and must restart from scratch for each vehicle programme.

The compounding claim is the interesting part: using transfer learning, the number of new simulations required for a subsequent SUV programme could drop to fewer than 300 by the third iteration, and the first sedan programme would need roughly 600 rather than 5,000. The stated accuracy is less than 3 percent RMSE on workflow-ready predictions.

The regulatory context matters and is favourable: in February 2026, the BMW Group became the first carmaker to have virtual crash simulations officially recognized as equivalent to physical tests in Germany. Luminary positions SHIFT-Crash inside that workflow, allowing screening against NHTSA NCAP, Euro NCAP 2026 and IIHS protocols in seconds rather than days.

Deployment and the architecture dependency

On February 24, 2026, Luminary launched Luminary Private Cloud, an on-premises deployment for secure, air-gapped defence and aerospace environments. It delivers the same capabilities as the cloud-hosted service while operating entirely within restricted environments, building on existing ITAR-compliant deployment options and SOC 2 Type 2 and ISO/IEC 27001:2022 certifications. Hybrid deployment between private and public infrastructure is described as a future possibility.

The published commercial terms describe prepaid or on-demand credits for compute and storage, following the Snowflake consumption model, with discounts for reserved capacity, plus commercial SHIFT model licensing. In 2024, Lango said pricing had not yet been specified but that the company followed "the Snowflake model with published pricing based on usage."

Luminary's own framing of the accuracy problem is worth quoting because it is more honest than most: "The hardest part of physics AI is not building the models. It is getting them to travel the last mile inside a real engineering organization." The company has also documented its own failure mode — a simplified propulsion assumption in an early combat-aircraft model after poorly constrained boundary conditions made the training data noisy. The response was to narrow the model's scope, not to pretend the physics had become easy.

Performance Benchmarks

The most informative comparison available is NVIDIA's own GeoTransolver paper, which benchmarks against DoMINO, Transolver and the literature-reported AB-UPT on DrivAerML, SHIFT-SUV and SHIFT-Wing.

On SHIFT-SUV, Estate geometry volumetric velocity, relative L1 error: GeoTransolver 1.36%, Transolver 1.87%, AB-UPT 2.25%. On SHIFT-Wing volumetric velocity, GeoTransolver reaches 1.92% at Mach 0.5 and 2.00% at Mach 0.85, against AB-UPT at 9.56% and 9.51% and Transolver at 3.60% and 3.72%. On drag and lift R-squared, GeoTransolver takes the best or comparable results, with CD R-squared of 0.98 on SHIFT-SUV Estate and 0.99 on Fastback.

The published task leaderboard for the GeoTransolver paper records: aerodynamic field prediction on SuperWing with relative L2 error 0.028; fluid flow field prediction on HiLiftAeroML with relative L2 error 0.0289; industrial-scale crash modelling on bumper beam at T=50 with relative L2 error 7.32 and on car crash at T=25 with relative L2 error 1.33; aerodynamic coefficient prediction on SHIFT-Wing drag at Ma=0.5; SHIFT-SUV Estate R-squared 0.98 and Fastback 0.99 for CD; SHIFT-SUV Estate pressure deviation 0.0057; and full-vehicle crash prediction on CarCrashNet with RMSE 21.773 mm for the Toyota Yaris and 79.23 mm for an unseen Chevrolet Silverado.

The read-through for Luminary is specific: the SHIFT-SUV drag error of roughly 3.45 percent sits in the same range as the field-level errors that the newer architectures reach below 2 percent on. Luminary's response was to adopt the new architecture wholesale rather than to defend the old one, and to rebrand the result as "near-simulation-level precision at AI speed."

Other published figures:

• SHIFT-Wing speedup of 45% to 138% from the newest PhysicsNeMo release in Luminary's training and inference pipelines.
• SHIFT-Crash less than 3 percent RMSE.
• A separate report cites an average drag prediction error of about 3.45% on Luminary's published SHIFT-SUV test set, with the explicit caveat that such figures describe a particular dataset and test, "not every SUV a designer might invent."

Product and Commercial Model

The product stack has four layers:

1. Simulation cloud. GPU-native solvers on rented NVIDIA capacity, elastic to hundreds of GPUs, with Python notebooks, published APIs and a software development kit.
2. Physics AI Factory. The data-to-deployment pipeline: generate high-fidelity synthetic data, train, validate, deploy, govern.
3. SHIFT model family. Pretrained models and, in some cases, the datasets themselves. Automotive (SUV), aerospace (transonic wing), crash, and the three defense models.
4. Developer surface. A Model Context Protocol Server, an AI-assisted notebook for rapid prototyping and post-processing, and a secure REST inference API.

The commercial model is consumption-based, following Snowflake, with prepaid or on-demand credits for compute and storage, discounts for reserved capacity, and separate commercial licensing for SHIFT models. As of 2026 the platform is marketed around three outcomes: rapid design iteration, design exploration across thousands of candidates in parallel, and optimization "across many design requirements with fast model predictions trained on differentiable solvers and adjoint sensitivity."

Named customers with public statements:

• Northrop Grumman Space Systems — Han Park, VP of Artificial Intelligence integration: "Physics AI is the next level of complexity in AI, and Northrop Grumman is bringing this technology to our design engineers to dramatically speed up hardware development." A separate partnership on rocket nozzle designs was signed in October 2025.
• Honda — Fong Loon Pan, Principal Aero Design Lead, on the SHIFT-SUV foundation model with Blender integration: it "provides immediate insights into vehicle performance that designers and engineers need."
• Otto Aviation — Scott Drennan, President and COO, on operating an aircraft in laminar flow using proprietary laminar algorithms on the Luminary platform.
• Rune Aero — Nadine Auda, Co-Founder, on a virtual wind tunnel reducing early development costs by over 80 percent against traditional wind tunnels.
• Joby Aviation — Gregor Veble Mikic, Chief Aerodynamicist: "With Luminary, we know which ideas to pursue and which ideas not to pursue very quickly. We don't spend months thinking about these problems."
• Sceye — Sidharth Arunkumar, design engineering manager, on stratospheric platforms; a named customer application in the GeoTransolver announcement.
• Natilus — Alexey Matyushev, CEO and Co-founder, on meshing being nearly turnkey.
• Piper Aircraft — Buddy Sessoms, Chief Product Officer: "The speed of simulation in Luminary compared to anything I've used previously is just bonkers."
• Trek — John Davis, Aerodynamics Lead: "With Luminary, we are moving twice as fast and at least half the cost."
• Mueller — Tim Stern, Mechanical Design Engineer, on transient solutions infeasible on current hardware; Mueller estimated avoiding $100,000 in physical prototyping costs and two months of predicted engineering time within two months of adopting the software.
• Welbilt, Flighthouse Engineering, M4 Engineering, Cobra Golf (a Puma subsidiary) also appear as customers.

One blind test is worth recording because it is the only instance of a customer-run accuracy check found in this entire category. Luminary principal solutions engineer Ian Lockley described an automotive customer asking for a blind transient aeronautics simulation: the customer's previous effort took 24 hours, Luminary ran it in less than an hour and came within 3 percent of the reference result.

Competitive Landscape

Luminary's competitive position is unusual because it is not primarily a neural surrogate vendor — it is a data and simulation infrastructure provider that also ships models.

The direct Physics AI competitors are PhysicsX (London, $300 million Series C in June 2026 at a $2.4 billion valuation, with Siemens' Simcenter data behind its LGM-Aero corpus), Neural Concept (EPFL, $138.1 million raised, a customer roster including GM, Safran, Airbus and four Formula 1 teams), Emmi AI (Linz, acquired by Mistral in May 2026 at up to EUR 330 million, whose AB-UPT became the accuracy benchmark Luminary's models were measured against), and Monolith AI (acquired by CoreWeave in 2025).

The deeper competition is the incumbents: Ansys inside Synopsys, Siemens Simcenter and BeyondMath, Dassault SIMULIA and 3DEXPERIENCE, Altair PhysicsAI. Note that Siemens is simultaneously a Luminary partner and a PhysicsX investor, and NVIDIA is simultaneously Luminary's supplier, its architecture provider, its benchmark author and, through NVentures, a shareholder.

The distinguishing structural point is the Physics AI Factory: the data generation loop sits upstream of the model. PhysicsX got there by owning the Simcenter relationship; Luminary got there by owning the solver cloud; Emmi got there by owning the research architecture and then being bought. A company that owns the solvers can generate a new dataset for any vertical in weeks. That is a genuinely different position from a company that owns one good model.

Funding and Team

The funding record has three distinct transactions, and the reporting on it is unusually muddled:

• December 2020, Series A, $15 million. SEC Form D records an offering amount of $15 million, amount sold $15 million, Rule 506(b), with Juan Alonso, Jason Lango and Michael Speiser listed as officers and directors. Sutter Hill Ventures is the sole investor. Forge lists a post-money valuation of $50 million at $0.94 per share.
• March 13, 2024, $115 million announced at stealth exit. The company announced it had launched from stealth with $115 million led by Sutter Hill Ventures. This was not a single round — Forge's cap table shows a Series A-1 of $54.41 million at $1.00 per share and a $176.03 million post-money, and Crunchbase, PitchBook and Form D data all record the 2024 event as a combination of equity and a debt facility. Crustdata explicitly labels the 2024 event as two entries: a $15 million venture round and a $100 million debt financing, both in March 2024, both led by Sutter Hill Ventures.
• September 15, 2025, Series B, $72 million. Announced by the company as $72 million led by N47 (formerly Next47), with Sutter Hill Ventures continuing and NVentures participating for the first time. N47's T.J. Rylander joined the board. Crustdata records the date as September 2025; Forge records a Series B of $80 million on June 27, 2025 at $0.69 per share and a $249.23 million post-money; PitchBook records the Series B on June 27, 2025 with $172 million raised to date; Fundraising Fox, working from SEC filings, reports the 2025 Form D showing an offering amount of $188.8 million and amount sold of $172 million, first sale June 2025, Rule 506(b).

The official company disclosure is $72 million for the Series B. The $80 million, $172 million and $188.8 million figures are different aggregators measuring different things — most likely the full Form D offering amount including a greenshoe, a portion of which remained unsold.

Valuation is not disclosed by the company. Forge's private-market data, which is a data vendor rather than a primary source, shows post-money valuations of $50 million (2020), $176.03 million (2024) and $249.23 million (2025) at $0.94, $1.00 and $0.69 per share respectively. On that sequence, the 2025 round was a down round on price per share while post-money rose. PitchBook lists the latest deal amount as $172 million. LeadIQ's $50 million to $100 million revenue estimate is a database estimate and should not be relied on.

NVIDIA's role is the one to watch. Bill Dally, NVIDIA's chief scientist, endorsed the company at the 2024 launch. Tim Costa, NVIDIA VP and GM of computational engineering, appears in the March 2026 GeoTransolver release. Luminary's platform integrates NVIDIA CUDA-X libraries including PhysicsNeMo, Transformer Engine and accelerated computing. And NVentures invested in the Series B. The dependency is therefore structural, not incidental — Luminary's models are built on NVIDIA's framework, benchmarked against NVIDIA's architecture, deployed on NVIDIA hardware, and partly owned by NVIDIA.

Discrepancies

Twelve conflicts were identified and adjudicated.

1. Total funding, $149.41 million versus $172 million versus $187 million versus $115 million plus $72 million. All four are defensible measurements of different things. Forge says $149.41 million (its cap table rounds and omits the debt facility). PitchBook says $172 million raised to date. Crustdata and LeadIQ say $187 million across three rounds, which includes a $100 million debt facility. Fundraising Fox, working from SEC Form D filings, reports $187 million disclosed across two of three rounds. Adjudication: the company-disclosed capital announced is $15 million in 2020 plus $115 million in 2024 plus $72 million in 2025 = $202 million, but the 2024 headline itself mixes $15 million of equity with a $100 million debt facility, so it cannot be read as equity; the SEC-filed amount sold totals $187 million; the $149.41 million figure is Forge's own cap table arithmetic and understates the record. No single number is "the" total.

2. Series B amount, $72 million versus $80 million. The company announced $72 million on September 15, 2025. Forge's $80 million on June 27, 2025 is a different data source measuring the filed offering. Adjudicated as $72 million officially disclosed, with the larger figure being a filed-offering measure, not a contradiction.

3. Series B date, June 27, 2025 versus September 15, 2025. Both trace to real filings — the Form D first sale was in June, the public announcement was in September. Adjudicated: first sale June 2025, announcement September 15, 2025.

4. Series A date, December 11, 2020 versus 20 December 2020. The same December 2020 Form D, with a one-week discrepancy in database transcription. Adjudicated as December 2020.

5. The 2024 $115 million, one round or two. The company announced a single "$115 million in funding." Form D and Crustdata show it decomposed into a $15 million venture round and a $100 million debt financing. Adjudicated: the $115 million headline includes debt, and is not a single equity round.

6. Founding date, 2019 versus 2019-10-23. All sources agree on 2019. Yuannian records 2019-10-23; the Delaware incorporation and first Form D are December 2020. Adjudicated as 2019, incorporated in Delaware.

7. Headcount, 73 versus 77 versus 84 versus 86. Four aggregators, four counts, all measuring different things — PitchBook is a 2024 figure, LeadIQ and Crustdata are LinkedIn-derived employee counts, Unify is observed professional profiles including contractors and non-employees. Adjudicated as roughly 80 to 85 people as of late 2026, growing 28 to 65 percent year on year depending on the measure.

8. CEO, Jason Lango or Pete Schlampp. Both are correct for different periods. Lango was co-founder and CEO at the March 2024 launch; Schlampp, previously Sutter Hill's managing director, is the current CEO. Adjudicated as a founder-to-investor CEO handover in 2024.

9. "Up to 10x greater accuracy." This is NVIDIA's GeoTransolver claim relative to prior architectures, adopted by Luminary, not a Luminary-specific measurement of a 10x improvement over its own previous models. The specific measurable gains Luminary published are a 45% to 138% speedup on SHIFT-Wing. The 10x figure is a category-level architecture claim, not a Luminary product delta.

10. "Less than 1% prediction error by SHIFT models." The homepage claim. The published SHIFT-SUV test-set figure is 3.45% mean drag error with a maximum around 6.5%, and NVIDIA's own leaderboard records SHIFT-SUV Estate pressure deviation at 0.0057 and CD R-squared 0.98. The sub-1% figure is not supported by the published SHIFT-SUV results and should not be repeated as a Luminary accuracy claim.

11. Whether Luminary or Emmi held the accuracy lead in late 2025. Luminary's SHIFT models were the reference datasets against which Emmi's AB-UPT was benchmarked, and Emmi's own technical report claimed top position on all five. NVIDIA's GeoTransolver paper then reported AB-UPT behind GeoTransolver and Transolver on SHIFT-SUV and SHIFT-Wing. The sequence: Luminary published the datasets, Emmi published the SOTA claim on them, NVIDIA published the architecture that beat both, and Luminary adopted it. No single company held a durable lead.

12. "Formerly known as Silicon Box." PitchBook lists this former name; it appears nowhere in Luminary's own materials. Low-confidence database metadata, retained as a lead rather than a fact.

Risks

The technical edge is rented from NVIDIA. Luminary's models are built on PhysicsNeMo, benchmarked against GeoTransolver, deployed on NVIDIA hardware, and partly owned by NVentures. When NVIDIA released a better architecture, Luminary's response was to adopt it within weeks and reposition. That is good execution and weak defensibility. The value captured is the data, the factory and the customer relationships, not the architecture.

Revenue has never been disclosed. No ARR, no contract values, no customer count, no net retention. The only point estimate is LeadIQ's $50 million to $100 million, which is a database estimate. The pricing model is consumption-based, which makes revenue lumpy and hard to forecast. The 2024 launch disclosed 33 prepaid customers by year end, and nothing comparable has been published since.

Defence revenue brings long sales cycles and programme risk. The company itself acknowledged that defence and aerospace sales cycles stay long and that a single stalled Northrop programme delays revenue. It is also exposed to procurement policy: the January 2026 announcement positions models for the "Department of War," a naming that itself signals reliance on the current US defence establishment's framing.

Generalisation beyond the training distribution is the unsolved technical problem, and Luminary documents it better than anyone. SHIFT-SUV degrades on out-of-distribution shapes such as station wagons. The specialised-versus-general trade-off is real and quantified. SHIFT-Crash's transfer learning promise across vehicle classes is a company claim not yet independently reproduced. And the company's own record includes a model whose scope had to be narrowed because poorly constrained boundary conditions made the training data noisy.

The GPU dependency cuts both ways. The entire thesis — that simulation becomes cheap because GPUs are abundant — depends on continued NVIDIA hardware economics and on rented capacity remaining available at the required scale. Luminary does not own its infrastructure; it rents it.

Platform risk in the AI builder market. A Physics AI Model Factory is a tool for building models. Foundation-model labs and the incumbents are both building the same tooling. If model construction becomes a commodity, the value concentrates in the data loop and the deployment, which is where Luminary is strong; if it concentrates in the model, Luminary is exposed.

Scoring

• Technology, 7.5. A genuinely complete engineering stack from solver to surrogate to deployment, the only dataset in the category published with the transparency of SHIFT-SUV, a documented and quantified specialisation-versus-generalisation trade-off, defensible deployments for air-gapped and ITAR environments, and the industry's most honest articulation of where the technology actually fails. Deducted for no Luminary-authored surrogate architecture, for the accuracy position having been surrendered to NVIDIA's architecture in March 2026, for the sub-1% claim being unsupported by published results, and for the crash transfer-learning promise remaining unreproduced.
• Commercialization, 7.0. More than fifteen named customers with direct executive quotes across aerospace, defence, automotive, sporting goods and industrial equipment, a blind-test accuracy check run by a customer, a documented consumption model, a private-cloud product for regulated deployments, and defence models opened to the community. Deducted for zero disclosed revenue, ARR, contract values or customer count, a consumption model that makes forecasting hard, long defence sales cycles, and dependence on a single GPU vendor across architecture, hardware and equity.
• Funding, 8.0. $187 million disclosed across SEC filings, an unusually large single-investor commitment from Sutter Hill that also supplied the CEO, a $72 million Series B led by N47 with the strategic addition of NVentures, and a private-market cap table showing steady post-money progression from $50 million to roughly $249 million. Deducted for the absence of any company-disclosed valuation, for the awkward $115 million headline that mixes equity with debt, and for the fact that the 2025 round appears to have been down on price per share.
• Product maturity, 7.5. A clear product ladder from simulation cloud through the Physics AI Factory to a family of eight named SHIFT models spanning automotive, aerospace, crash, and three defense domains, plus an MCP server, an AI-assisted notebook, a secure inference API, published datasets, and an open-source commitment for at least two defense models. Deducted for the pivot still being in progress, for the 2026 defence and crash models having no independent validation, and for the platform's dependence on components it does not own.
• Market position, 7.5. The best distribution and credibility position among the independent Physics AI companies, a named defence contract with Northrop Grumman, a data relationship with Honda, unique datasets that competitors benchmark against, ITAR and air-gapped capability in an era of sovereignty-driven procurement, and a brand that the largest GPU vendor publicly endorses. Deducted for a business whose defensible layer is a data loop rather than a model, for valuation that is not disclosed and cannot be tested against revenue, for the risk that the whole category consolidates into a handful of platform owners, and for the structural conflict in being supplier, benchmark author and shareholder at once.

References

1. Luminary Cloud, "Luminary Cloud Emerges from Stealth, Empowering R&D with Realtime Engineering" (2024-03-13)
2. Luminary Cloud, "Luminary Cloud Secures $72M Series B to Lead the Physics AI Era" (2025-09-15)
3. Luminary Cloud, "Physics AI Will Redefine Innovation in the Physical World" (2025)
4. Luminary Cloud, "Luminary Launches New Secure Advanced Physics AI Models to Power U.S. Department of War" (2026-01-12)
5. AIAA, "Advancing the Field: Luminary Cloud Announces New Physics AI Models at AIAA SciTech Forum" (2026-03-20)
6. Luminary Cloud, "Luminary Adopts NVIDIA GeoTransolver AI Model Architecture" (2026-03-16)
7. Luminary Cloud, "Luminary Launches Private Cloud" (2026-02-24)
8. Luminary Cloud / GlobeNewswire, "Luminary Launches SHIFT-Crash" (2026-04-14)
9. NVIDIA, "GeoTransolver: Learning Physics on Irregular Domains Using Multi-scale Geometry Aware Physics Attention Transformer", arXiv:2512.20399
10. SiliconANGLE, "Luminary Cloud launches with $115M funding for cloud CAE platform" (2024-03-13)
11. Crunchbase News, "The Week's 10 Biggest Funding Rounds" (2024-03)
12. SEC EDGAR Form D filings, Luminary Cloud Inc., December 2020 and June 2025
13. AIAA / Washington Technology coverage of the three defense Physics AI models
14. Crustdata, LeadIQ, PitchBook, Forge Global and Fundraising Fox company profiles, various access dates
15. Emmi AI / arXiv:2510.15808, AB-UPT for Automotive and Aerospace Applications, as the comparison point for SHIFT-SUV and SHIFT-Wing

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.