Vinci AI Series B

Vinci AI: Turning Physics Simulation from a Checkpoint into a Continuous Constraint

Vinci AI runs thermal-mechanical simulation directly at manufacturing resolution using a physics foundation model and GPU-native kernels. Its peer-reviewed EPTC 2025 paper reports under 2% deviation from commercial solvers at 240x acceleration. The company closed a $250M Series B at a $1.5B valuation in October 2026, with about $296M raised in total. The technical validation is solid; cross-domain expansion and named customers remain unproven.

Executive Summary

Vinci AI is an American physical AI company founded in late 2023 and headquartered in Palo Alto, with founder and CEO Hardik Kabaria, previously vice president of software at Carbon, alongside co-founders Sarah Osentoski as CTO with robotics and machine learning background and Vincent Rerolle as CCO. In December 2025 the company emerged from stealth disclosing 46 million dollars of seed and Series A funding and published a peer-reviewed advanced packaging thermal study at EPTC 2025, the Electronic Components and Technology Conference.

On October 6, 2026 the company announced a 250 million dollar Series B at a 1.5 billion dollar valuation, co-led by Advent, Temasek and Xora Innovation with AMD Ventures, Madrona, Eclipse and Khosla Ventures participating, bringing disclosed funding to about 296 million dollars.

The technical thesis is Continuous Physics Reasoning: moving simulation from a checkpoint at the end of design into a continuous constraint within the design process. The stack has four parts, namely automated design preparation, agentic orchestration, a Foundation Model for Physics and GPU-native physics kernels. The company claims three capabilities: no meshing, no approximations and no customer-specific training, described as zero-shot, while retaining solver-grade accuracy.

Verifiable performance comes from the EPTC 2025 paper and website cases. On a three-dimensional stacked package with ten layers spanning nanometre BEOL features through the full package, Vinci ingested industrial layout formats OASIS, GDS and IPC-2581 and resolved BEOL features below seven nanometres, ran 9,101 simulations averaging 218 million elements each in 26 hours total at 10.29 seconds each, and compared against commercial solvers at under two percent deviation, 240 times faster, with over 100 million degrees of freedom. A case with 117.4 million degrees of freedom took 20 seconds against two hours for a commercial FEA solver, with maximum temperature 100.00 versus 101.11 degrees Celsius, average 56.49 versus 56.50 and minimum 26.18 versus 25.93.

On customers the company says more than half of the top 20 semiconductor companies have benchmarked Vinci with results confirmed and that it is deployed in production programs at more than a dozen semiconductor and electronics organizations, without naming any. That is stronger evidence than pure marketing but remains company framing.

In physical AI terms Vinci occupies manufacturing-resolution physical AI: not general simulation but thermal, thermo-mechanical and convective analysis at chip, package and board level, a domain with extreme geometric complexity spanning five orders of magnitude, mesh sizes in the hundreds of millions and frequent design iteration. The value is moving simulation from hours per run to seconds per run so thermal design enters the design loop.

One framing point matters. This is the company with the hardest paper evidence and the earliest commercialization in this series. The EPTC paper gives checkable accuracy and speed at under two percent deviation and 240 times faster, far stronger than most AI simulation claims. Yet customers are anonymous, revenue is undisclosed, patent status is unverified, and the combination of zero-shot, meshless and deterministic is an extremely strong technical claim requiring independent reproduction. The key questions are robustness of the meshless method on arbitrary geometry, generalization of zero-shot inference to new materials and processes, and the feasibility of extending from semiconductor thermal analysis to automotive, aerospace and satellites.

Company Overview

Origins and Team

Vinci was founded in late 2023 and is headquartered in Palo Alto. Kabaria was vice president of software at Carbon, the 3D printing company, an experience relevant to the technical route, since Carbon used numerical simulation of the photopolymerization process to control printing and understood that simulation must be fast enough to enter the control loop. Osentoski brings robotics and machine learning, Rerolle leads business and John Bruggeman leads marketing.

The combination is not an academic project commercialized but industrial software veterans paired with physicists. Kabaria's Carbon experience taught two things: industrial customers do not pay for slightly faster simulation, only for an order of magnitude gain that changes workflows, and simulation must enter the design loop by removing manual preparation such as meshing and boundary setting.

Stealth and First Appearance

In December 2025 the company emerged disclosing 46 million dollars across seed and Series A, with the A round involving Xora, Khosla and Eclipse without disclosed valuation, and published at EPTC 2025. The sequence of publishing before fundraising matters, using peer review to build credibility and credibility to raise capital.

EPTC is a top conference in electronic packaging with peer review. Publishing industrial validation there is a strategy of self-certifying by academic standard, more credible than a white paper though still short of independent third-party evaluation, since authors are employees and test design is company-led.

Series B and Expansion

The 250 million dollar Series B of October 6, 2026 at 1.5 billion valuation was co-led by Advent, Temasek and Xora with strategic participation from AMD Ventures. Proceeds extend the platform from semiconductors to memory, advanced computing, automotive, aerospace and satellites, and to more physical phenomena including fluid and electromagnetic.

AMD Ventures participation is strategic, since AMD is a major chip and packaging customer. Brian Amick, AMD senior vice president of technology and engineering, said in the release that bringing fast accurate simulation into design lets teams evaluate trade-offs earlier, iterate faster and make better architectural decisions while there is still time to act, which is industrial endorsement of product value.

Technology

Three Bottlenecks of Conventional FEA

Understanding the value requires understanding three bottlenecks of finite element analysis in semiconductor packaging.

Meshing comes first. FEA discretizes continua into elements whose quality determines accuracy and convergence. Packaging geometry is extremely complex, with BEOL line widths below seven nanometres and package dimensions in centimetres spanning five orders of magnitude, and resolving both nanometre features and the full package yields meshes of hundreds of millions to billions of elements requiring hours to days and failing easily.

Solution time comes second. A thermo-mechanical solve with hundreds of millions of degrees of freedom takes hours on CPU clusters, so engineers simulate only at key milestones rather than after every change.

Manual setup comes third. Boundary conditions including convection coefficients, radiation and contact resistance, material properties varying with temperature and layer-by-layer interfaces all require manual specification, with heavy experience dependence and error risk.

Together these make simulation a checkpoint at the end of design rather than a constraint inside it.

What Meshless, Zero-Shot and Deterministic Mean

Vinci's three claims map onto the three bottlenecks: meshless solves meshing, zero-shot solves per-design retraining and deterministic solves reproducibility for certification.

Meshless methods are not new academically, with meshfree finite element variants, boundary elements and spectral methods all avoiding meshes, but achieving industrial-grade complex geometry without approximation is hard. Three implementations are possible: implicit geometry such as signed distance fields with numerical integration, point cloud or voxel discretization avoiding body-fitted meshes, or neural operators learning geometry-to-field mappings and skipping discretization entirely.

Zero-shot means a pretrained model applies to new designs without fine-tuning, which technically requires learning operator representations of physical laws rather than fitting specific geometry, unlike conventional surrogates that resample and retrain per design space.

Deterministic means identical output for identical input, which matters because certification and review require reproducibility while sampling-based methods such as diffusion or Monte Carlo vary run to run. If inference is deterministic forward computation with fixed weights and inputs the claim holds, and if any sampling is involved it does not. Implementation details are undisclosed.

Evidence Strength of the EPTC 2025 Paper

The paper is the hardest evidence and deserves itemized reading.

The test object is a 3D stacked package with ten layers spanning nanometre BEOL features through the full package, a design that deliberately tests both nanometre resolution and package scale, precisely where conventional FEA struggles.

Input handling reads OASIS, GDS and IPC-2581 industrial layout formats directly without manual conversion, a detail showing automated preparation handles real design files rather than simplified imports.

Scale and speed covers 9,101 simulations averaging 218 million elements in 26 hours at 10.29 seconds each, showing single runs in seconds and batch capability equivalent to parameter sweeps.

Accuracy against commercial solvers shows thermal field deviation under two percent, similar hot-spot-to-average ratios and effectively identical isotherm contours under comparable modeling assumptions.

The methodology is rigorous with peer review, explicit comparison and disclosed scale and timing, but two limits remain: the commercial solver is unnamed, preventing judgment of the baseline, and testing is company-led without third-party reproduction.

Difference From Surrogate Routes

Vinci repeatedly distinguishes itself from conventional surrogates, a technically important distinction.

Surrogates sample high-fidelity results within a specific design space, train a fast approximation and use it for optimization in that space, limited by retraining per space and poor extrapolation.

Vinci's logic is a pretrained physics foundation model learning the laws themselves, covering conduction, convection and radiation operators, so it applies to any new geometry without retraining. If true the commercial implication is large, converting project-based modeling services into a standardized software product.

It is also the claim most in need of verification, since zero-shot generalization to arbitrary geometry at two percent accuracy is a very strong statement in machine learning. Possible explanations are training on large simulation corpora yielding robust representations of geometric variation, or an internal fast numerical solver, GPU-native kernels, with AI handling acceleration and automation. The company mentions both a foundation model and GPU-native physics kernels, implying a hybrid architecture.

Technical Shortfalls

First, physics coverage. Validated domains are thermal, thermo-mechanical and convective, with fluid and electromagnetic planned. Different domains have different equation types, elliptic versus hyperbolic versus parabolic, and whether one model unifies them is unknown.

Second, material and process generalization. Packaging involves dozens of materials including silicon, copper, dielectrics, solder and underfill with temperature-dependent properties, and whether the model covers these and extrapolates to new materials is undisclosed.

Third, patent status. Third-party review found no verified patents, a signal for a company claiming a foundation model: either protection relies on trade secrets or patent filing is incomplete.

Fourth, missing third-party reproduction. The EPTC test is company-led without independent replication by a body such as NIST or a semiconductor research institute.

Product and Commercial Model

Product Form and Delivery

The product is a software platform delivered either in the cloud or behind customer firewalls, the latter essential for semiconductor customers whose layouts are core secrets and cannot leave.

The capability chain runs from layout ingestion through OASIS, GDS and IPC-2581 to automated design preparation covering geometry cleanup and material identification, agentic orchestration deciding physical models and boundary conditions, foundation model solving and result output.

Each link carries commercial meaning: automated layout ingestion reduces manual conversion, automated preparation reduces modeling time, agentic orchestration reduces dependence on specialists and foundation model solving cuts time from hours to seconds.

Customers and Deployment Scale

The company says more than half of the top 20 semiconductor companies have benchmarked with confirmed results and that it is deployed in production at more than a dozen organizations, naming none.

The disclosure style is common in semiconductors where customers avoid publicly acknowledging new tools for competitive reasons, but it also prevents external verification. Indirect verifiable evidence is AMD Ventures investment and public endorsement by an AMD executive, since AMD would not invest strategically in unproven technology.

Inferred Business Model

Three models are possible: software licensing by seat or compute, cloud services billed per simulation and enterprise deployment with one-time license plus annual maintenance. Semiconductor customers typically prefer enterprise deployment for data residency, making the third most likely.

Pricing is undisclosed. Industrial software of this type usually prices against engineering time saved, since reducing a run from two hours to 20 seconds allows hundreds of iterations per team per year and the value is quantifiable.

Expansion Path

Series B proceeds extend physics and industries toward memory, advanced computing, automotive, aerospace and satellites.

The logic is same physics, different scenario: thermo-mechanical simulation is needed in semiconductor packaging, automotive electronics, avionics and satellite thermal control with identical equations but different geometry and boundary conditions. If the foundation model truly learned physics rather than geometry the expansion is sound.

Expansion carries risk since automotive and aerospace certification regimes such as AEC-Q, DO-178C and AS9100 are stricter than semiconductors and demand more validation, and moving from semiconductor thermal analysis to airworthiness certification requires substantial verification work.

Competitive Landscape

Direct Rivals

The closest rival is PhysicsX of the United Kingdom, which raised 300 million dollars at a 2.4 billion valuation in 2026, also doing physics AI simulation starting from aerospace and automotive. Others include BeyondMath in the United Kingdom for AI acceleration of CFD, Luminary Cloud in the United States with 72 million dollars for cloud CFD and Neural Concept in Switzerland with 100 million dollars for CAD-native graph neural network simulation.

The routes differ in acceleration method: PhysicsX uses geometric deep learning, BeyondMath differentiable solvers, Neural Concept graph neural networks and Vinci a foundation model with GPU kernels. Which wins depends on the best balance of accuracy, speed and generalization.

Relationship With EDA and CAE Incumbents

EDA vendors Synopsys, Cadence and Siemens EDA and CAE vendors led by Ansys, now part of Synopsys, all pursue simulation acceleration, but their path is optimizing existing solvers through GPU acceleration, parallelization and cloud rather than rewriting the solving paradigm.

Vinci's opportunity lies in paradigm difference: if meshless plus foundation models genuinely bypass meshing and per-design training it holds a structural speed advantage, while incremental optimization would struggle to displace Ansys.

Relationship With Chip Companies

AMD Ventures investment shows chip industry interest. Thermal management challenges are intensifying as AI accelerator power density rises from hundreds to over a thousand watts per die and 3D stacking and chiplets complicate thermal paths, making package-level thermo-mechanical stress a reliability issue.

The demand is real and growing, and it explains why semiconductors were chosen as the first market.

Financing and Capital Structure

Funding History

Founded in late 2023, the company emerged in December 2025 disclosing 46 million dollars across seed and Series A and closed a 250 million dollar Series B at 1.5 billion valuation in October 2026, for about 296 million disclosed.

The Series B investor structure is notable, with Advent as global private equity, Temasek as a sovereign fund and Xora as deep tech venture co-leading, AMD Ventures participating strategically and Madrona, Eclipse and Khosla following. It marks a move from venture-backed validation to private equity and sovereign-backed scaling.

Valuation and Efficiency

A 1.5 billion valuation against 296 million raised implies roughly 20 percent dilution post-money. That sits mid-to-upper in AI simulation, between PhysicsX at 2.4 billion and Neural Concept near one billion.

The valuation rests on technical validation plus customer deployment, with the EPTC paper providing accuracy evidence and production deployment at dozens of organizations providing commercial evidence, though unnamed.

Use of Proceeds and Burn

A 250 million dollar round is ample for a company founded in 2023 with undisclosed headcount. At typical AI burn covering compensation and GPU compute it funds three to four years, and since proceeds extend physics and industries, revenue will not grow sharply near term while investment continues.

DISCREPANCIES

1. Two speedup figures. The homepage claims 1000 times faster while the EPTC 2025 paper reports 240 times faster. The two correspond to different tests, with 1000 times likely a best case and 240 times the paper average, but media conflate them. This report labels source and scenario. Source tier: OFFICIAL (website and paper) with differing definitions.

2. Degree of disclosure on customer count and identity. The company says more than half of the top 20 semiconductor companies benchmarked and deployment at more than a dozen organizations without naming any, and some interpretations treat this as many paying customers. This report labels it benchmarking and deployment without inferring paid scale. Source tier: OFFICIAL (company statements) distinguished from inference.

3. Unnamed baseline for the accuracy comparison. The EPTC paper reports under two percent deviation against commercial solvers without naming which. That affects interpretation, since deviation against the industry leader is more persuasive than against a secondary tool. This report notes the missing baseline. Source tier: OFFICIAL (paper) with baseline absent.

4. The meshless and approximation-free claims. The company claims no meshing and no approximations. Strictly, any numerical method involves discretization approximation whether through meshes, point clouds or neural representations, so approximation-free is marketing rather than technical fact. This report decomposes it into no body-fitted meshing and no numerical approximation, two different claims. Source tier: company marketing requiring technical reading.

5. Definitional boundary of zero-shot. The company says the model is zero-shot with no customer-specific training, but zero-shot can mean no data at all or no customer data while requiring general pretraining data. This report adopts the latter reading and notes training data sources are undisclosed. Source tier: OFFICIAL (statement) plus inference.

6. Patent status. Third-party review found no verified patents and the company publishes no list, creating tension with foundation model claims. This report labels patent status undisclosed. Source tier: third-party verification (MEDIA) with company disclosure absent.

7. Wording of the Series B leads. The company describes Advent, Temasek and Xora as co-leads while some coverage names only one. This report uses the three co-lead framing. Source tier: OFFICIAL (company announcement).

8. Timeline from semiconductors to automotive and aerospace. The company says proceeds support expansion to automotive, aerospace and satellites without a timeline or statement that validation has begun, while some interpretations treat these markets as entered. This report labels them planned expansion. Source tier: OFFICIAL (strategy statement) with progress data absent.

9. Two degree-of-freedom figures. The 6 October 2026 Series B release states the platform has processed more than 15 billion degrees of freedom in total, whereas the EPTC 2025 paper and website cases cite over 100 million degrees of freedom for individual runs (for example 117,440,512). The two are cumulative production volume versus single-run scale and are not the same metric; they should not be substituted for each other. Adjudication: record both with the caliber noted. Source tier: OFFICIAL (release and paper).

10. Headcount range. A third-party platform lists Vinci’s team size at about 101 to 250 people as of 2026, while the company discloses no headcount. This range is a third-party inference consistent with a 2023-founded company whose burn is driven by GPU compute and compensation. Adjudication: label as third-party and not a company disclosure. Source tier: PUBLIC (third-party platform).

Risks and Diligence Gaps

First, validation of zero-shot generalization. Applying a pretrained model to arbitrary new geometry at two percent accuracy is a very strong claim requiring third-party reproduction on independent datasets, while only a company-led EPTC paper exists.

Second, technical difficulty of physics expansion. Moving from thermo-mechanical to fluid and electromagnetic involves different equation types, and whether one architecture unifies them is unknown.

Third, anonymous customers and missing revenue. Deployment at a dozen organizations cannot be verified and revenue is undisclosed, so if customers are trialing rather than paying the commercial evidence weakens considerably.

Fourth, competition with EDA incumbents. Synopsys and Cadence hold complete toolchains and customer relationships, and integrating AI acceleration into existing products would compress Vinci's differentiation.

Fifth, certification regimes. Semiconductor reliability qualification, automotive functional safety and aerospace airworthiness all require traceable and verifiable simulation, and the black-box nature of AI models conflicts structurally with certification.

Sixth, patents and intellectual property. With patent status undisclosed, reliance on trade secrets raises copying risk.

Seventh, valuation decoupled from revenue. A 1.5 billion valuation against undisclosed revenue leaves material downside if revenue trails expectations.

Composite Assessment

Technology scores 8.0. The EPTC 2025 paper provides checkable accuracy under two percent deviation and speed at 240 times, the meshless plus zero-shot plus deterministic combination targets three real FEA bottlenecks, and automated layout ingestion and design preparation solve key industrial adoption steps. Deductions for missing third-party reproduction of zero-shot generalization, narrow physics coverage and unknown patent status.

Business scores 6.5. Production deployment at more than a dozen organizations and benchmarking by over half of the top 20 semiconductor companies show real-scenario entry, AMD Ventures strategic investment and executive endorsement provide industrial validation, and expansion from semiconductors to memory, automotive and aerospace is logical. Deductions for anonymous customers, undisclosed revenue and unknown pricing.

Finance scores 7.0. About 296 million dollars raised with a 250 million Series B at 1.5 billion valuation, and an investor base upgraded from venture to private equity and sovereign funds. Deductions for undisclosed revenue and heavy burn from expansion investment.

Product scores 7.5. The capability chain from layout ingestion through automated preparation and agentic orchestration to foundation model solving is complete, on-premises deployment meets semiconductor data requirements and 20-second runs change workflows. Deductions for narrow physics coverage, undisclosed interface and usability, and insufficient certification support.

Market scores 7.5. Rising AI chip power density and 3D stacking with chiplets make package thermal management essential, semiconductor customers pay well and expansion into automotive, aerospace and satellites is large. Deductions for EDA giant competition, certification thresholds and market education cost.

Overall 7.4. This is among the strongest paper evidence in this series: the peer-reviewed EPTC 2025 work demonstrates two percent deviation and 240 times speedup for semiconductor packaging thermal simulation, the meshless plus zero-shot plus deterministic combination attacks three real FEA bottlenecks and AMD strategic investment validates industrial value. Its weaknesses are missing independent reproduction of zero-shot generalization and opaque customers and revenue. The judgment on physical AI is that its physical core of GPU-native numerics plus a physics foundation model is a real engineering asset and that manufacturing-resolution simulation is a correct and defensible track, while moving from semiconductor thermal analysis to a multi-physics, multi-industry platform requires clearing generalization validation and certification.

References

  1. Vinci Series B announcement, October 6, 2026: 250 million dollars at a 1.5 billion valuation co-led by Advent, Temasek and Xora Innovation with AMD Ventures, Madrona, Eclipse and Khosla Ventures participating; about 296 million dollars of disclosed funding; proceeds for expanding physics domains and industries including memory, advanced computing, automotive, aerospace and satellites
  2. Vinci website, accessed 2026: positioning as physics-accurate design and simulation at manufacturing resolution; stack including automated design preparation, agentic orchestration, a Foundation Model for Physics and GPU-native physics kernels; claims of 1000 times faster, no meshing, no approximations and no customer data required to train the model; on-premises deployment behind customer firewalls; statement that more than half of the top 20 semiconductor companies have benchmarked Vinci with results confirmed
  3. EPTC 2025 paper as cited on the website: a 3D stacked package with ten layers covering BEOL nanometre features through the full package; ingestion of OASIS, GDS and IPC-2581 layouts resolving BEOL features below seven nanometres; 9,101 simulations averaging 218 million elements each in 26 hours at 10.29 seconds each; under two percent deviation against commercial solvers, 240 times faster with over 100 million degrees of freedom; similar hot-spot-to-average ratios and effectively identical isotherm contours
  4. Website case with 117,440,512 degrees of freedom: Vinci predicting maximum 100.00 degrees Celsius, average 56.49 and minimum 26.18 in 20 seconds against a commercial FEA solver at 101.11, 56.50 and 25.93 in two hours
  5. Company history from public sources: founded in late 2023 and headquartered in Palo Alto; founder and CEO Hardik Kabaria previously vice president of software at Carbon; co-founders Sarah Osentoski as CTO and Vincent Rerolle as CCO; John Bruggeman as CMO; emerged from stealth in December 2025 disclosing 46 million dollars across seed and Series A
  6. AMD statement in the Series B release: Brian Amick, senior vice president of technology and engineering, saying that designing advanced systems requires understanding how thermal, physical and electrical behavior interact across chip, package and board, and that bringing fast accurate simulation into design lets teams evaluate trade-offs earlier, iterate faster and make better architectural decisions while there is still time to act
  7. Technical description from public reporting: Vinci says its approach differs from conventional surrogates that require extensive training for specific applications, analyzing new designs without customer-specific tuning while maintaining solver-level accuracy; the platform is commercialized in semiconductors covering thermal, thermo-mechanical and convective analysis
  8. Expansion plans from public reporting: semiconductors as the first market with plans for memory, advanced computing, automotive, aerospace and satellites; longer-term ambition to help engineers determine design changes, manage constraints and eventually generate designs from high-level requirements
  9. Competitive landscape from public sources: PhysicsX with a 300 million dollar round at a 2.4 billion valuation in 2026, BeyondMath for AI-accelerated CFD, Luminary Cloud with 72 million dollars for cloud CFD, Neural Concept with 100 million dollars for CAD-native graph neural network simulation, and incumbents including Synopsys, Cadence and Ansys
  10. Industry background from public sources: semiconductor packaging geometry spans five orders of magnitude from nanometres to centimetres, conventional FEA meshes reach hundreds of millions of degrees of freedom and take hours to days to generate; AI accelerator power density rises from hundreds to over a thousand watts and 3D stacking with chiplets complicates thermal paths, making package-level thermo-mechanical stress a reliability issue
  11. Technical background from public sources: the three FEA bottlenecks are meshing whose quality determines accuracy and which fails easily on complex geometry, solution time of hours for hundreds of millions of degrees of freedom, and manual setup of boundary conditions, material properties and contact interfaces that depends on experience; meshless methods include meshfree finite element variants, boundary elements and spectral methods; surrogates require resampling and retraining per design space with poor extrapolation
  12. Third-party verification from public sources: Vinci has a peer-reviewed EPTC 2025 paper but no verified patents; the company reports production deployment at more than a dozen semiconductor and electronics organizations without naming customers
  1. Vinci Series B release (6 October 2026): states the platform has processed more than 15 billion degrees of freedom in total, versus the EPTC 2025 paper and website cases citing over 100 million for individual runs; the two are cumulative production volume and single-run scale respectively.
  2. Third-party platform (2026): lists Vinci team size at about 101 to 250 people; the company discloses no headcount, so the figure is a third-party inference.
This report is compiled from public sources for informational purposes only and is not investment advice. The company is private and revenue and customer data are undisclosed; the EPTC paper is a company-led test without independent third-party reproduction; the no-approximation and zero-shot claims are marketing statements decomposed technically in this report.

Related reports66

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.