Yunjing Zhifang Seed

Yunjing Zhifang Deep Dive: A Peking-University-Incubated Cloud-Native GPU CAE Company with the Simversus Platform and the YJ-DORGP Small-Sample Intelligent Solver, Reporting 275 Percent Order Growth, 160 Percent Revenue Growth and Breakeven in 2025

A Peking-University-incubated cloud-native GPU CAE company founded in April 2021 in Haidian, Beijing; its YJ-DORGP small-sample intelligent solver derives from PKU DORGP research; an angel round in December 2021 was invested jointly by Glodon and Guangji Venture Capital and a Pre-A round followed in 2023, with conflicting amount figures; it self-reports 275 percent order growth and 160 percent revenue growth in 2025.

Executive Summary

Beijing Yunjing Zhifang Information Technology Co., Ltd., branded Simversus, is a Peking-University-incubated Chinese company founded in April 2021 that builds cloud-native, GPU-parallel CAE simulation software. Its thesis is that CAE should be rebuilt from the algorithm layer up for the GPU and the browser rather than ported, and that a component-based platform can serve both end-user engineers and third-party software vendors who need simulation capability without rebuilding it. The company states its founding team comes from Peking University, that a National Distinguished Young Scholar co-founded the company and serves as chief scientist, that an academician serves as expert advisor, and that more than half of the team holds a master's degree or above.

The core product is Simversus, a cloud-native collaborative CAE platform the company describes as the industry's first to support full-workflow GPU parallel computing. Every stage of the simulation chain, from geometry and meshing through solving to post-processing and visualisation, has been redesigned around GPU parallelism, which the company claims delivers order-of-magnitude performance gains over existing algorithms. The platform supports multi-person online collaboration, twenty-plus mainstream commercial-software file formats with mutual conversion, a component library with low-code customisation, and a self-developed visualisation engine. In 2025 the company reported that its browser renderer stably handles models above one hundred million particles or cells, renders a two-gigabyte pre-processing file within ten seconds, and supports five-gigabyte-plus uploads in private deployments.

The most technically distinctive asset is the YJ-DORGP intelligent solver, jointly developed with Peking University researcher Chensen Ding on the basis of his DORGP research, short for Dual Order-Reduced Gaussian Process regression. The method trains an interpretable probabilistic surrogate from small samples and predicts large-scale complex spatio-temporal systems with quantified uncertainty, which the company positions as a low-energy AI path distinct from data-and-compute-driven approaches. The company states performance improvements of hundreds of times over finite-element methods for suitable real-time scenarios, support for training data generated by FEM, BEM, SPH and MPM as well as experimental data, and applications spanning multi-scale materials, linear and nonlinear structures, multiphase fluids, fluid-structure interaction and high-velocity impact and explosion, with validation reported on a major national water-conservancy project. A 2025 release removed the mesh-topology dependency so training samples no longer need identical meshes.

The commercial trajectory, as self-reported, accelerated sharply in 2025. The company's own year-end review states order value grew two hundred and seventy-five percent and revenue grew one hundred and sixty percent year on year, and that the company reached breakeven. It also discloses anonymous deployments: a research institute's component modal-analysis software and a state-owned enterprise's facility-simulation project for the new solver release, and named industry solutions for unnamed customers in shipbuilding, aerospace and power. Two funding rounds are on the record: an angel round in December 2021, invested jointly by Glodon and Guangji Venture Capital, and a Pre-A round that the company states it completed in 2023, with the angel amount recorded as close to ten million RMB in company-side material and as several million RMB in third-party databases. The company was certified as a national high-tech enterprise in October 2024 and appears on the 2026 Beijing innovative small and medium-sized enterprise list.

The 2026 pivot is explicitly toward physical AI. At the twenty-second China CAE Engineering Analysis Technology Annual Conference, founder and chief executive Zhu Fei presented a physics-informed intelligent simulation and simulation world-model framework, organised around five technical paths: physics-informed machine learning, neural-operator learning, geometric deep learning, physics foundation models, and world models with video-diffusion priors. Underpinning this is the YJPhysika physical-AI simulation engine, which the company describes as a geometry-AI plus physics-AI dual-drive system that automates geometry cleanup, repair and parametric modelling and achieves second-level prediction for hydrodynamics, structural strength and fluid-structure coupling, with claimed hundred-fold efficiency gains and design-iteration improvements of ten to one hundred times.

The positioning inside the physical-AI landscape is unusual: it is a platform-and-components company, not a single-product vendor. The company sells the Simversus cloud platform to engineering teams, a component family to software vendors, and custom simulation-software development to enterprises and research institutes in aerospace, shipbuilding and marine engineering. It has helped partners including CSSC Olantuo, PeriDyno and MultiFracS bring their own algorithms onto its cloud, joined the OpenUSD consortium, and integrated NVIDIA Omniverse data interfaces. Its Simversus platform has completed compatibility certification with Moore Threads M1000 GPUs on KylinOS, one of the few fully domestic CAE stacks on domestic GPU and domestic operating system.

Company Overview

Yunjing Zhifang was founded in April 2021 in Beijing and describes itself as incubated from Peking University, drawing on years of the university's technology and talent accumulation. Registration records place it at Room 1303, Tower A, Xueqing Jiachuang, No. 10 Xueqing Road, Haidian District, Beijing, and no Suzhou entity is associated with the company; the Wuxi name on record belongs to a customer, CSSC Olantuo Wuxi Software Technology. It states the mission of gathering professional strength to progressively realise the localisation of simulation products and the autonomy of simulation technology, and frames its vision as becoming an internationally advanced CAE simulation service and technology development platform. Founder and chief executive Zhu Fei holds a doctorate from Peking University and has worked on computer graphics, physics simulation, GPU parallel computing and deep learning, participating in national key research and development programmes, publishing papers, holding more than ten patent applications or grants, and winning a CCF technical invention first prize.

The founding posture is componentisation as a survival strategy, not only as an architecture choice. Zhu Fei has described the company's path as cloudification, componentisation, parallelisation and intelligentisation, and argued that covering the full simulation chain with flexibly combinable components serves everyone from end users to professional software vendors, calling it both a survival compromise and a co-development model. That framing matters because it explains why the company sells a platform, a component family and custom development simultaneously rather than betting on one.

The scientific lineage runs through Chensen Ding's group at Peking University. Ding is an assistant professor, researcher and doctoral supervisor in the Department of Mechanics and Engineering Science at Peking University, a national-level young talent and PKU Boya Young Scholar, who took his bachelor's and doctoral degrees at Hunan University with joint training at Peking University and the University of Minnesota's Minnesota Supercomputing Institute context, and held postdoctoral positions at the University of Luxembourg and the University of Exeter's Institute for Data Science and Artificial Intelligence. His DORGP line of work, including the dual order-reduced Gaussian-process emulator family published in CMAME and the ExDORGP lithium-battery uncertainty-quantification method featured on the cover of Acta Mechanica Sinica in 2026, is the academic base of the company's flagship solver component.

The product surface is broad for a company of this size. The Simversus platform covers the full chain; the component family includes the YJExchanger data-format converter, the YJSubdiv subdivision-modelling tool, the YJMesher meshing component and the YJVis scientific-visualisation component; the solver family includes YJStructure for structures, YJFluid for fluids, YJMultibody for multibody dynamics, YJProjdyna for elastic-rod dynamics and YJ-DORGP for small-sample intelligent solving; and the standalone YJViewer provides lightweight browser-based CAD and CAE viewing. In September 2026 the company launched a standalone YJSubdiv page with an AI modelling assistant that generates 3D models from natural-language instructions.

The ecosystem play is the most underrated part of the strategy. The company has helped CSSC Olantuo, PeriDyno and MultiFracS, the finite-discrete-element multiphysics fracture software developed by Professor Yan Chengzeng, run their own algorithms on Simversus, joined the OpenUSD consortium, and deeply integrated NVIDIA Omniverse data interfaces to connect CAE data into digital-twin and industrial-metaverse stacks. A platform that hosts third-party solvers accrues switching costs that a single-vendor platform does not.

Funding is disclosed at the round level, headcount at the registration level, and revenue only indirectly. Business-registration records give seventeen social-insurance participants for 2025, registered capital of 1.177212 million RMB and paid-in capital of 0.327212 million RMB, while the company's own 2026 milestone page describes a twenty-person full-time core team and a recruiting-platform listing places the unit at fewer than fifty people. On the commercial side, public procurement records show four awarded or candidate-award entries between 2024 and 2026, none of them disclosing an amount, alongside the self-reported 2025 growth and breakeven claims.

Common Misreadings: Four Errors That Change the Analysis

Yunjing Zhifang is not another full-suite CAE vendor, and it does not need to be. Its dual form of cloud platform plus algorithm components means its customer set includes other software companies, not only engineering end users. Reading it as a SimScale or Simversus-style platform competitor alone misses the component business, which is where the survival logic lives.

The flagship solver is jointly developed, not wholly self-invented. YJ-DORGP is built on Chensen Ding's DORGP research at Peking University, and the company describes the relationship explicitly as joint development based on the researcher's work. That lineage is the differentiation and also a dependency: the academic relationship, the researcher's own programme and any competing commercialisation of the same research are all part of the risk picture.

The hundreds-of-times performance claim is a scoped comparison, not a general replacement. It is stated against finite-element methods for suitable real-time scenarios, and the intelligent solver still requires training data, small-sample but not zero-sample, generated by numerical methods or experiments. A reader who treats the claim as a solver replacement for general CAE workloads will overstate the coverage.

Breakeven and the 2025 growth rates are company self-reports. The two-hundred-and-seventy-five-percent order growth, one-hundred-and-sixty-percent revenue growth and breakeven milestone come from the company's own annual review, with no disclosed base figures and no audited statement. They are directional evidence of commercial traction, not verified financials.

The reading that no financing has been disclosed is outdated. An angel round and a Pre-A round are both on the record, with Glodon and Guangji Venture Capital as investors, and the angel amount carries two figures: close to ten million RMB in company-side material and several million RMB in third-party databases.

Neither Wuxi nor Suzhou is the company's location. The registered address is in Haidian District, Beijing; Wuxi appears inside a customer's company name, and no Suzhou entity is associated with the company.

Technology

The technical claim has three layers: a GPU-native simulation chain, a small-sample probabilistic solver, and a 2026-vintage physics-AI engine that fuses geometry AI with physics AI.

The GPU-native chain is the platform-level differentiator. Simversus is described as the industry's first high-performance CAE collaborative simulation cloud platform supporting full-workflow GPU parallel computing, with the company stating that parallelisation is in the product's genes and that it rebuilt and redeveloped algorithms across the whole simulation chain on GPU-parallel principles, achieving order-of-magnitude performance improvements over existing algorithms. The visualisation layer is the most concretely evidenced part: the 2025 claims of stable rendering above one hundred million particles or cells in the browser, two-gigabyte pre-processing files rendered within ten seconds, five-gigabyte-plus uploads in private deployments, WebGPU support alongside WebGL in the YJVis component, and a restructured mesh file format with dynamic precision adjustment to cut first-frame latency and transfer cost.

YJ-DORGP is the method-level differentiator. Built on Bayesian inference and dual order-reduced Gaussian-process regression, it trains an interpretable surrogate from small samples, predicts large-scale complex spatio-temporal systems, and provides uncertainty estimates, which the company frames as low-energy AI against the data-and-compute-driven mainstream. It accepts training data from FEM, BEM, SPH and MPM simulations as well as experimental data, and the company states coverage of multi-scale materials, linear and nonlinear structures, multiphase fluids and fluid-structure interaction, and high-velocity impact and explosion, with validation on a major national water-conservancy project. The 2025 version removes the mesh-topology dependency, so training samples no longer require consistent meshes and the algorithm learns mesh shape features automatically, and the company reports applications in a research institute's component modal-analysis software and a state-owned enterprise's facility-simulation project.

The 2026 physics-AI layer is the strategic bet. The YJPhysika engine is presented as a geometry-AI plus physics-AI dual-drive system that automates geometry cleanup, repair and parametric modelling to lower the pre-processing barrier, and fuses neural operators with physics foundation models to deliver second-level prediction for hydrodynamics, structural strength and fluid-structure coupling, with the company claiming hundred-fold efficiency gains over traditional simulation and design-iteration improvements of ten to one hundred times, landing in design-simulation integration, real-time digital-twin rollout and embodied-intelligence training. The accompanying simulation-world-model framing argues that pure generative models have weak physical logic while traditional simulation models lack flexibility, and positions a physics-consistent, generative and visually faithful world model as the intelligent base connecting digital simulation to the physical world.

The honest reading is that the evidence is strongest at the visualisation and platform layer, credible but thin at the solver layer, and forward-looking at the physics-AI layer. The rendering numbers are specific and testable; the solver claims are anchored in a real peer-reviewed research line with a named academic and a published cover-paper result, but the company's own performance multiples have no third-party benchmark; the YJPhysika claims are the newest and least externally validated.

Product and Commercial Model

The commercial model has three surfaces: the Simversus cloud platform sold to engineering teams as SaaS, the YJ component family sold to software vendors as integrable capability, and custom simulation-software development sold to enterprises and research institutes as project work.

The platform business is the volume play. Simversus offers a public-cloud trial path, multi-person collaboration, twenty-plus commercial-software format conversions, component-library co-development with low-code customisation, GPU acceleration with claimed order-of-magnitude gains over CPU, fast sharing and a self-developed post-processing visualisation engine. The component and solution pages target aerospace, shipbuilding and marine engineering, energy and power, and ordnance, with anonymous customer cases in ship, aerospace and power describing full-stack solutions built from the company's own solvers, meshing algorithms and post-processing.

The component business is the margin and leverage play. YJExchanger has been spun out as a standalone page, added usdc and hdf5 support and customer-proprietary formats, and reports users across earthquake simulation, water conservancy and hydropower, bridges and tunnels, and additive manufacturing; YJSubdiv has been spun out with an AI modelling assistant; YJViewer ships as a lightweight viewer with frequent version updates. Components sold into other vendors' products create revenue that does not compete with the platform's own end-user funnel.

The product surface widened again in 2026, and commercial presence is now visible in procurement records. The YJPhysika engine has expanded into a solver family of YJCollision, YJMultibody, YJFem, YJProjdyna, YJSph and YJMpm, and the YJ-DORGP intelligent solver is presented on the Simversus product site as a small-sample data-driven intelligent solver. On 2026-08-18 the company and the State Key Laboratory of Advanced Design and Manufacturing Technology for Complete Vehicles at Hunan University released a physical-AI solution for end-to-end intelligent vehicle design and simulation, and on 2026-09-24 it published a dedicated custom simulation software development solution together with four delivered cases. Platform-side releases include YJViewer v1.3 on 2026-08-04 and v1.5 on 2026-09-16, Simversus v1.23 on 2026-09-07 with VTP sequence animation, and the YJSubdiv v1.1 update on 2026-10-26, which added reference-image upload of up to four images per conversation, a built-in multi-view scene snapshot capability in the assistant, and completed interface adaptation for the Zhipu GLM model service. Procurement records add an awarded result on 2026-08-11 for the inquiry result on a split-negotiation and split-signature composite-structure modelling module functional test with CSSC Olantuo Wuxi Software Technology, a candidate-award entry on 2025-09-18 for point-to-point and transfer-learning aerothermal prediction algorithm integration and platform construction, a candidate-award entry on 2025-05-15 for integrated optimisation and development of a self-developed hydrodynamic CFD solver, and a candidate-award entry on 2024-09-05 for a multi-source heterogeneous aerodynamic data fusion model algorithm, with no amounts disclosed in any of them. The 2025 tax credit rating was A.

The custom-development business is the services play. A September 2026 solution page positions one-stop custom development of dedicated simulation software for enterprises and research institutes in aerospace, shipbuilding and marine engineering, covering requirements analysis, function selection, custom development, parallel deployment and operations, explicitly aimed at customers who face inadequate commercial-software fit, high barriers to autonomous software development and large-scale simulation compute bottlenecks. This is the surface that most directly monetises the domestic-substitution demand.

The 2025 self-reported financials are the traction evidence, and they are unaudited. Order value up two hundred and seventy-five percent, revenue up one hundred and sixty percent, and breakeven reached. For a company of fewer than fifty people on one recruiting listing, breakeven at that growth rate is a meaningful signal of commercial efficiency; the absence of disclosed bases, margins and audited figures is the reason the claim cannot be treated as verified.

The domestic-stack certification is a real moat in a specific segment. The Simversus compatibility certification with Moore Threads M1000 GPUs on KylinOS, and the national high-tech enterprise certification in October 2024, position the company for procurement contexts where foreign software licensing and data-security risk are explicit constraints, the same demand that benefits domestic CAE peers such as Intesim and Suochen.

Competitive Landscape

Yunjing Zhifang competes on three fronts simultaneously, which is unusual for a company of its size.

Against domestic cloud-CAE and AI-CAE peers, the comparison is architecture against breadth. Intesim offers a full self-developed solver stack with fifteen years of accumulation and a hundred-plus million RMB B round; Suochen, Yundao, Tianfu and Tenfong each bring deeper single-domain depth or larger balance sheets. Yunjing's defence is that it is cloud-native and GPU-native from the algorithm layer, that its component model lets it ride on partners' domain software, and that its small-sample solver targets a different operating point, real-time and low-data, than accuracy-first incumbents.

Against international cloud-native simulation platforms, the comparison is focus against scale. SimScale is the reference cloud-native CAE platform and has integrated deep-learning prediction through third parties; Luminary Cloud and Rescale occupy adjacent GPU and HPC-cloud positions. Yunjing's argument is domestic data residency, domestic GPU and operating-system certification, and a component model that international platforms do not offer to Chinese software vendors.

Against AI-solver and surrogate vendors, the comparison is method class against method class. PhysicsX and Neural Concept represent the deep-learning surrogate path at far larger scale; Yunjing's DORGP-based small-sample path is the probabilistic, interpretable, low-data alternative, closer in spirit to Secondmind's Bayesian framing than to the neural-surrogate mainstream. The two paths are complementary in principle, and no public integration exists.

Against in-house and open-source stacks, the comparison is productisation against cost. OpenFOAM and the JAX ecosystem give sophisticated teams a free GPU-capable base, and large enterprises run internal simulation platforms. Yunjing's answer is the full-chain packaging, collaboration, visualisation and format-conversion layer, which is where most engineering organisations actually spend pain.

The channel and ecosystem position is the least copyable asset. Hosting CSSC Olantuo, PeriDyno and MultiFracS algorithms on Simversus, joining OpenUSD, and integrating Omniverse interfaces create a partner graph that a single-vendor competitor cannot replicate quickly, and the custom-development surface keeps the company close to the aerospace and shipbuilding customers that anchor the domestic-substitution market.

Financing and Operations

Two funding rounds are on the record, and the amount of the first one is not settled. Third-party databases date an angel round to 2021-12-27, describe it as several million RMB and name Glodon and Guangji Venture Capital as investors; company-side material from the GAMES 2022 conference page describes the same round as close to ten million RMB, jointly invested by Glodon and Guangji Venture Capital. Registration records show the Glodon-affiliated vehicles entering on 2022-04-25 and 2022-05-17. The company's own development history records a Pre-A round completed in 2023, with no amount stated. This report records the phase as seed, the closest available label for a company at angel and Pre-A stage.

The financial picture mixes self-reported growth with procurement records. The 2025 annual review states order value growth of two hundred and seventy-five percent and revenue growth of one hundred and sixty percent year on year, and that the company reached breakeven, which it frames as a key milestone achieved through strong growth momentum and efficient operations. Public tender records add four awarded or candidate-award entries between 2024 and 2026 with no amount disclosed, and a 2025 tax credit rating of A. No base figures, margin structure, cash balance or burn rate accompanies these claims.

The 2025 to 2026 product cadence is unusually dense for a company at seed stage. YJExchanger standalone, the mesh-topology-independent solver release, WebGPU support in YJVis, the CAD-CAE integration push, cloud subdivision modelling, the YJPhysika physics-AI engine, YJSubdiv with an AI modelling assistant, YJViewer versions 1.3 and 1.5, Simversus versions 1.19 through 1.23, a valve parametric-design solution, and a dedicated custom-development solution page, all inside roughly eighteen months. Either the company is capital-efficient at small scale or it is running on services revenue and early-stage capital; both readings point to a business still close to its founding economics.

The open financing questions are now specific rather than absolute. A company claiming breakeven while shipping a physics-AI engine, a world-model framework and a component family is either under-invested relative to its ambition or deliberately funding growth from services margins. The amount, valuation and cash position behind both disclosed rounds are not public, so neither reading can be tested.

DISCREPANCIES

1. Funding status. Convention A (company and databases): an angel round and a Pre-A round are on the record, the angel round dated December 2021 with investors Glodon and Guangji Venture Capital, and the company development history recording a Pre-A round completed in 2023. Convention B (registration records): Beijing Glodon Chuangyuan Investment Center holds 10.04 percent and Tianjin Glodon Jianke Investment Management Partnership holds 5.02 percent, with entry dates 2022-04-25 and 2022-05-17. Ruling: the disclosure exists and the earlier no-financing reading is withdrawn; the frontmatter phase moves from bootstrapped to seed, the closest available label for an angel and Pre-A stage company in this library's enumeration.

2. Angel-round amount. Convention A (company side): an angel investment of close to ten million RMB jointly from Glodon and Guangji Venture Capital, received at the company's founding, whose only source is the GAMES 2022 conference company page, repeated in the company's public account and in its exhibitor note for the 19th China CAE Annual Conference. Convention B (third-party databases): Qingke VC, InnoHere and Fengchao all record several million RMB for a round dated 2021-12-27. Ruling: the two conventions are reported side by side and not merged; the company-side figure has no investor or round announcement behind it and the database figure is unconfirmed by the company, so the range stays at several million to close to ten million RMB.

3. Angel-round investors. Convention A: company-side wording names Glodon and Guangji Venture Capital as joint investors. Convention B: Baidu Baike records that the investors include a Glodon-affiliated institution, a more cautious wording. Ruling: reported as a joint investment by Glodon and Guangji Venture Capital, not an either-or choice, which the registration records support, with Glodon Chuangyuan Investment Center at 10.04 percent and Glodon-affiliated Tianjin Glodon Jianke at 5.02 percent.

4. Financing date. Convention A (database financing date): 2021-12-27. Convention B (registration entry dates): 2022-04-25 and 2022-05-17. Ruling: both are reported; the difference may reflect a signing date against a registration-change date, and no merge is attempted.

5. Headcount. Convention A (social insurance): seventeen participants in the 2025 annual report. Convention B (company statement): a twenty-person full-time core team in the 2026 milestones. Convention C (recruiting platform): fewer than fifty people, undated. Convention D (education mix): more than half of the team holds a master's degree or above. Ruling: all four are reported side by side and are not converted into one another; the social-insurance figure is the narrowest and the platform listing the widest, and each mention states its source type.

6. Year of the high-tech enterprise title. Convention A (company development history): the national high-tech enterprise item appears among the 2026 milestones. Convention B (registration qualification records): certificate GR202411001267, obtained in 2024 and valid to 2027-10-29. Ruling: recorded as obtained in 2024 per the registration records, with a note that the company timeline is imprecise on this point.

7. Company location versus Wuxi and Suzhou. Convention A (registration and company about page): registered address Room 1303, Tower A, Xueqing Jiachuang, No. 10 Xueqing Road, Haidian District, Beijing. Convention B (procurement records): CSSC Olantuo Wuxi Software Technology appears in the inquiry result announcement of 2026-08-11. Ruling: Wuxi is a customer location and unrelated to the registered address; no Suzhou entity is associated with the company.

8. Level of the innovative SME recognition. Convention A (government announcement): the ninth batch of innovative small and medium-sized enterprises published by the Beijing Municipal Bureau of Economy and Information Technology on 2026-09-07, where the company is listed as entry 10 among 384 in Haidian District. Ruling: recorded as a Beijing municipal recognition, and the recognition has no connection to any Jiangsu provincial list.

9. Performance claims. Convention A: hundreds-of-times improvement over finite-element methods for suitable real-time scenarios (company, YJ-DORGP). Convention B: order-of-magnitude gains over existing algorithms from GPU parallelism (company, Simversus). Convention C: hundred-fold efficiency and ten-to-one-hundred-times design-iteration gains from YJPhysika (company, 2026). Ruling: all three are company claims against different baselines and scopes, none with a third-party benchmark; they are reported as scoped claims with their baselines stated rather than merged into a single speed number.

10. 2025 financials and revenue disclosure. Convention A (company self-report): order value up two hundred and seventy-five percent, revenue up one hundred and sixty percent, breakeven reached. Convention B (absence of disclosure): no audited or filed financial statement and no base figures. Convention C (public procurement): an awarded result on 2026-08-11 and candidate-award entries on 2025-09-18, 2025-05-15 and 2024-09-05, none disclosing an amount. Ruling: the self-reported figures are reported as self-reports with undisclosed bases, commercial presence is supported by the procurement records, and revenue amounts remain undisclosed, recorded as several bid awards with no amounts published.

11. Customer identity. Convention A: named industry solutions for unnamed customers in shipbuilding, aerospace, power and vehicles (company site). Convention B: named academic collaborators including Chensen Ding at Peking University and Yan Chengzeng for MultiFracS (public). Ruling: the industrial customers are deliberately anonymous in company materials and are reported as anonymous cases; the academic collaborators are public and named.

12. Founded year and lineage. Convention A: founded April 2021 (company). Convention B: incubated from Peking University with the seed germinating there over years (company narrative). Convention C: the underlying DORGP research line publishes from 2023 onward in the academic record (public literature). Ruling: these are sequential, not conflicting; the company is founded in 2021 on a research base that continues to publish, and the joint-development framing is retained throughout.

Risks and Due Diligence Gaps

The largest gap is financial verification. The two disclosed rounds carry no reconciled amount or valuation, and no cash, margin or audited revenue is disclosed, while the growth and breakeven claims have no published bases. Due-diligence action: request audited or management accounts for 2024 and 2025, order-backlog composition and concentration, and the cash position behind the physics-AI build-out.

The flagship solver's performance claims lack third-party benchmarks. The hundreds-of-times figure against finite-element methods, the order-of-magnitude GPU gains and the YJPhysika hundred-fold claims are all company-stated against different baselines. Due-diligence action: an independent benchmark on a standard structural and a standard fluid case, with error tolerances defined before testing.

The academic dependency is a real concentration. YJ-DORGP's joint-development relationship with a single Peking University researcher ties the flagship component to one person's research programme and availability, and the company's chief scientist is a National Distinguished Young Scholar whose name the company does not publish. Due-diligence action: map the IP ownership and licensing terms of the DORGP line, and the researcher's other commercial commitments.

The team is small against the product surface. A full-chain platform, a component family, a solver family, a physics-AI engine and custom-development services, delivered by a unit that business registration places at seventeen social-insurance participants, company material describes as a twenty-person full-time core team, and one external listing places at fewer than fifty people, implies either narrow per-product depth or heavy partner reliance. Due-diligence action: engineering headcount by product line and the partner-dependency map.

The domestic-substitution market is crowded and better capitalised. Intesim, Suochen, Yundao, Tianfu and Tenfong all compete for the same sovereignty-driven procurement, several with far larger financing and longer track records, and the company's anonymous customer cases cannot be independently scaled. Due-diligence action: named reference customers and contract values.

The physics-AI pivot is early and crowded at the top. Neural operators, physics foundation models and world models are the most contested claims in the industry, and the company's YJPhysika results are the least externally validated part of its portfolio. The strategic risk is spending physics-AI credibility before the platform business is profitable at verified scale.

Overall Assessment

  • Technology 7.5. A genuinely differentiated method asset in YJ-DORGP, built on a named, publishing academic line with a cover-paper result, packaged as a small-sample, interpretable, low-energy solver; a GPU-native full-chain platform with specific, testable visualisation performance claims; and a coherent 2026 physics-AI framing in YJPhysika. Deducted points: the performance multiples are company claims against different baselines with no third-party benchmark, the physics-AI layer is the newest and least validated, and the flagship solver depends on one academic relationship.
  • Commercialisation 6.0. Self-reported two-hundred-and-seventy-five-percent order growth, one-hundred-and-sixty-percent revenue growth and breakeven in 2025; anonymous but concrete deployments in shipbuilding, aerospace, power and vehicles; a component business spun out into standalone products with cross-industry users; and a dense product cadence. Deducted points: no audited financials, no disclosed bases, no named industrial customers, and a revenue model that leans on services and component sales whose margins are unknown.
  • Financing 5.5. Two disclosed early rounds, an angel round in December 2021 backed by Glodon and Guangji Venture Capital and a Pre-A round in 2023 stated by the company; registered capital of 1.177212 million RMB with 0.327212 million RMB paid in; seventeen social-insurance participants; a national high-tech enterprise certification; and a self-reported breakeven that implies capital efficiency at small scale. Deducted points: the angel amount is unresolved between company-side and database figures, no valuation is disclosed, and for a company shipping a physics-AI engine and a world-model framework the size and cash position behind the build-out remain the single largest unknown in the file.
  • Product maturity 7.0. A delivered platform with versioned releases through 1.23, a component family with two standalone spins, a five-member solver family including the flagship intelligent solver, a lightweight viewer, a domestic GPU and operating-system certification, and an ecosystem that hosts third-party solvers. Deducted points: no third-party benchmark of the headline performance, anonymous industrial references, and per-product depth unverifiable at the stated team size.
  • Market position 6.0. A distinctive operating point at the intersection of cloud-native GPU CAE, small-sample probabilistic solving and component supply, with a real partner graph and a domestic-stack certification in a sovereignty-driven market. Deducted points: a crowded domestic CAE field with better-capitalised peers, an unreconciled financing record, an academic dependency on the flagship asset, and physics-AI claims in the most contested corner of the industry.

References

  1. Yunjing Zhifang company site and news centre, including the company profile, the Simversus platform description, the YJ component and solver family, the industry solutions for unnamed shipbuilding, aerospace, power and vehicle customers, the Moore Threads M1000 and KylinOS compatibility certification, and the 2025 annual review with order, revenue and breakeven claims
  2. Yunjing Zhifang product announcement for the YJ-DORGP intelligent solver, including the joint development with Peking University researcher Chensen Ding on DORGP, the Bayesian basis, small-sample training, hundreds-of-times performance over finite-element methods, support for FEM, BEM, SPH and MPM data, the application scope, and the national water-conservancy project validation
  3. Yunjing Zhifang 2025 annual review, including the order-value and revenue growth rates, the breakeven milestone, the mesh-topology-independent solver release with research-institute and state-owned-enterprise applications, the rendering and WebGPU upgrades, the OpenUSD and Omniverse integration, and the partner algorithm cloud enablement for CSSC Olantuo, PeriDyno and MultiFracS
  4. Yunjing Zhifang news items from 2026, including the YJSubdiv standalone page and AI modelling assistant, the dedicated simulation-software custom-development solution, YJViewer 1.3 and 1.5, Simversus 1.23, and the valve parametric-design solution
  5. Yunjing Zhifang conference report on physics-informed intelligent simulation and simulation world models at the twenty-second China CAE Engineering Analysis Technology Annual Conference, including the five technical paths, the YJPhysika engine description and the claimed efficiency and design-iteration gains
  6. Peking University faculty and group pages for Chensen Ding, including his position, national-level young talent and Boya Young Scholar status, the DORGP and ExDORGP research lines, and the representative publications
  7. Acta Mechanica Sinica 2026 cover article on ExDORGP for lithium-battery uncertainty quantification, including the ten-sample accuracy result, the thirty-fold acceleration and the author list
  8. Simulation-industry coverage of emerging AI-CAE companies placing Yunjing Zhifang among the domestic AI-simulation entrants alongside Tianfu, Intesim and Tenfong
  9. Zhu Fei speaker profiles and talks, including the Peking University doctorate background, the CCF technical invention first prize, the cloudification-componentisation-parallelisation-intelligentisation path, and the Simversus SaaS plus PaaS model
  10. Third-party vendor and recruiting listings, including the company profile with fewer-than-fifty headcount and the high-tech enterprise certification announcement of October 2024
  11. Company page at the GAMES 2022 conference, stating that the company received an angel investment of close to ten million RMB jointly from Glodon and Guangji Venture Capital at its founding
  12. Third-party financing databases Qingke VC, InnoHere and Fengchao, each recording an angel round of several million RMB dated 2021-12-27
  13. Business-registration records on Qichacha, including registered capital of 1.177212 million RMB, paid-in capital of 0.327212 million RMB, seventeen social-insurance participants in 2025, the angel-round label, the Glodon-affiliated fund entries dated 2022-04-25 and 2022-05-17, and qualification records including GR202411001267
  14. Beijing Municipal Bureau of Economy and Information Technology announcement of the ninth batch of innovative small and medium-sized enterprises dated 2026-09-07, together with the Haidian District public notice
  15. Company about page, including the registered address, the development history recording a Pre-A round in 2023, the twenty-person full-time core team in the 2026 milestones, and the national high-tech enterprise entry placed in 2026
  16. Company news item on YJSubdiv v1.1 dated 2026-10-26, covering reference-image upload of up to four images, the built-in multi-view scene snapshot capability, and the completed Zhipu GLM model service interface adaptation
  17. Company news index and product site, covering the joint physical-AI solution with the Hunan University vehicle-design laboratory dated 2026-08-18, the dedicated simulation software custom-development solution with four delivered cases dated 2026-09-24, YJViewer v1.3 and v1.5, Simversus v1.23 with VTP sequence animation, the YJPhysika solver family, and the 28th Beijing Science and Technology Expo
  18. Investment-network company profile recording registered capital of 1.1772 million RMB
  19. Baidu Baike entry recording the December 2021 angel round with investors including a Glodon-affiliated institution, and twenty employees in 2023
This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of 2 November 2026; please refer to official company announcements and filings.

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.