TenFong Deep Dive: An Academician-Led Chinese Industrial Simulation Software Vendor Whose Official Timeline and Database Records Use Incompatible Round Names, With an Official Financing PDF and an Investment Database Both on File

Executive Summary

Shenzhen TenFong Technology Co., Ltd. is a Chinese industrial simulation software company whose most distinctive asset is not a product or a funding round but a person. Chen Shiyi, an academician of the Chinese Academy of Sciences and an academician of the Third World Academy of Sciences, one of the founders of turbulent boundary layer theory and one of the inventors of the lattice Boltzmann method, serves as chief scientist and chief adviser. Around that academic core the company has assembled a commercial industrial simulation stack of nearly twenty products, three self-developed neural operators, a set of institutional credentials that are difficult for a startup to assemble, and a financing history that is documented in two mutually inconsistent forms.

The company is unusual in the Chinese simulation market because its scientific pedigree is genuinely first tier. Chen Shiyi is not a figurehead imported for fundraising credibility; the lattice Boltzmann method that underpins several of the company's core products was developed by the group he leads. The company's own technical claims are consistently described in terms of solving real industrial problems with methods that trace back to that lineage.

The product line is broad and, unusually, the claims come with named quantitative targets. Fluid products include TF-QFLUX in a GPU-accelerated version, TF-Lattice for lattice Boltzmann work, TF-CFlow and TF-SPH. Structural products include TF-Struct, TF-Dyna and TF-DCAMS. Multidisciplinary products include TF-eMag, TF-Acoustics and TF-DEM. The platform layer includes TF-AIDEA, released on 20 September 2026, together with TF-AIMDO and TF-Pandroid, and a grid product named ZF-Grid V2026 released on 25 September 2026. A bundled suite named TF-FEAST was upgraded on 13 October 2026. Three self-developed neural operators carry the AI claim: SATO, a spatially aware Transformer with dual attention; GeoFormer, which maps CAD geometry to flow fields; and DragSolver, presented at ICML 2025 with a reported relative error of 0.0005 and an R-squared of 0.97.

Every performance number is a vendor statement. The claim that a single 12GB GPU can carry a ten-million-cell mesh, that an hours-long computation becomes a result in under ten seconds, and that the error stays below five percent, all come from the company without a published model, mesh, baseline solver, or error tolerance. The most concrete application figure is an automotive account in which aerodynamic optimisation went from seven person-days to half a person-day, a fifteen-fold reduction, with drag coefficient deviation below two percent and wind resistance reduced by more than five counts. For a vehicle aerodynamic map the company reports a mean absolute percentage error of 1.16% and a trend accuracy of 94.34%, and across five physical quantities it reports normalised root mean square errors below five percent. None of these has been reproduced by a third party.

The financing record is the single largest documentary problem in this file, and it cannot be resolved by choosing one source. The company's own financing materials describe a timeline of three rounds: an A round of over CNY 100 million in December 2020 from Hillhouse, Green Pine and Nanshan; an A-plus round of several hundred million in December 2021 from China Development Bank Capital, Knowledge City, and the Shenzhen angel fund; and a B round of several hundred million in May 2023 from China Development Bank Manufacturing and others. Investment databases describe a different sequence with more steps: an angel round in April 2021, a pre-A round in May 2021 from Nanshan, Jingcheng and Hillhouse, an A round in August 2021 from Nanshan, Green Pine and Linear, an A-plus round in March 2022 of several hundred million from a nine-firm syndicate including China Development Bank Capital, Hillhouse, Basis, Green Pine, Haipeng and Renzhi, a B round in December 2023 from Jinglin, China Development Bank Manufacturing, Hangshi and Wuchuang Institute, and a B-plus round on 27 March 2024 from Xietong Investment with the amount undisclosed. One Chinese industry outlet recorded three rounds totalling CNY 75 million, which is an order of magnitude below every other figure on record.

The company has not disclosed revenue, and it is not yet profitable in any documented period. What is documented instead is institutional: a national high-technology enterprise designation, CMMI level three, a specialised-and-sophisticated small giant status, a Shenzhen postdoctoral practice base established in January 2025, a Guangdong provincial engineering research centre for multidisciplinary intelligent optimisation CAE established in July 2025, and a 2026 national case selection for domestically independent industrial simulation software. Headcount is given as 128 in the 2024 annual report with more than seventy percent holding master's or doctoral degrees, alongside more than thirty software copyrights and more than three hundred granted patents.

The physical-AI relevance of this company is concentrated in one direction: the last few percent of aerodynamic performance. In a vehicle programme where the drag coefficient is already measured in hundredths, a tool that shortens an aerodynamic optimisation loop from seven person-days to half a day changes the number of design iterations a team can afford, and that is a real economic benefit to a manufacturing customer. The company's own claim, that a single RTX 4070 performs second-level inference on an eight-million-node case, is the metric that would matter to that customer, and it is also a claim with no published benchmark behind it.

Company Overview

The entity is Shenzhen TenFong Technology Co., Ltd., incorporated on 17 December 2020 with unified social credit code 91440300MA5GJ4M99E and registered capital of CNY 16.729421 million. The registered address is room 2301, building D1, Nanshan Zhiyuan, 1001 Xuyuan Avenue, Nanshan District, Shenzhen. Zhang Rikui is the legal representative and general manager, holding a doctoral degree in physics from the School of Engineering at Peking University.

The academic core is the company's defining feature and it is verifiable. Chen Shiyi is an academician of the Chinese Academy of Sciences and of the Third World Academy of Sciences, and is widely regarded as one of the founders of the modern theory of turbulent boundary layers. The lattice Boltzmann method, one of the most widely used approaches for simulating complex fluid behaviour on GPUs, is closely associated with his research group. His role in the company is chief scientist and chief adviser.

The commercial organisation was assembled around that core over roughly three years. The product naming convention is a TF- prefix for the physics products and a separate line for platform products. Institutional recognition arrived in a recognisable sequence: national high-technology enterprise status and CMMI level three early on, specialised-and-sophisticated small giant status, a Shenzhen postdoctoral practice base in January 2025, and a Guangdong provincial engineering research centre for multidisciplinary intelligent optimisation CAE in July 2025.

Scale disclosures are specific but narrow. The 2024 annual report gives 128 employees with more than seventy percent holding postgraduate degrees, more than thirty registered software copyrights, more than three hundred granted patents, more than ten enterprise customers, more than twenty chain-leader strategic partnerships, and more than fifteen regional industrialisation bases. These are the numbers a counterparty should use, and they are the only ones published.

Ecosystem relationships are unusually concrete and are a real part of the competitive position. The company's solver technology is integrated with NVIDIA PhysicsNeMo, which hosts DoMINO, Transolver and GeoTransolver, with acceleration through NVIDIA AmgX and in-situ visualisation through OptiX, and NVIDIA carries a case study page for the company. The company co-founded the Ningbo Eastern Institute of Technology with a partner, established a joint innovation laboratory for automotive fluid simulation software with FAW, and its subsidiary Jingfeng Technology together with the Beijing Humanoid Robot Innovation Center released a platform called Horizon of Tomorrow, with the Tian Gong Jingfeng entry running 100 metres in 9.34 seconds at the 2026 World Humanoid Robot Games.

There is no public information on revenue, gross margin, burn rate, employee equity, or valuation. The most recent financing event on record is the B-plus round of 27 March 2024, whose amount was not disclosed.

Common Misreadings: Five Errors That Change the Analysis

The company's financing rounds are not the database's financing rounds. The official financing document and the investment database record the same company with different round labels and different round counts. Neither can be used to count rounds, and quoting one as the company's history while citing the other for amounts produces a factually incoherent answer.

The round count is not seven. The official materials describe three rounds. The database describes seven or more. The media figure of three rounds totalling CNY 75 million is a third count again, and the lowest amount. All three are reported below and none is presented as the answer.

Chen Shiyi is not the legal representative or the chief executive. The legal representative and general manager is Zhang Rikui, and the chief scientist role is a scientific and advisory position. Reading Chen Shiyi's standing as evidence of a specific corporate management structure is a category error, although his scientific authority is the source of the company's technical credibility.

Jingfeng Technology is a subsidiary, not a second business. The humanoid robotics results associated with Jingfeng and the Beijing Humanoid Robot Innovation Center belong to a subsidiary's collaboration, and should not be read as evidence that TenFong itself operates a robotics product line.

The company is not an LLM or agent company. Its AI layer consists of three neural operators embedded in a solver product line. There is no disclosed foundation model, no disclosed agent product, and no disclosed large language model component. Reading it against that category is a mis-mapping.

Technology

The technical story has three layers, and separating them is necessary because they have different maturity.

Layer one is the classical solver portfolio, and it is the substrate the company actually sells. Fluid products are TF-QFLUX, with a GPU-accelerated version, TF-Lattice for lattice Boltzmann computation, TF-CFlow for general incompressible and compressible flow, and TF-SPH for smoothed particle hydrodynamics. Structural products are TF-Struct for static structural analysis, TF-Dyna for transient dynamics, and TF-DCAMS for dynamic and acoustic analysis. Multidisciplinary products are TF-eMag for coupled electromagnetics, TF-Acoustics for acoustics, and TF-DEM for discrete element method handling. Industry-specific products include TF-Thermal for thermal analysis, TF-SimFARM for wind farm simulation, and TF-SimCITY for urban scale simulation. This breadth is the most defensible part of the offer, because it is the part that a customer must replace to move away.

Layer two is the AI layer, and it consists of three named self-developed neural operators. SATO is a spatially aware Transformer with a dual-attention mechanism, aimed at capturing long-range spatial correlations in flow fields. GeoFormer converts CAD geometry directly into flow field prediction, which is the operation that matters most in a design loop because it removes the meshing and solving step from the inner iteration. DragSolver was presented at ICML 2025 and is the most concrete result: a reported relative error of 0.0005 and an R-squared of 0.97 against reference solutions. The published basis for these figures is not available in the sources consulted, so they are treated as peer-reviewed-track claims whose review status could not be independently confirmed.

Layer three is the platform and grid layer. TF-AIDEA, released on 20 September 2026, is the software platform; ZF-Grid V2026, released on 25 September 2026, is the mesh generation and grid technology; TF-AIMDO and TF-Pandroid are module and application frameworks; and TF-FEAST, upgraded on 13 October 2026, is a bundled suite positioned for rapid adoption. The release cadence across a single month in September and October 2026 is itself informative about how the company is currently organised around product shipping.

The integration with NVIDIA's physics machine learning stack is a genuine engineering endorsement and a dependency. The company's solvers are integrated with PhysicsNeMo, including DoMINO, Transolver and GeoTransolver, with acceleration through AmgX and in-situ visualisation through OptiX, and NVIDIA carries a case study page for the company. In exchange the company is tied to one vendor's physics machine learning roadmap and to that vendor's hardware pricing.

The performance claims are uniform in character and uniform in weakness. A single 12GB GPU carrying a ten-million-cell mesh, hours reduced to under ten seconds, error below five percent, an eight-million-node case on a single RTX 4070 in seconds, roughly five seconds for battery liquid cooling, more than one thousand wind fields computed by TF-SimFARM, a mean absolute percentage error of 1.16% and a trend accuracy of 94.34% on a vehicle aerodynamic map, normalised root mean square error below five percent across five physical quantities, and a fifteen-fold reduction in person-days on an aerodynamic optimisation case. Every one of these is a company statement with no published model, mesh, baseline solver, hardware specification, or error definition, and the metric definitions in particular are not comparable across the list because the underlying quantities differ.

One competitive result is verifiable in outline. In August 2024 the company took a strong result in an AI-based arbitrary three-dimensional vehicle configuration drag coefficient prediction competition, an external event with a third-party organiser. The placement number is not consistently reported across sources, so the result is cited as strong rather than as a specific rank.

The physical-AI framing is appropriate for the aerodynamic use case and thin elsewhere. The economic argument for the customer is that in vehicle aerodynamics, where the drag coefficient moves in hundredths and the design loop is run against manufacturing deadlines, reducing an optimisation cycle from seven person-days to half a person-day directly increases the number of design iterations affordable. The company's evidence that it can do this rests on its own case studies, and the more aggressive claim, that the tool can predict drag coefficient for arbitrary three-dimensional configurations, is the one that would need the most careful validation before a customer relies on it.

Product and Commercial Model

The commercial model is direct and largely enterprise. The company sells licences and project deployments to manufacturers, testing bodies, and research institutions, and it is moving toward bundling through TF-FEAST, which is positioned as a way to shorten the adoption curve.

No revenue figure is public. The only quantitative business information in the file is the headcount of 128 in the 2024 annual report, the patent and software copyright counts, and the customer and partnership counts. There is no disclosed revenue, gross margin, backlog, customer concentration, or renewal rate, and there is no indication of whether the business is profitable.

The customer and partnership counts are the most usable commercial indicators. More than ten enterprise customers, more than twenty chain-leader strategic partnerships, and more than fifteen regional industrialisation bases. The nature of the ten-plus enterprise customers is not disclosed individually, with the automotive accounts appearing through case material rather than a customer list.

The institutional channel is a second commercial route. The postdoctoral practice base, the provincial engineering research centre, and the 2026 national case selection for domestically independent industrial simulation software all convert into procurement visibility, and the latter in particular gives the company a referenceable credential in a market where procurement departments have begun to ask about domestic alternatives.

The humanoid robotics activity, although it sits in a subsidiary, is a real commercial adjacency. Jingfeng Technology with the Beijing Humanoid Robot Innovation Center released Horizon of Tomorrow, and the Tian Gong Jingfeng entry ran 100 metres in 9.34 seconds at the 2026 World Humanoid Robot Games. A simulation company that can supply the fluid and structural tools to a humanoid programme has a plausible route into a market whose budgets are currently being absorbed by a small number of well-funded model companies.

The pricing structure is not disclosed. No list prices, no licence model description, no subscription terms, no maintenance or support rates. Any revenue model assessment therefore has to be inferred from the product architecture rather than observed.

Competitive Landscape

The competitive set for TenFong is the Chinese domestic industrial simulation software segment, which has a specific structure: international incumbents with full product suites on one side, domestic challengers with point solutions on the other, and a customer base that is increasingly required to consider domestic alternatives for procurement reasons.

Against international incumbents, the company's position rests on two things. The first is price and procurement accessibility, which is real but not a technology argument. The second is the lattice Boltzmann and turbulence lineage of its scientific leadership, which gives it a specific area of genuine depth rather than a general claim of capability.

Against domestic challengers, the differentiator is breadth combined with an academic founder, which is unusual in this cohort. Most domestic simulation startups are product-led with a commercial or academic but non-founder scientific advisor; here the scientific founder is directly the source of the core method. The weakness of the academic advantage is the one that comes with it: an organisation built around a distinguished academic research group is not obviously the structure that commercialises quickly, and the release cadence suggests the platform layer is still being assembled.

Against the physics AI and model-first cohort, the comparison is sharper than for most domestic simulation vendors because the AI layer overlaps directly. A company selling a learned operator that predicts a flow field from geometry is competing with model-first vendors on the same axis, and it does so with the significant advantage that its operator is embedded in a solver product line with an installed workflow and a customer base. It has the disadvantage that its AI results are not published to the standard the field's conference venues require, which is a reputational cost in exactly the community that would validate its claims.

The automotive and humanoid robotics direction is where the most attractive optionality sits. Chinese original equipment manufacturers running their own aerodynamic development, and humanoid robot programmes that need coupled simulation, are both large and both have an increasing willingness to buy domestic tooling.

The structural risk is the same one that faces every domestic simulation vendor: the domestic software policy environment can pull in both directions. It creates procurement demand that would not otherwise exist, and it can also create customer behaviour where localisation is pursued to satisfy a process rather than to improve a result, which produces revenue that does not renew.

Financing

The official financing record consists of three rounds. An A round of over CNY 100 million in December 2020 from Hillhouse, Green Pine and Nanshan; an A-plus round of several hundred million in December 2021 from China Development Bank Capital, Knowledge City and the Shenzhen angel fund; and a B round of several hundred million in May 2023 from China Development Bank Manufacturing and other investors. Only the first round has a specific order of magnitude, and only in the form of over one hundred million.

The database record consists of a longer sequence with different labels. An angel round in April 2021 from Xietong Innovation; a pre-A round in May 2021 from Nanshan, Jingcheng and Hillhouse; an A round in August 2021 from Nanshan, Green Pine and Linear; an A-plus round in March 2022 of several hundred million from a nine-firm syndicate comprising China Development Bank Capital, Hillhouse, Basis, Green Pine, Haipeng, Renzhi and others; a B round in December 2023 from Jinglin, China Development Bank Manufacturing, Hangshi and the Wuchuang Institute; and a B-plus round on 27 March 2024 from Xietong Investment, amount undisclosed. The final round falls after the official B round, which is the clearest evidence that the two sequences cannot be aligned label by label.

A third account exists at a completely different magnitude. One Chinese industry outlet recorded three rounds totalling CNY 75 million. This is roughly an order of magnitude below the official figure for the A round alone, and it is reported here as a media account rather than reconciled.

The investor list is itself the most useful output of the financing record. It contains both market investors (Hillhouse, Green Pine, Linear, Jinglin, Basis, Haipeng, Renzhi) and state-linked capital (China Development Bank Capital, China Development Bank Manufacturing, the Shenzhen angel fund, Knowledge City, Nanshan, Hangshi, the Wuchuang Institute). The presence of two separate China Development Bank vehicles is a meaningful signal of policy alignment, and the presence of Linear, which is known for industrial software and manufacturing focus, is a meaningful signal of sector fit.

No valuation is disclosed for any round, and no financing figure exists for the most recent B-plus round. This report therefore states a financing history in two parallel columns and does not produce a cumulative total.

DISCREPANCIES

1. Round naming and round count. Convention A, the company's own financing materials: A round (December 2020), A-plus round (December 2021), B round (May 2023), three rounds in total. Convention B, investment databases: angel (April 2021), pre-A (May 2021), A (August 2021), A-plus (March 2022), B (December 2023), B-plus (27 March 2024), six or more rounds. Ruling: both are reported in parallel columns and are not merged. A is preferred for the funding narrative because it is the company's own account. B is preferred for the investor list and for the B-plus round, which appears only in B. Neither count is presented as the number of rounds the company has completed.

2. Series A timing and size. Convention A: December 2020, over CNY 100 million (OFFICIAL company materials). Convention B: April 2021 angel, May 2021 pre-A and August 2021 A (DB). Ruling: the December 2020 event and the 2021 sequence are presented as separate records and the question of whether they are the same money is stated as unresolved. The database has no entry for December 2020, and the company has no entry for the 2021 angel and pre-A rounds.

3. Cumulative financing. Convention A: over CNY 100 million plus two rounds of several hundred million each, so a minimum in the order of CNY 700 million (OFFICIAL, derived from stated magnitudes). Convention B: three rounds totalling CNY 75 million (MEDIA). Convention C: database records that state amounts for the A-plus round of several hundred million in March 2022 without a total. Ruling: no single cumulative figure is given. A and B differ by roughly an order of magnitude and no source reconciles them; the minimum in A is a derived floor from the phrase several hundred million and is not a disclosed total. This report states that the company has raised at least over CNY 100 million with two further rounds of several hundred million each, and separately that one outlet recorded three rounds totalling CNY 75 million.

4. Series B-plus round. Convention A: the most recent official round is the B round of May 2023 (OFFICIAL). Convention B: a B-plus round on 27 March 2024 from Xietong Investment with the amount undisclosed (DB). Ruling: both are reported, and the existence of a post-official-round event is stated plainly. The phase field for this report follows the official naming, which places the company at Series B, while the 2024 B-plus event is recorded as a further undisclosed round.

5. Investor attribution for the first institutional round. Convention A: Hillhouse, Green Pine and Nanshan in the A round (OFFICIAL). Convention B: Nanshan, Jingcheng and Hillhouse in the pre-A round of May 2021 (DB). Ruling: the investor lists are given as separate records. The overlap between Hillhouse and Nanshan suggests related activity, but no source describes a restructuring, and the report does not assert one.

6. Vehicle aerodynamic case figures. Convention A: person-days reduced from seven to half a day, drag coefficient deviation below two percent, wind resistance reduced by more than five counts (company case material). Convention B: mean absolute percentage error of 1.16%, trend accuracy of 94.34%, wind resistance reduced by four to ten percent with 7.43% for one sport utility vehicle, normalised root mean square error below five percent across five physical quantities (company disclosure). Ruling: both sets are reported as separate metrics with their own definitions. They measure different things, and the report does not merge a person-day saving with an error percentage into a single accuracy claim.

7. Hardware claim. Convention A: a single 12GB GPU supports a ten-million-cell mesh and an eight-million-node case runs in seconds on one RTX 4070 (company statement). Convention B: no published model, mesh, baseline solver, or hardware specification accompanies either claim. Ruling: the claims are reported as vendor statements with the absence of the test matrix stated explicitly. This is not a case of two conflicting measurements but of one claim whose verification basis is absent.

8. Employee count. Convention A: 128 in the 2024 annual report, with more than seventy percent postgraduate (company disclosure). Convention B: corporate registry and aggregator records that do not carry a comparable headcount figure. Ruling: 128 with the 2024 label is used. No source in the file provides a conflicting headcount, and the registry information does not include one.

9. DragSolver result. Convention A: a relative error of 0.0005 and an R-squared of 0.97, presented at ICML 2025 (company and conference claim). Convention B: the underlying paper, dataset, and evaluation protocol are not available in the sources consulted. Ruling: the figures are reported with the conference venue named and the verification status explicitly stated as unconfirmed. The report treats the numbers as a track-record claim rather than a published result.

10. Humanoid robotics attribution. Convention A: the 100 metres in 9.34 seconds result belongs to the Tian Gong Jingfeng entry developed by the subsidiary Jingfeng Technology with the Beijing Humanoid Robot Innovation Center (company disclosure). Convention B: no source attributes the result to TenFong's own product line. Ruling: the result is reported as a subsidiary and collaboration achievement, and is not counted as a TenFong product capability. This distinction is maintained deliberately, because it is exactly the kind of association that a reader would otherwise carry into a competitive assessment.

Risks and Due Diligence Gaps

Revenue and profitability are entirely undisclosed. There is no revenue, gross margin, backlog, burn rate, or customer concentration figure anywhere in the file. For a company that has raised at least in the order of CNY 700 million according to its own account, this is the largest gap. Due diligence action: audited or reviewed financial statements, revenue split by product line and by customer, and the ageing profile of receivables.

The financing record cannot be reconciled and no cumulative total can be stated. The official document and the investment database disagree on the number of rounds, the labels, and in one account the total by an order of magnitude. Any counterparty relying on a single data aggregator will therefore be working with a figure that at least one credible record contradicts.

Every performance number lacks a test matrix. The ten-million-cell mesh on a 12GB GPU, the ten-second target, the error below five percent, the 1.16% mean absolute percentage error, the 94.34% trend accuracy, the sub-five-percent normalised root mean square error across five quantities, and the eight-million-node RTX 4070 result all lack the model, mesh, baseline solver, hardware specification, and metric definition needed to check them. Due diligence action: request the complete test matrix for each claim and reproduce at least one case independently.

The DragSolver result is the most quotable and the least verifiable. A relative error of 0.0005 and an R-squared of 0.97 presented at a top-tier machine learning conference is unusually strong for an industrial claim, and it is exactly the kind of number that a customer will treat as verified once it has been seen. The underlying paper and evaluation protocol were not available in the sources consulted, so its review status could not be confirmed here.

The academic founder is both the asset and the concentration risk. Chen Shiyi's scientific authority is the source of the lattice Boltzmann and turbulence depth, and it is also the reason the company is investable at all. No source clarifies the governance terms of his advisory role, whether his institution retains intellectual property, or how the company's core method relates to the academic lineage from which it came.

The scale mismatch with international incumbents is real. Siemens, Ansys, Dassault Systemes and Cadence all sell simulation products with revenue in the billions and full enterprise footprints. TenFong has 128 employees and a national footprint of fifteen industrialisation bases. The commercial question is not whether the technology works but whether the support organisation can serve a national OEM customer at the standard a multinational expects.

The subsidiary and robotics adjacency carries execution risk that is not visible. Jingfeng Technology's humanoid activity is a plausible route into a large market, and a 9.34 second 100 metres result is a strong reference for a simulation supplier. But nothing in the public record describes the revenue, the contract, the ownership split, or the degree to which the simulation tools were actually used rather than the entry being primarily hardware.

The domestic policy tailwind cuts both ways. Localisation-driven procurement creates demand that would not otherwise exist, and it also creates customers who buy to satisfy a process. The second category does not renew, and the public record contains no renewal or retention data with which to separate the two.

There is no exit route and no valuation anchor. No valuation is disclosed for any round, the most recent round amount is undisclosed, and the company is private with no listing timetable. Any assessment of current value rests on assumptions about the last round price rather than on a documented figure.

Overall Assessment

• Technology 7.5. A classical solver portfolio spanning fluid, structural, multidisciplinary, and industry-specific products with breadth that raises switching costs; three self-developed neural operators including a lattice Boltzmann lineage that traces to a founder academician; a reported ICML 2025 result with a 0.0005 relative error and 0.97 R-squared; verified integration with NVIDIA PhysicsNeMo, AmgX and OptiX; and a mesh and grid product, ZF-Grid V2026, that addresses a real production bottleneck. Deducted points: no published test matrix behind any performance number, the DragSolver verification status unconfirmed, no published method for the AI layer's training data, and the platform layer still visibly under assembly across a concentrated September and October 2026 release cadence.
• Commercialisation 6.0. A real enterprise customer base of more than ten named-count organisations, more than twenty chain-leader strategic partnerships, more than fifteen regional industrialisation bases, recognisable institutional credentials including a provincial engineering research centre and a 2026 national domestic-software case selection, and a bundled suite strategy aimed at shortening adoption. Deducted points: zero disclosed revenue, no gross margin, no backlog, no customer concentration, no renewal data, no pricing structure, and a customer base whose individual names are not published.
• Financing 6.5. A documentable institutional shareholder base that includes both market investors and two China Development Bank vehicles, a policy-aligned investor list, a final round in March 2024, and an implied minimum cumulative raise in the order of CNY 700 million on the company's own account. Deducted points: an irreconcilable conflict between the official record and the database record, a media account an order of magnitude lower, no disclosed amount for the most recent round, no disclosed valuation at any round, and no cumulative total this report is willing to state.
• Product maturity 7.5. Nearly twenty shipped products across three physics families plus platform, grid, and industry solutions, with named releases in September and October 2026 and a bundled suite for rapid adoption, plus a genuine deployment reference in the humanoid robotics programme. Deducted points: the newest platform layer has no reference customers disclosed, the AI layer's accuracy claims are unverifiable from public sources, and no release notes, user documentation, or version history is publicly available for inspection.
• Market position 6.5. A differentiated position built on an academician chief scientist and lattice Boltzmann depth that few domestic vendors can match, full-suite breadth rather than point solutions, verified NVIDIA ecosystem integration, and credible adjacency to Chinese automotive aerodynamic development and to humanoid robotics simulation budgets. Deducted points: no third-party market share ranking, no evidence of international customer adoption, a structural resource gap against international incumbents that support organisations can only narrow over years, and the same localisation-policy double edge that applies to the domestic cohort.

References

1. TenFong corporate registry record, unified social credit code 91440300MA5GJ4M99E, including incorporation on 17 December 2020, registered capital of CNY 16.729421 million, legal representative Zhang Rikui and the Nanshan District registered address
2. TenFong company and product materials on TF-QFLUX, TF-Lattice, TF-CFlow, TF-SPH, TF-Struct, TF-Dyna, TF-DCAMS, TF-eMag, TF-Acoustics, TF-DEM, TF-Thermal, TF-SimFARM and TF-SimCITY, including the release dates for TF-AIDEA on 20 September 2026, ZF-Grid V2026 on 25 September 2026 and the TF-FEAST upgrade on 13 October 2026
3. TenFong materials on SATO, GeoFormer and DragSolver, including the dual-attention description, the CAD-to-flow-field mapping function, and the ICML 2025 presentation with relative error 0.0005 and R-squared 0.97
4. TenFong 2024 annual report, including 128 employees with more than seventy percent postgraduate, more than thirty software copyrights, more than three hundred granted patents, more than ten enterprise customers, more than twenty chain-leader strategic partnerships and more than fifteen regional industrialisation bases
5. TenFong official financing materials and timeline, containing the December 2020 A round of over CNY 100 million, the December 2021 A-plus round of several hundred million and the May 2023 B round of several hundred million
6. Investment database records for the angel round of April 2021, the pre-A round of May 2021, the A round of August 2021, the A-plus round of March 2022, the B round of December 2023 and the B-plus round of 27 March 2024, including the investor lists and the undisclosed amount of the final round
7. Chinese industry press coverage of TenFong financing recording three rounds totalling CNY 75 million
8. TenFong institutional credentials and announcements, including national high-technology enterprise status, CMMI level three, specialised-and-sophisticated small giant status, the Shenzhen postdoctoral practice base of January 2025, the Guangdong engineering research centre of July 2025, and the 2026 domestic industrial simulation software case selection
9. TenFong vehicle aerodynamic case material, including the seven person-day to half person-day optimisation cycle, drag coefficient deviation below two percent, wind resistance reduced by more than five counts, the 1.16% mean absolute percentage error, the 94.34% trend accuracy, the 7.43% sport utility vehicle figure and the sub-five-percent normalised root mean square error across five quantities
10. NVIDIA PhysicsNeMo and NVIDIA case study material on TenFong, covering DoMINO, Transolver, GeoTransolver, AmgX acceleration and OptiX in-situ visualisation
11. Announcements of the joint innovation laboratory for automotive fluid simulation software with FAW, the collaboration with Ningbo Eastern Institute of Technology, and the TF-SPH Auto V1.0 release
12. Jingfeng Technology and Beijing Humanoid Robot Innovation Center material on Horizon of Tomorrow and the Tian Gong Jingfeng 100 metres in 9.34 seconds result at the 2026 World Humanoid Robot Games
13. TenFong materials on the August 2024 AI-based arbitrary three-dimensional vehicle configuration drag coefficient prediction competition, reported as a strong result without a consistent placement number across sources
14. Chinese Academy of Sciences and Third World Academy of Sciences records for Chen Shiyi, including the turbulence and lattice Boltzmann research lineage

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of October 23, 2026; please refer to official company announcements and filings.