Tianfu (Nanjing Tianfu Software) Deep Dive: A 2011-founded Chinese Intelligent-CAE Veteran That Closed a Several-Hundred-Million-RMB Series C, Pivoted to an Industrial World Model in 2024, With Unreconcilable Financing and Valuation Records and No Disclosed Revenue

Executive Summary

Nanjing Tianfu Software Co., Ltd. (Tianfu) is one of the oldest surviving Chinese intelligent-CAE vendors: incorporated on 20 May 2011 and headquartered in Nanjing, it has spent more than a decade building a self-developed stack of simulation, optimisation and industrial-AI software under an explicit "fully autonomous code" doctrine. Its defining commercial asset is breadth plus longevity rather than a single breakthrough model. The product line spans general-purpose tools (AICFD for intelligent thermal-fluid simulation, AIFEM for intelligent structural simulation, AIPOD for intelligent optimisation, DTEmpower for intelligent data modelling, AITOPT for heat-transfer topology optimisation, and the TFIIF industrial-AI base) and a long list of industry-specific tools (AIPIPE, AIPump, AIFan, AIBlade, AIWind, EnvCloud, AIICS intelligent monitoring, AIPHM predictive maintenance, AITurbo turbomachinery). The company reports having served more than 400 customers since inception.

The most distinctive recent move is the Industrial World Model narrative. From 2024 onward Tianfu has reframed itself as a "provider of an Industrial World Model framework and infrastructure", positioning its TFIIF base, its simulation solvers, mechanism models and real-time plant data as a single layer that can not only mirror a physical asset (the digital-twin function) but also understand physical laws, run multiple possible futures, and assist operating decisions. The architecture it publishes decomposes into four cooperating parts: a large language model as commander, domain-specific small models for precise calculation, mechanism models that enforce physical constraints, and an industrial agent that organises execution. A deployed reference is the Shandong smart-boiler-island platform, which visualises in three dimensions the furnace temperature, flue-gas composition, wall heat load and pollutant distribution; a Zhuhai power-plant project is reported to have built more than 350 operating-condition models, 130 indicator models, 1050 threshold models and 210 parameter models.

Every performance number in the public record is a vendor statement without a published test matrix. AICFD is reported to predict a four-cooling-tower wind field in three minutes versus eight hours for a single CFD run; ship-resistance models are reported within 3% of experiment; an electronics-conjugate-heat-transfer case is reported with component-average-temperature deviations below 3 degrees Celsius versus a commercial code; a battery liquid-cooling case is reported within 0.1 degree Celsius of a foreign code; and a ship-hull optimisation is reported to cut resistance by 3%, saving close to ten million RMB per year in fuel. None of these has a published mesh, baseline solver, hardware specification or error definition behind it, and the company has not published peer-reviewed accuracy results comparable to those of some newer domestic peers.

The financing record cannot be reconciled and no revenue is disclosed. The company completed a Series C of several hundred million RMB in July 2022 led by GGV Capital and Legend Capital with follow-on from Yunqi Partners, after a Series B of several hundred million in 2021 and earlier angel and A rounds. Investment databases disagree on the angel and A-round amounts (one records an angel round of RMB 2 million from Dongpingheng Capital; another records an undisclosed angel from Jiangning Science-and-Technology Venture Capital; one records an A round of RMB 15 million from Lenovo; another records an undisclosed A round from Lenovo Capital), and one database attaches a stale RMB 150 million "market value / valuation" that is inconsistent with a post-Series-C several-hundred-million scale. The company has never disclosed revenue, gross margin, backlog, customer concentration or renewal rate.

The physical-AI relevance of Tianfu is genuine but mostly at the surrogate-and-assist level, not at the foundation-model level. Its AI layer accelerates and automates classical CAE (AI mesh refinement, AI-accelerated solving, AI reduced-order prediction, LLM Q&A assistants), and its Industrial World Model is an architecture that wires large, small and mechanism models together rather than a single published physical foundation model. It is therefore best read as an AI-accelerated CAE and industrial-digital-twin vendor that has adopted the world-model vocabulary, rather than as a physical-foundation-model company.

Company Overview

The entity is Nanjing Tianfu Software Co., Ltd., incorporated on 20 May 2011, unified social credit code 913201155715955464, registered at 19 Suyuan Avenue, Jiangning Economic and Technological Development Zone, Nanjing. The legal representative and founder is Zhang Ming, a PhD who previously served as technical director of ANSYS Japan and general manager of Blue Ridge Numerics Japan, a background that, together with a founding team of multiple Tsinghua PhDs, is the source of the company's simulation credibility. The company states that it has maintained fully autonomous core code and self-developed core solvers since inception.

The product organisation is broad and has shipped continuously for over a decade. General-purpose products are AICFD (intelligent thermal-fluid simulation, released in 2021), AIFEM (intelligent structural simulation), AIPOD (intelligent optimisation design), DTEmpower (intelligent data modelling) and AITOPT (intelligent heat-transfer topology optimisation). The industrial-AI base TFIIF integrates "simulation + AI + mechanism". Industry-specific products cover turbomachinery (AITurbo, AIPump, AIFan, AIBlade), piping (AIPIPE), wind farms (AIWind), environment (EnvCloud), and operations and maintenance (AIICS intelligent monitoring, AIPHM predictive maintenance).

Scale and credential disclosures are specific but narrow. The company reports serving more than 400 customers since inception, with customers across energy and power, shipbuilding and maritime, vehicles and transportation, aerospace, new-energy vehicles, power batteries, consumer electronics, and petroleum and petrochemicals. Milestones include national high-tech enterprise status in 2016, Jiangsu provincial specialised-and-sophisticated SME status in 2023, national little-giant SME status in 2024, and Nanjing municipal enterprise technology centre status in 2024. One database records headcount in the 51 to 100 range, which is the only size figure in the file and is not reconciled against the 400-plus-customer claim.

No revenue, gross margin, burn rate, valuation or listing timetable is public. The most recent disclosed financing event is the July 2022 Series C, whose amount is stated only as "several hundred million RMB".

Common Misreadings: Six Errors That Change the Analysis

The Series C is not the whole financing history. The company's own announcements describe angel, A, B and C rounds, but investment databases add a 2019 equity round and disagree on the angel and A amounts. Quoting only the company's narrative or only a database produces an incomplete or contradictory answer.

"Several hundred million RMB" is not a number. The B and C rounds are each described only as "several hundred million" or "over one hundred million". No source states an exact figure for either, so any cumulative total derived from them is a floor, not a disclosed amount.

The Industrial World Model is an architecture, not a published foundation model. Tianfu's world-model material describes a framework of large, small and mechanism models with agents; it does not publish a single trained physical foundation model, weights, or benchmark. Reading it as a physical-foundation-model company is a category error.

Digital twin and Industrial World Model are not the same claim, and Tianfu does both. The company states plainly that the world model sits on top of digital-twin capability and adds understanding, future simulation and decision support. Conflating the two hides the fact that the digital-twin deployments (boiler island, power plant) are the more concrete and verifiable asset.

The 51 to 100 headcount figure does not settle the scale question. A 51 to 100 employee count alongside a 400-plus-customer, multi-product, multi-province operation and a national little-giant designation is internally striking, and the company has not published revenue to resolve it. The figure should be treated as one database's estimate, not as a measured fact.

Tianfu is not primarily an LLM company. Its 2025R1 release adds an LLM-based Q&A and modelling assistant to AICFD, and its world-model architecture uses an LLM as commander, but the core product is classical CAE with AI acceleration. Reading it against the foundation-model or agent-only category mis-maps the business.

Technology

The technical story has three layers, and they have different maturity and different verification status.

Layer one is the classical solver and product portfolio, and it is what the company actually sells. AICFD covers incompressible and compressible flow (the 2025R1 release materially improved compressible and hypersonic capability), heat transfer, multiphase flow, noise and combustion; AIFEM covers structural finite-element analysis on a self-developed solver; AIPOD covers design optimisation with interfaces to more than 40 domestic and foreign CAD/CAE tools and more than 70 self-developed AI and optimisation algorithms; DTEmpower covers data modelling and mechanism-model fusion; AITOPT covers heat-transfer topology optimisation. This breadth is the most defensible part of the offer because it is the part a customer must replace to switch away.

Layer two is the AI-acceleration layer inside the solvers. AICFD 2025R1 introduces AI mesh (flow-solution-driven adaptive refinement with geometry-conforming enforcement), AI-accelerated solving, AI reduced-order prediction, and an LLM-based intelligent assistant for Q&A and modelling. AIPOD's multimodal surrogate models cover parameter, contour and time-series prediction. The company's long-standing claim is that it was the first in the industry to adopt an AI-accelerated algorithm system to speed up fluid and structural simulation, and that its "AI plus simulation" approach speeds up classical CFD analysis by several times.

Layer three is the Industrial World Model and the TFIIF base. The published architecture has four cooperating parts: a large model as commander that understands natural language, decomposes tasks and dispatches tools; industrial small models trained on specific equipment, processes and tasks for precise calculation; mechanism models that enforce physical laws and safety boundaries; and an industrial agent that organises execution. The company states that useful industrial intelligence is a library of many narrow specialist models rather than one super-model, and cites a Zhuhai power-plant project with 350-plus condition models, 130-plus indicator models, 1050-plus threshold models and 210-plus parameter models.

The performance claims are uniform in character and uniform in weakness: they are all vendor statements without a published test matrix. Ship-resistance error below 3% versus experiment, electronics component-average-temperature deviation below 3 degrees Celsius, battery liquid-cooling within 0.1 degree Celsius of a foreign code, cooling-tower wind-field prediction in three minutes versus eight hours, and ship-hull resistance reduced by 3% with close to ten million RMB annual fuel saving. Every one of these lacks the mesh, baseline solver, hardware specification and error definition needed to check it, and no peer-reviewed result is in the public record.

One deployment is verifiable in outline. The Shandong smart-boiler-island platform and the Zhuhai power-plant predictive-maintenance system are described in company and third-party (Nanjing Daily, China Economic News) coverage, with the model counts for Zhuhai stated explicitly. These are the most concrete evidence that the world-model framing has reached production, even though the underlying accuracy is still a vendor claim.

Product and Commercial Model

The commercial model is direct enterprise licensing plus project-based custom development. The company sells licences and deployments to manufacturers, power and energy operators, research institutes and universities, and provides custom development (CAE pre- and post-processing, dedicated solvers, industrial-design platforms, AI platforms). It is also expanding into education, supporting university formula-student racing teams with AICFD, AIFEM and AIPOD.

No revenue figure is public. The only quantitative business information is the more-than-400-customer count, the product and credential list, and one database's 51 to 100 headcount estimate. There is no disclosed revenue, gross margin, backlog, customer concentration, renewal rate or pricing structure.

The customer and vertical footprint is the most usable commercial indicator. Energy and power (boiler island, cooling towers, power plants), shipbuilding and maritime (hull resistance optimisation), automotive and new-energy vehicles (aerodynamics, battery thermal management), aerospace, and consumer electronics. The education channel (university racing teams) is a deliberate pipeline investment for future engineers.

The operations-and-maintenance and world-model route is where recurring revenue would live, if it materialises. AIICS intelligent monitoring and AIPHM predictive maintenance, wired into DCS control systems, are positioned for "few-operator" smart power plants; this is the part of the offer with the most plausible annuity character, but no contract value or renewal data is disclosed.

The pricing structure is not disclosed. No list prices, licence model, subscription terms or support rates are public, so any revenue-model assessment must be inferred from architecture rather than observed.

Competitive Landscape

Tianfu's competitive set is the Chinese domestic industrial-simulation and intelligent-industrial-software segment, which has a specific structure: international incumbents with full suites on one side, domestic challengers with point or broad solutions on the other, and a customer base increasingly required to consider domestic alternatives.

Against international incumbents, Tianfu's position rests on price, procurement accessibility and autonomous controllability, which are real but not technology arguments, plus a long track record and a self-developed-solver claim that gives it a specific credibility in the domestic procurement context.

Against domestic challengers, the differentiators are longevity and breadth: founded in 2011 versus most domestic simulation startups that are far younger, and a product matrix that spans general, industry-specific and operations-and-maintenance tools rather than a single point solution. The weakness is the same one that faces every domestic vendor: international incumbents have billions in revenue and full enterprise footprints, and Tianfu has a 51 to 100 employee estimate and no disclosed revenue.

Against the physics-AI and model-first cohort, the comparison is sharper because Tianfu's AI layer overlaps directly. A company selling AI-accelerated CFD and surrogate models competes on the same axis as model-first vendors, with the advantage that its models sit inside a shipped solver-and-optimisation workflow with an installed customer base, and the disadvantage that its accuracy claims are not published to the conference standard the field expects. The Industrial World Model vocabulary also overlaps with World Labs, Skild and Physical Intelligence narratives, but Tianfu's version is an enterprise-software architecture rather than a research foundation model.

The structural risk is the same double-edged policy tailwind as the domestic cohort. Localisation-driven procurement creates demand that would not otherwise exist and also creates customers who buy to satisfy a process rather than to improve a result, and the second category does not renew.

Financing

The official and database records describe the same company with different round details. A July 2022 Series C of several hundred million RMB led by GGV Capital and Legend Capital with follow-on from Yunqi Partners is the most consistent event across sources. A 2021 Series B of several hundred million (or "over one hundred million") from SDIC Co-investment, Yunqi, Lenovo Capital and others is broadly agreed but differs in exact amount and date between sources. Earlier rounds are where the records diverge.

The angel and A rounds are the largest documentary problem. One database records an angel round of RMB 2 million from Dongpingheng Capital on 25 December 2013 and an A round of RMB 15 million from Lenovo on 26 October 2017; another records an angel round from Jiangning Science-and-Technology Venture Capital (undisclosed) and an A round from Lenovo Capital (undisclosed), with dates shifted to December 2017. A third source adds a December 2019 equity round from SDIC Co-investment, Zhongxing Venture Capital and Nanjing Fengling, undisclosed, that the other databases do not carry. Neither the angel nor the A amount can be stated with confidence.

A stale valuation figure exists and should not be used. One database attaches a "market value / valuation" of RMB 150 million to the company, but that figure is inconsistent with a post-Series-C several-hundred-million scale and is almost certainly a stale or mislabelled field; no other source gives a valuation.

No cumulative total can be stated. The company describes B and C rounds each as "several hundred million" with no exact figure, one database totals disclosed financing at RMB 267 million which is inconsistent with two several-hundred-million rounds, and the angel and A amounts conflict across sources. This report states the investor base and the round sequence qualitatively and does not produce a cumulative total.

DISCREPANCIES

1. Angel round investor and amount. Convention A (investment database): Dongpingheng Capital, RMB 2 million, 25 December 2013. Convention B (venture database): Jiangning Science-and-Technology Venture Capital, undisclosed, December 2013. Ruling: both are reported and neither amount is reliable. The investor names and the amount directly conflict, and no source reconciles them.

2. Series A amount and date. Convention A (database): Lenovo, RMB 15 million, 26 October 2017. Convention B (venture database): Lenovo Capital, undisclosed, December 2017. Ruling: the amount and date are reported as two separate records. The RMB 15 million figure appears in only one source; the other records the round as undisclosed, so the exact A-round size is unresolved.

3. 2019 equity round existence. Convention A (venture database): a December 2019 equity round from SDIC Co-investment, Zhongxing Venture Capital and Nanjing Fengling, undisclosed. Convention B (the financing database that carries the angel and A amounts): no such round. Ruling: the 2019 round is reported as present in one source and absent in another. It is listed here as a possible additional round rather than confirmed, because only one record carries it.

4. Series B amount magnitude and date. Convention A (financing database): several hundred million RMB, 29 June 2021, investors including SDIC Co-investment, Lenovo Capital and Yunqi. Convention B (venture database): over one hundred million RMB, July 2021, with additional investors SDIC Investment and Jucheng Capital not listed by the other. Ruling: both are reported, and the investor list is given as separate records. The magnitude differs (several hundred million versus over one hundred million) and the date differs by a month; no source reconciles them.

5. Valuation figure. Convention A (financing database): a "market value / valuation" of RMB 150 million. Convention B: no valuation disclosed by the company or any other source. Ruling: the RMB 150 million figure is reported as a stale or mislabelled field and is not used as a valuation anchor. It is inconsistent with a post-Series-C several-hundred-million scale.

6. Cumulative financing total. Convention A (financing database): total disclosed financing RMB 267 million. Convention B (company and venture records): B and C rounds each several hundred million RMB, which would imply a total well above RMB 267 million. Ruling: no cumulative figure is given. The RMB 267 million total is inconsistent with two several-hundred-million rounds and is not presented as a reconciled amount.

7. Operating scale tension. Convention A (venture database): headcount in the 51 to 100 range. Convention B (company): more than 400 customers served since inception, a multi-product and multi-province operation, and a national little-giant designation. Ruling: the 51 to 100 figure is reported as one database's estimate and is not reconciled against the 400-plus-customer claim. The company has not published revenue to resolve the tension.

8. Branch and subsidiary count. Convention A (company website): branches or offices in Beijing, Dalian, Ningbo, Shanghai and Qingdao, five locations. Convention B (encyclopedia): six locations including Changzhou. Ruling: both are reported and the count is given as five per the company, six per the encyclopedia, without reconciliation.

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 founded in 2011 with more than 400 customers, 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 angel and A amounts conflict between databases, a 2019 equity round appears in only one source, a stale RMB 150 million valuation figure exists, and the RMB 267 million total is inconsistent with two several-hundred-million rounds. Any counterparty relying on a single aggregator will be working with a figure at least one credible record contradicts.

Every performance number lacks a test matrix. The 3% ship-resistance error, the below-3-degree-Celsius electronics temperature deviation, the 0.1-degree-Celsius battery liquid-cooling match, the three-minutes-versus-eight-hours cooling-tower prediction, and the 3% hull-resistance reduction all lack the mesh, baseline solver, hardware specification and error definition needed to check them. Due diligence action: request the complete test matrix for each claim and reproduce at least one case independently.

The Industrial World Model is an architecture, not a published foundation model. The world-model material describes large, small and mechanism models with agents, but no single trained physical foundation model, weights or benchmark is published, and no peer-reviewed accuracy result is in the public record. The claim should be read as an enterprise-software framework, not as a research-model capability comparable to published physical foundation models.

The long track record is both the asset and the concentration risk. Fifteen years of autonomous-code development give Tianfu a credibility that newer entrants lack, but the public record contains no detail on governance terms for the founders, ownership concentration, or how the core solver relates to the founders' prior international employment. The 42.98% founder ownership recorded in one database is the only equity figure available.

The scale mismatch with international incumbents is real. Siemens, Ansys, Dassault Systemes and Cadence sell simulation products with revenue in the billions and full enterprise footprints; Tianfu has a 51 to 100 employee estimate and a national footprint of five to six branches. 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 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 except a stale RMB 150 million figure, 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 self-developed classical solver and product portfolio spanning thermal-fluid, structural, optimisation, data-modelling, topology-optimisation and an industrial-AI base, with AI-acceleration features (AI mesh, AI-accelerated solving, AI reduced-order prediction, LLM assistant) in the 2025R1 release, and a verifiable-in-outline Industrial World Model deployment in power plants. Deducted points: no peer-reviewed accuracy result in the public record, every performance number lacks a published test matrix, and the world model is an architecture rather than a published foundation model.
• Commercialisation 6.5. More than 400 customers served since inception across energy, shipbuilding, automotive, aerospace and electronics, a real operations-and-maintenance and world-model route with deployed power-plant references, and a deliberate education pipeline. Deducted points: zero disclosed revenue, no gross margin, no backlog, no customer concentration, no renewal data, no pricing structure, and a 51 to 100 headcount estimate that is not reconciled with the customer count.
• Financing 6.5. A documentable institutional shareholder base including GGV Capital, Legend Capital, Yunqi Partners, SDIC vehicles and Lenovo Capital, a clear Series C event in July 2022, and a policy-aligned investor list. Deducted points: irreconcilable angel and A amounts between databases, a 2019 equity round present in only one source, a stale RMB 150 million valuation figure, a RMB 267 million total inconsistent with two several-hundred-million rounds, and no disclosed valuation at any round.
• Product maturity 7.5. A broad, continuously shipped matrix of more than ten products across general, industry-specific and operations-and-maintenance categories, named 2025R1 releases with AI mesh and LLM assistant, and real deployment references in power plants. Deducted points: the AI-layer accuracy claims are unverifiable from public sources, no published test matrix, and no public release notes or version history for inspection.
• Market position 7.0. A differentiated position built on the longest track record among domestic intelligent-CAE vendors, full product breadth rather than a point solution, a verifiable Industrial World Model narrative that aligns with the site's physical-AI theme, and strong vertical depth in power and energy. Deducted points: no third-party market-share ranking, no evidence of international customer adoption, a structural resource gap against international incumbents, and the same localisation-policy double edge that applies to the domestic cohort.

References

1. Nanjing Tianfu Software corporate registry record, unified social credit code 913201155715955464, incorporation on 20 May 2011, registered address 19 Suyuan Avenue, Jiangning Economic and Technological Development Zone, Nanjing, legal representative Zhang Ming
2. Tianfu company and product materials on AICFD, AIFEM, AIPOD, DTEmpower, AITOPT and the TFIIF industrial-AI base, including the AICFD 2025R1 release covering AI mesh, compressible flow, automatic-optimisation interface and LLM assistant
3. Tianfu product case materials, including ship-resistance error below 3% versus experiment, electronics component-average-temperature deviation below 3 degrees Celsius, battery liquid-cooling within 0.1 degree Celsius of a foreign code, cooling-tower wind-field prediction in three minutes versus eight hours, and 3% hull-resistance reduction
4. Tianfu Industrial World Model materials, including the four-part architecture (large model commander, small models, mechanism models, industrial agent), the Shandong smart-boiler-island platform, and the Zhuhai power-plant project with 350-plus condition models, 130-plus indicator models, 1050-plus threshold models and 210-plus parameter models
5. Tianfu company announcement of the July 2022 Series C of several hundred million RMB led by GGV Capital and Legend Capital with follow-on from Yunqi Partners
6. Investment database records for the angel round of RMB 2 million from Dongpingheng Capital on 25 December 2013, the A round of RMB 15 million from Lenovo on 26 October 2017, the B round of several hundred million on 29 June 2021, and the stale RMB 150 million market-value figure
7. Venture database records for the angel round from Jiangning Science-and-Technology Venture Capital (undisclosed, December 2013), the A round from Lenovo Capital (undisclosed, December 2017), the B round of over one hundred million in July 2021, and the December 2019 equity round from SDIC Co-investment, Zhongxing Venture Capital and Nanjing Fengling
8. Tianfu milestone and credential announcements, including national high-tech enterprise status in 2016, Jiangsu provincial specialised-and-sophisticated SME status in 2023, national little-giant SME status in 2024, and Nanjing municipal enterprise technology centre status in 2024
9. Third-party coverage of Tianfu, including Nanjing Daily on Industrial World Model deployment (29 September 2026) and China Economic News on the distinction between digital twin and Industrial World Model
10. Records on founder Zhang Ming's prior roles as technical director of ANSYS Japan and general manager of Blue Ridge Numerics Japan, and the Tsinghua-PhD composition of the founding team

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.