Songying Deep Dive: A Huawei-Cloud-Alumni Startup Building a Domestic Physical AI Operating System, With ORCA OS Positioning Against NVIDIA Omniverse, Three Incompatible Financing Narratives, and a Verifiable Third-Party Anchor in a CNY 20 Million Minority Investment

Executive Summary

Shanghai Songying Technology Co., Ltd. is a Chinese physical AI software company whose product is positioned as a direct answer to NVIDIA Omniverse rather than as an extension of it. The company builds ORCA OS, a physical AI operating system that connects spatial scene construction, real-time multiphysics simulation, synthetic data generation, algorithm development, parallel reinforcement learning training, software-in-the-loop and hardware-in-the-loop verification, and Sim2Real deployment into a single engineering chain. Its founder and chief executive, Nie Kaixuan, spent years at Huawei Cloud, where he was responsible for interactive entertainment and cloud gaming and served as deputy commander for Kunpeng solution delivery. The technical team is drawn largely from Huawei Cloud, Tencent, NetEase, HiSilicon, Intel, and Unity.

The positioning claim is the most consequential thing in the file, and it is a claim rather than a verified comparison. The company describes ORCA OS as the first physical AI operating system for multi-form robot collaborative training, presented at the 2026 World Artificial Intelligence Conference in Shanghai in July. Its subsystem ORCA SIM is described as the first domestically self-developed real-time multiphysics coupling simulation system, fusing rigid body, fluid, and deformable body simulation in one solver. No benchmark against Omniverse, no independent reproduction of the multiphysics claim, and no third-party evaluation of the operating system layer is available in the sources consulted.

The financing record exists in three mutually incompatible forms and cannot be reduced to a single sequence. The company announced in August 2026 that it had completed an A round and an A1 round totalling several hundred million renminbi, led by CICC Capital and Wuhan Hongshan Capital, and described this as its third round within half a year. Corporate registry and database records show a seed round on 5 August 2022 from Qiming Capital and Sequoia Seed Fund, an angel round on 20 March 2025 led by Zhongke Chuangxing, and a separate disclosure in February 2026 of a Pre-A and Pre-A+ round led by Dian Shi Capital and E Fund Private. The company's own account of "three rounds in half a year" cannot be reconciled with the sequence in which the Pre-A and Pre-A+ rounds are dated months before the A round, and the count that emerges from the registry is at least four.

There is one hard, verifiable third-party anchor in the file. On 12 June 2026 Shanghai Fengyuzhu Culture Technology Co., Ltd., listed on the Shanghai Stock Exchange, announced a board resolution to invest no more than CNY 20 million of its own funds in Songying for a minority stake, with the resulting holding percentage explicitly undetermined, and to jointly establish a joint venture company in which Fengyuzhu intends to hold control. This is a disclosed transaction price between an arm's-length listed company and the target, and it is the only number in the file that comes from a regulated filing rather than from the company itself.

The gap between the reported client footprint and the reported headcount is the largest analytical tension in the company. The company states that ORCA has served close to one hundred central and state-owned enterprises, national-level robot innovation centres, the national AI application pilot base, national supercomputing centres, and leading embodied AI enterprises. The registry record for the Shanghai entity shows a registered capital of CNY 1.73035 million and a social insurance headcount of seventeen people as reported in the 2025 annual report. A seventeen-person entity that reports serving close to a hundred state-owned clients is not an impossibility, but it is a specific and unusual commercial structure that requires explanation, because the delivery of on-site scenario building, private deployment, and integrated hardware appliances requires people.

No revenue, gross margin, order book, customer concentration, or valuation is disclosed for any period. The only quantitative business facts available are the client counts, the Fengyuzhu transaction price, the headcount, and a portfolio of institutional credentials.

The company's stated ambition is ecosystem leadership rather than product leadership, and that is a deliberate and coherent strategy. Its founder describes the goal as building the Android of physical AI, not replicating Apple's closed model. The strategy is expressed concretely: a free developer edition that shares the same technical base as the enterprise edition, a ten-million-yuan-plus open source special fund, an open source incubation centre in Wuhan, and free SimReady asset libraries offered alongside a paid enterprise tier.

Company Overview

The entity is Shanghai Songying Technology Co., Ltd., incorporated on 6 December 2024 with unified social credit code 91310112MAE7NGHR77, registered in Minhang District, Shanghai, at room 3 of building B, 940 Jianchuan Road. The legal representative and chairman is Nie Kaixuan. The registry lists Guo Siqiang as financial officer and director, and An Kewei, Huang Wei, and Yuan Bo as directors. The registered capital is recorded as CNY 1.73035 million with paid-in capital of CNY 609,910, and the latest approval date on file is 8 September 2026. Social insurance participation in the 2025 annual report is seventeen people. The registered business scope covers software development, artificial intelligence software development and applications, data processing services, cloud computing equipment services, network and information security software, big data services, digital content production, animation and game development, and import and export of goods and technology.

The group structure has an important wrinkle that is easy to misread. A separate entity, Beijing Songying Technology Co., Ltd., was incorporated in 2021 and is wholly owned by the Shanghai entity. The company itself dates the ORCA programme to 2021, and describes a core technical team assembled in that year, with the Shanghai entity formed in late 2024. The registry therefore shows a 2024 Shanghai company holding a 2021 Beijing company, and both readings appear in company materials as "founded in 2021" and "established in December 2024". Neither is wrong; they describe the programme and the legal entity respectively.

The founder's background is specific and is the company's most credible asset. Nie Kaixuan was at Huawei Cloud in roles including head of interactive entertainment and cloud gaming business and deputy commander of the Kunpeng solution portfolio. The technical team is described as coming from Huawei Cloud, Tencent, NetEase, HiSilicon, Intel, and Unity. This matters more than a generic founder pedigree because the physics simulation and graphics engineering required for a real-time multiphysics operating system is exactly the skill set that cloud gaming and cloud rendering pipelines demand, and that is where the founding team came from.

Institutional credentials are extensive and partly verifiable through third-party announcements. The company lists selection as the first batch of high-quality specialist service providers for AI plus manufacturing in Shanghai in 2025, a role as a unit that has answered a public call for proposals in the new generation AI innovation task programme, and a contract to build a major digital twin training site. It has participated in constructing the national AI application pilot base for intelligent manufacturing and in a Ministry of Industry and Information Technology industrial biomimetic robotics programme. Awards include first prize in the national finals of the China Innovation and Entrepreneurship Competition organised by the Ministry of Industry and Information Technology and the Ministry of Science and Technology, a top-30 placement in the SAIL award at the 2026 World Artificial Intelligence Conference, and first prize at the tenth Maker China and Maker Shanghai competition.

Third-party commercial validation is genuine and is the strongest evidence in the file. Fengyuzhu, a Shanghai Stock Exchange listed company in the cultural technology and spatial computing business, disclosed on 12 June 2026 a board resolution and a framework agreement to make a minority equity investment in Songying of up to CNY 20 million, together with a jointly established joint venture that Fengyuzhu intends to control. The filing explicitly notes the accumulated twelve-month outbound investment ceiling of CNY 220 million, roughly 10.61 percent of audited net assets. The filing also states that Songying's existing shareholders bear no related-party relationship to Fengyuzhu.

There is no public information on revenue, gross margin, burn rate, employee equity, valuation, or customer concentration.

Common Misreadings: Five Errors That Change the Analysis

The company's founding date is not a single date. The ORCA programme began with a core team assembled in 2021; the Shanghai legal entity was incorporated on 6 December 2024. Reading "founded in 2024" as the start of the technical work understates the programme, and reading "founded in 2021" as the incorporation date of the contracting entity overstates it.

The number of financing rounds is not three. The company describes its August 2026 A and A1 rounds as the third round within half a year. The registry and database records show a seed round in August 2022, an angel round in March 2025, a Pre-A and Pre-A+ round announced in February 2026, and the A and A1 rounds in August 2026. Counting by distinct disclosed events gives at least four and, treating the two February rounds separately, five.

"Third round within half a year" and "the Pre-A round is the immediately preceding round" cannot both be true. The February 2026 Pre-A disclosure is roughly six months before the August 2026 announcement, and the company characterises the A and A1 rounds as the third within that half year. The two statements are both present in company-sourced material.

ORCA OS is not a robot, and the company is not a foundation model company. The product is an operating system and simulation stack for training and validating robots. There is no disclosed robot body, no disclosed large language model, and no disclosed general-purpose embodied model. Mapping it onto the model-company category misreads what is being sold.

"Serving close to a hundred clients" is a coverage statement, not a revenue statement. It counts relationships including national innovation centres, pilot bases, and supercomputing centres, which are typically joint research and platform relationships rather than paid software contracts. It is also inconsistent with a seventeen-person social insurance headcount unless most delivery is partner-delivered.

Technology

The technical story is best described as a layered stack, and the layering is the substance of the company's competitive argument.

The lowest layer is the heterogeneous compute framework. It includes ORCA ENGINE for real-time ray tracing and physically based rendering, ORCA SIM for real-time physics simulation, and ORCA DataLink for real-time data, together with an ATEEM heterogeneous computing cluster. The design principle is hardware decoupling: the software is adapted to international mainstream GPUs and to domestic vendors including Moore Threads, MetaX, Iluvatar CoreX, Enflame's domestic counterpart Horizoncore, Cambricon-adjacent vendors, and Biren-adjacent vendors, with AMD and Intel also listed. The domestic adaptation list is the part of this that has independent strategic value, because it is the part that does not depend on a single foreign vendor's roadmap.

The physics layer is where the differentiation claim sits. ORCA SIM is described by the company as the first domestically self-developed real-time multiphysics coupling simulation system, solving rigid body, fluid, and deformable body interaction in a unified manner. The hard part of that claim is not rigid body physics, which is well trodden, but real-time deformable body and fluid interaction at training throughput, since the whole economic argument for simulation-based training depends on how many robot-hours per wall-clock-hour the platform can produce. No independent measurement of that throughput is available in the sources consulted.

The middle layers are training and validation. ORCA Gym is the reinforcement learning framework, ORCA Replicator is the multi-dimensional data synthesis system, ORCA Kit is the development kit and interfaces, and Orca Connect is the software connector. Verification runs through a dual loop: software-in-the-loop and hardware-in-the-loop. This combination is the part of the stack that most closely resembles what an industrial customer actually needs, because the failure mode in physical AI is rarely the policy and frequently the gap between the simulator and the machine.

The top layers are the developer and business surface. ORCA Lab is the developer platform, Orca Store is the asset store, Orca Agent is the AI scene generation model, Orca Playground is the embodied robot simulation example set, and ORCA Studio is the enterprise edition. A library of more than one hundred thousand high-precision SimReady simulation assets is offered. The developer edition was released in March 2026 jointly with leading robot platforms including Unitree, Fourier, Beijing Humanoid Innovation Center, Norinco 58 Intelligent, DEEP Robotics, and Lingchu Intelligent, and is free for education and individual developers.

The data strategy is the single most quotable technical idea in the company, and it is a ratio rather than a result. The one-to-eight-to-one hybrid strategy allocates ten percent of training data to expert demonstration, eighty percent to simulation synthesis, and ten percent to real-world fine-tuning. The founder is explicit that simulation supplements rather than replaces reality, and the corresponding engineering statement is that a single real-world industrial part can be reconstructed in simulation with its mass, friction, and material properties, then varied in pose, orientation, illumination, and occlusion to generate practice tasks. The ratio is a policy decision, and no published measurement of what accuracy is achieved at that mixture is available.

Parallel training scale is claimed but not measured. Company materials state support for concurrent training of ten thousand to one hundred thousand robots, distributed multi-node architectures, and multi-sensor support across RGB, depth, IMU, and lidar, with a unified openUSD dataset format and compatibility with mainstream reinforcement learning platforms. These are capability statements about scale rather than throughput measurements, and the distinction matters for a customer computing unit economics.

The ORCA OS launch at the 2026 World Artificial Intelligence Conference is the single most important product event in the file. In July 2026 the company presented ORCA OS as the first physical AI operating system for multi-form robot collaborative training, demonstrating humanoid, quadruped, wheeled, and unmanned aerial robots sharing one digital factory scene, with humanoid robots performing picking and sorting, wheeled robots relaying transport, quadrupeds conducting inspection, and drones supplementing perception from the air, and with the system re-planning tasks and paths in response to environment change. The presence of the director general of the Ministry of Industry and Information Technology's small and medium enterprise administration and the district head of Minhang at the booth indicates institutional attention rather than independent technical validation.

Product and Commercial Model

The commercial model is a three-tier software business with a deliberate free tier at the bottom.

The paid enterprise tier is ORCA Studio, sold on a subscription model priced by user count and subscription term. The company states that the enterprise edition provides full-process control of simulation and training and is intended for organisations that operate their own research infrastructure. No list prices, no renewal rates, and no revenue split between subscription, private deployment, and services are disclosed.

Physical AI simulation appliances are a hardware-coupled delivery channel. An integrated system pre-installs ORCA 2.0 with a stable runtime environment, positioned for laboratories and early-stage companies that want a ready-to-run deployment rather than a software licence. This is a meaningful hedge for a Chinese customer worried about foreign supply chain interruption, and it also converts some of the economics from pure software into systems integration.

Virtual training ground construction and embodied AI synthetic data services are the two service lines. Scene construction covers high-fidelity factory, production line, warehouse, and logistics environments for smart manufacturing and logistics; synthetic data services generate customised robot manipulation simulation data for operation, locomotion, and navigation tasks. These are people-intensive and therefore the lines most in tension with a seventeen-person headcount.

The declared vertical scope is broad: smart manufacturing, low-altitude economy, eldercare services, logistics and warehousing, and commercial services. The electricity sector has a concrete instance. Together with China Southern Power Grid, the company launched the 2026 Electricity Embodied Intelligence Application Competition, with ORCA as the designated simulation platform for the online simulation event, using substation operation and equipment maintenance scenarios to assess autonomous perception, decision making, and dexterous manipulation under generalisation.

The reference partnerships with large industrial enterprises are concrete and verifiable through announcements. In July 2026 at the World Artificial Intelligence Conference, Songying signed a strategic cooperation agreement with Siemens and entered the Siemens Xcelerator ecosystem, covering physical simulation, industrial digital twins, and embodied intelligence for advanced manufacturing. In the same month it signed a strategic cooperation with the National Robot Testing and Evaluation Center (Headquarters) around simulation dataset CR certification, a joint simulation training and testing platform, and embodied dataset platform construction. It has built an embodied intelligence computing joint laboratory with the National Supercomputing Center in Wuxi, has brought ORCA online at the national embodied intelligence application pilot base in Hangzhou, and in September 2026 signed a strategic cooperation with Yu Jia Zhishu Technology (Shanghai) covering the fusion of industrial multi-source heterogeneous data with the ORCA OS.

The open source track is a deliberate commercial strategy rather than an ideological position. In June 2026 the company donated CNY 300,000 to the Tiangong Kaiwu Open Source Foundation to establish the first embodied intelligence open source special fund in the country, supporting talent development, open source tooling, and data research, and established an open source incubation centre in Wuhan. The logic is stated plainly by the founder: free tools, public assets, and customer scenes are the entry point, and paid services arrive when a scheme needs deeper technical support, robot adaptation, and on-site delivery.

Competitive Landscape

The competitive question for Songying is not whether it can match Omniverse feature by feature, which is not the plan, but whether the position it has taken is defensible.

Against NVIDIA Omniverse, the argument is ecosystem and supply chain rather than capability. The company's own framing is that global physical AI currently depends heavily on an Omniverse-centred software ecosystem, and that Huang Renxun described Omniverse as the operating system of the physical AI era and the successor to CUDA, while a single-vendor dependency concentrates the application layer. Songying's counter-position is openness plus domestic hardware adaptation. This is a real and politically durable argument in the Chinese market, and it is also an argument that does not require Songying to be better than NVIDIA at rendering or physics.

The distinction between a company with a differentiated product and a company with a differentiated position matters here. ORCA's strongest verifiable features are the domestic GPU adaptation breadth, the real-time multiphysics claim, and the combined SIL and HIL verification loop. The parts most likely to be evaluated against Omniverse, namely rendering fidelity, ecosystem breadth, and the size of the third-party asset library, are the parts where a four-year-old company with a small team is furthest behind.

Against domestic CAE and simulation vendors, the comparison is less direct than the framing suggests. TenFong, Suochen, and the other domestic industrial simulation vendors sell solvers for fluid, structural, and thermal analysis to engineers who make engineering decisions. Songying sells a training and deployment platform to teams that train robot policies. These overlap at the level of physics simulation capability and diverge completely at the level of customer, decision, and purchase centre. Reading them as direct competitors overstates the overlap.

Against the embodied intelligence platform companies, the competition is for the same budget. Model companies and robot body companies increasingly build or buy their own simulation and data infrastructure. Songying's answer is to be the layer beneath them, which is precisely the position Lightwheel occupies from the data side, and which NVIDIA occupies from the compute and platform side.

Against open source robotics simulation stacks, the moat is engineering, not licence. Isaac Lab, MuJoCo, and their successors are free and improving quickly. The defensible ground for a commercial vendor is the integration depth, the industrial scenario library, the hardware adaptation and support commitment, and the certification and evaluation relationships with national institutions, none of which a user could assemble from open components alone.

The structural risk is the same one that faces every domestic localisation vendor. Policy-driven procurement creates demand that would not otherwise exist, and it also creates customers who localise to satisfy a process rather than to improve a result. Those customers do not renew, and the public record contains no renewal or retention data with which to separate the two populations.

Financing

The August 2026 rounds are the most recent disclosed events. The company announced that it had consecutively completed an A round and an A1 round totalling several hundred million renminbi, led by CICC Capital and Wuhan Hongshan Capital, with Shangrong Capital, Aux Investment, Gaoxin Capital, the Shanghai Angel Association, Fengyuzhu, and Qiaobei Capital participating as new and existing investors, and with Zhongxin Fund and other existing shareholders oversubscribing. It described the transaction as its third financing round within half a year, with proceeds directed to continued iteration of the ORCA OS physical AI operating system, deep adaptation to the domestic chip ecosystem, developer ecosystem construction, and scaled deployment in priority industrial scenarios.

Earlier rounds are documented across several sources with different dates and different naming. A seed round is recorded on 5 August 2022 at several million US dollars, jointly led by Qiming Capital and Sequoia Seed Fund. An angel round is recorded on 20 March 2025 with investors including Zhongke Chuangxing, Ruichuang Venture, MiraclePlus, and the Shanghai Angel Association technology service company. A Pre-A and Pre-A+ round was announced in February 2026, led by Dian Shi Capital and E Fund Private, with Danbei Capital and Zhongxin Capital participating, described as several hundred million renminbi raised over the preceding year.

The February and August disclosures are not a single continuous narrative. The February round is dated roughly six months before the August announcement and was described at the time as the company's most recent institutional round. The August announcement describes the A and A1 rounds as the third round within half a year. The registry-based count of distinct financing events is at least four, and five if the two February rounds are counted separately, and none of the sources provides a consolidated round table.

The one disclosed transaction price is a minority investment, not a valuation. Fengyuzhu's June 2026 announcement sets a ceiling of CNY 20 million for a minority stake with the holding percentage explicitly undetermined, to be fixed in the final investment agreement. No implied valuation can be derived from this figure, because the stake percentage is not disclosed. It is nonetheless a regulated, arm's-length, third-party price for equity in this company, and the report treats it as an anchor on the scale of a single ticket rather than as a valuation.

No valuation is disclosed for any round, and no cumulative financing total is stated by any source. The company's own language is "several hundred million renminbi" for the two August rounds and "several hundred million renminbi over the past year" for the February disclosure. This report therefore states the rounds individually and does not produce a cumulative figure.

DISCREPANCIES

1. The number of financing rounds. Convention A, the company's own August 2026 announcement: the A and A1 rounds are the third round within half a year. Convention B, corporate registry and investment database records: a seed round in August 2022, an angel round in March 2025, a Pre-A and Pre-A+ round announced in February 2026, and the A and A1 rounds in August 2026. Ruling: both are reported in parallel and neither is used as the company's round count. Convention B is preferred for the existence and dating of individual events because it is tied to registry records, and Convention A is preferred as the company's own characterisation of its funding velocity. The report states at least four distinct disclosed financing events.

2. Round naming for the most recent event. Convention A, the company: an A round and an A1 round completed consecutively. Convention B, the same company's February 2026 disclosure: Pre-A and Pre-A+ rounds. Ruling: the labels are reported exactly as each source gives them. The A and A1 naming follows the company's August announcement and the phase field of this report follows it, placing the company at Series A. The February disclosure is recorded as a separate event rather than being renumbered into the A sequence, because no source describes a restructuring or a renaming.

3. Cumulative financing. Convention A: several hundred million renminbi across the two August rounds, described as the third round within half a year. Convention B: several hundred million renminbi raised over the year preceding the February 2026 announcement. Convention C: a seed round of several million US dollars in August 2022. Ruling: no single cumulative total is given. The two several-hundred-million statements cover overlapping and different windows, the seed round is denominated in US dollars, and no source reconciles them. The report states the rounds individually with their amounts and does not add them.

4. Registered capital. Convention A, the Fengyuzhu Shanghai Stock Exchange filing of 12 June 2026: CNY 1.635735 million. Convention B, the corporate registry record with an approval date of 8 September 2026: CNY 1.73035 million. Ruling: the later registry figure of CNY 1.73035 million is used, with the filing date noted. The difference is consistent with a capital increase between June and September 2026, which is what a financing round would produce, but no source states that directly.

5. Founding date. Convention A, company materials: the ORCA programme and core technical team date from 2021. Convention B, the registry: Shanghai Songying Technology was incorporated on 6 December 2024, and Beijing Songying Technology was incorporated in 2021 as a wholly owned subsidiary. Ruling: both are reported and the distinction between programme start and legal entity incorporation is maintained. Reading the 2021 date as the incorporation date of the contracting entity is the common error, and reading the 2024 date as the start of the technical work is the other one.

6. Client count. Convention A, company material from August 2026: ORCA has served close to one hundred central and state-owned enterprises, national robot innovation centres, the national AI application pilot base, national supercomputing centres, and leading embodied AI enterprises. Convention B, company material from February 2026: several tens of central and state-owned enterprises, national and provincial robot innovation centres, humanoid and embodied intelligence pilot bases, leading manufacturing enterprises, and embodied model vendors. Ruling: both counts are reported with their dates, and neither is presented as current. The growth between the two statements is consistent with commercial ramp, and the counts include non-paying platform relationships.

7. Headcount against client count. Convention A: close to one hundred client organisations. Convention B: social insurance participation of seventeen people in the 2025 annual report, with registered capital of CNY 1.73035 million. Ruling: both are reported side by side and the tension is stated explicitly rather than resolved. No source explains the gap, and the most likely mechanisms, namely partner-delivered implementation and relationship counts that include joint labs and bases, are inference and are flagged as such.

8. The multiphysics first-claim. Convention A, company and government publicity: ORCA SIM is the first domestically self-developed real-time multiphysics coupling simulation system, fusing rigid body, fluid, and deformable body in one solver. Convention B: no benchmark, no comparison against domestic multiphysics solvers, and no independent reproduction appears in the sources consulted. Ruling: the claim is reported as a vendor statement with the absence of an independent test basis stated explicitly. This is a single unverifiable claim rather than a conflict between two measurements.

9. The operating system first-claim. Convention A: ORCA OS is the first physical AI operating system for multi-form robot collaborative training, presented at the 2026 World Artificial Intelligence Conference. Convention B: the term physical AI operating system is used loosely across the Chinese market by several vendors, and no authoritative body assigns the designation. Ruling: the claim is reported with its venue and its absence of an independent arbiter stated. The substantive demonstration, multi-form collaborative training inside one digital factory, is reported separately because it does not depend on the naming.

10. Fengyuzhu transaction and implied valuation. Convention A: an investment of up to CNY 20 million for a minority stake with the percentage undetermined. Convention B: no valuation figure is implied or derivable, since the stake percentage is not disclosed. Ruling: the CNY 20 million ceiling is reported as a disclosed transaction price and explicitly not treated as a valuation anchor. Any valuation figure inferred from it would be an arithmetic exercise, not evidence.

Risks and Due Diligence Gaps

Revenue is entirely undisclosed, and the headcount does not support the reported client footprint. There is no revenue, gross margin, backlog, or customer concentration figure anywhere in the file. The seventeen-person social insurance headcount against a claimed footprint approaching one hundred organisations is the first thing a counterparty should reconcile, because it determines whether the delivery model is software or services. Due diligence action: a revenue split between subscription, appliance, and services; the headcount at the present date rather than the 2025 annual report; and a named list of paying customers with contract values.

The financing narrative cannot be reconciled and no cumulative total can be stated. Three conventions describe the company's history differently, and two of them are company-sourced. Any counterparty relying on a single aggregator will work with a round count that a company statement contradicts.

The valuation cannot be derived from the only disclosed transaction. Fengyuzhu's CNY 20 million is a ceiling for an undisclosed minority percentage in a company whose later financing rounds are described only as several hundred million renminbi. The implied valuation spread across plausible stake percentages is wide enough that the figure is not usable for pricing. Due diligence action: the final investment agreement and the agreed percentage, and the fully diluted cap table.

The ORCA SIM multiphysics claim and the ORCA OS first-claim both lack an independent test basis. No throughput measurement, no accuracy benchmark against a competing solver, and no reproduction of the deformable-body and fluid interaction behaviour are available. Due diligence action: a benchmark of robot-hours per wall-clock-hour against Isaac Lab and MuJoCo, and a deformable-object accuracy study against measured ground truth.

The one-to-eight-to-one data mixture is a policy, not a result. Eighty percent synthetic data is a defensible choice in a domain where real data is scarce and expensive, and it is also the point at which simulation fidelity becomes the binding constraint on policy quality. No published measurement of downstream task success at that mixture is available. Due diligence action: task success rates stratified by data source, and a measurement of how performance degrades as the synthetic share rises.

Domestic chip adaptation breadth is a claim about roadmap coverage rather than measured performance. The list of adapted vendors is long, and each adaptation implies sustained engineering effort across changing driver stacks. A single vendor's silicon release cycle can invalidate an adaptation. Due diligence action: the number of chips on which the full feature set is certified rather than merely booting, and the support commitment for each.

The Siemens relationship is a cooperation agreement, not a contract. The agreement covers physical simulation, industrial digital twins, and embodied intelligence for advanced manufacturing, and it places Songying inside the Xcelerator ecosystem, which is meaningful validation. No revenue, no scope, and no duration are attached to it in any source.

The free-tier strategy has an unquantified conversion path. The company's economics depend on individual developers and educational users adopting the free edition and later converting, or on partner ecosystems creating downstream paid demand. There is no disclosed number of registered developers, no conversion rate, and no revenue from the enterprise tier attributable to the free funnel. Due diligence action: registered developer counts over time, and the share of enterprise revenue traceable to the open source channel.

The competition it has chosen is the hardest available. Competing against a platform ecosystem is a different proposition from competing against a point solution. The relevant risk is not that ORCA loses a feature comparison, but that the decision to build rather than adopt leaves it permanently behind on the axes where scale accumulates, namely rendering fidelity, asset library breadth, and third-party ecosystem depth.

There is no exit route and no valuation anchor. No valuation is disclosed for any round, the company is private with no listing timetable, and the only transaction price is a capped minority ticket with an undisclosed percentage. Any assessment of current value rests on assumptions about the last round price rather than on a documented figure.

Overall Assessment

• Technology 7.0. A genuinely layered engineering stack covering real-time ray tracing, a real-time multiphysics solver, a reinforcement learning framework, multi-dimensional data synthesis, SIL and HIL dual-loop verification, and Sim2Real deployment, with hardware decoupling across a long list of domestic GPU vendors that has genuine strategic value in the Chinese market, a coherent one-to-eight-to-one hybrid data strategy, a free SimReady asset library of more than one hundred thousand assets, and a demonstrated multi-form collaborative training scenario at the 2026 World Artificial Intelligence Conference. Deducted points: the multiphysics first-claim and the operating system first-claim both lack any independent test basis, no throughput measurement supports the parallel training scale claim, the data mixture has no published downstream result, and the layer most exposed to direct comparison with Omniverse, namely rendering fidelity and ecosystem breadth, is where the resource gap is widest.
• Commercialisation 5.5. A clear three-tier model with a free developer edition and a paid enterprise subscription, a hardware appliance channel, two service lines, a long list of named institutional relationships including Siemens, China Southern Power Grid, the National Robot Testing and Evaluation Center, the National Supercomputing Center in Wuxi, and the national embodied intelligence pilot base in Hangzhou, and a genuine third-party commercial validation in a disclosed listed-company investment. Deducted points: zero disclosed revenue, no gross margin, no backlog, no customer concentration, no renewal data, no pricing, a seventeen-person headcount that does not visibly support the reported footprint, and a stated count that mixes paying customers with platform relationships.
• Financing 6.0. Five disclosed financing events from a 2022 seed round through an August 2026 A and A1, an investor list that spans market investors such as CICC Capital and Qiming Capital, state-linked capital including Wuhan Hongshan Capital and Zhongxin Fund, and industrial buyers including Fengyuzhu, Aux Investment, and Huayu, and one regulated arm's-length transaction price from a listed company. Deducted points: an irreconcilable round-count conflict between company statements and registry records, no disclosed valuation at any round, no cumulative total that this report is willing to state, a minority ticket whose percentage is undisclosed and from which no valuation can be derived, and a company that describes the same period with two different round counts.
• Product maturity 6.5. A shipped product matrix including ORCA Studio, ORCA Lab, the appliance, and vertical solutions, a versioned product history from ORCA 1.0 in June 2024 through ORCA 2.0 in December 2024 to ORCA OS in July 2026, a developer edition distributed jointly with seven leading robot platforms, an open source incubation centre in Wuhan, and a national special fund for embodied intelligence open source. Deducted points: no disclosed reference customer list for the enterprise tier, no published documentation, release notes, or version history available for inspection, no disclosed service and support commitments behind the industrial delivery, and a services-heavy delivery model that is inherently constrained by headcount.
• Market position 6.0. A clear and durable position as the domestic alternative to Omniverse, backed by breadth of domestic chip adaptation that few peers can match, a named Siemens Xcelerator ecosystem placement, participation in the national embodied intelligence pilot base, a designated platform role in a China Southern Power Grid industry competition, and a coherent ecosystem strategy built on openness rather than closed hardware lock-in. Deducted points: the differentiator is positioning and ecosystem rather than a verified technical advantage, the company chooses the hardest competitive arena available, no third-party market share ranking exists for domestic physical AI operating systems, no evidence of international customer adoption, and the same localisation-policy double edge that applies to the entire domestic cohort.

References

1. Shanghai Songying Technology corporate registry record, unified social credit code 91310112MAE7NGHR77, including incorporation on 6 December 2024, registered capital of CNY 1.73035 million with paid-in capital of CNY 609,910, the Minhang District registration, the 8 September 2026 approval date, the seventeen-person social insurance participation from the 2025 annual report, and the listed directors
2. Fengyuzhu Culture Technology Shanghai Stock Exchange announcement 2026-039 of 12 June 2026, including the minority equity investment of up to CNY 20 million, the undetermined holding percentage, the intended joint venture, the accumulated twelve-month outbound investment ceiling of CNY 220 million at approximately 10.61 percent of audited net assets, and the related-party clearance statement
3. Songying company financing announcements of August 2026 covering the A and A1 rounds, the investor list led by CICC Capital and Wuhan Hongshan Capital, the several-hundred-million renminbi total, and the third-round-within-half-a-year characterisation
4. Songying financing disclosure of February 2026 covering the Pre-A and Pre-A+ rounds led by Dian Shi Capital and E Fund Private with Danbei Capital and Zhongxin Capital participating, and the several-hundred-million renminbi raised over the preceding year
5. Investment database records for the seed round of 5 August 2022 of several million US dollars jointly led by Qiming Capital and Sequoia Seed Fund, and the angel round of 20 March 2025 with Zhongke Chuangxing, Ruichuang Venture, MiraclePlus, and the Shanghai Angel Association technology service company
6. Songying company materials on ORCA OS, ORCA Studio, ORCA Lab, ORCA SIM, ORCA ENGINE, ORCA Gym, ORCA Replicator, ORCA Kit, ORCA DataLink, Orca Connect, Orca Store, Orca Agent, Orca Playground, and the ATEEM heterogeneous computing cluster, including the more than one hundred thousand SimReady asset library
7. Songying materials on the one-to-eight-to-one hybrid data strategy, comprising ten percent expert demonstration, eighty percent simulation synthesis, and ten percent real-world fine-tuning, and on the reconstruction of industrial parts with mass, friction, and material properties
8. Songying announcement of the March 2026 ORCA Lab 1.0 developer edition, released jointly with Unitree, Fourier, Beijing Humanoid Innovation Center, Norinco 58 Intelligent, DEEP Robotics, and Lingchu Intelligent, and the parallel training and multi-sensor support claims covering RGB, depth, IMU, and lidar
9. Science and Technology Daily, China High-Tech News, and Shanghai Observer coverage of the July 2026 World Artificial Intelligence Conference launch of ORCA OS as a multi-form robot collaborative training operating system, and the demonstration of humanoid, quadruped, wheeled, and aerial robots in one digital factory
10. Songying announcements of the Siemens strategic cooperation agreement and Xcelerator ecosystem entry of July 2026, the strategic cooperation with the National Robot Testing and Evaluation Center (Headquarters) on simulation dataset CR certification, the embodied intelligence computing joint laboratory with the National Supercomputing Center in Wuxi, and the national embodied intelligence application pilot base in Hangzhou
11. Songying and China Southern Power Grid material on the 2026 Electricity Embodied Intelligence Application Competition, with ORCA as the designated simulation platform for the online simulation event on substation operation and maintenance scenarios
12. Songying announcement of the September 2026 strategic cooperation with Yu Jia Zhishu Technology (Shanghai) on industrial multi-source heterogeneous data
13. Songying material on the CNY 300,000 donation to the Tiangong Kaiwu Open Source Foundation establishing the embodied intelligence open source special fund in June 2026, and the Wuhan open source incubation centre
14. Songying corporate registry information on Beijing Songying Technology Co., Ltd., incorporated in 2021 and wholly owned by the Shanghai entity, and the correspondence between the 2021 programme start and the December 2024 incorporation
15. Songying institutional credential and award records, including the Shanghai 2025 high-quality specialist service provider selection, the new generation AI innovation task call for proposals, the China Innovation and Entrepreneurship Competition national first prize, the 2026 SAIL award top-30 placement, and the tenth Maker China first prize
16. China Science Daily interview of 19 October 2026 with founder and chief executive Nie Kaixuan, covering the Huawei Cloud background, the 2021 core team assembly, the 2024 ORCA 1.0 and ORCA 2.0 releases, the funding trajectory including the March 2025 Zhongke Chuangxing angel round and the August 2026 A and A1 rounds, the one-to-eight-to-one data strategy, and the stated goal of building the Android of physical AI

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.