Legal name: TensorWorld (Beijing Wuyuan Technology Co., Ltd.)
TensorWorld Deep Dive: A Physics-Native Foundation Model Built on GeoPT and Transolver, Over $100M Raised Four Months After Founding, with Product and Customers Still Unverified
TensorWorld, founded in Beijing in July 2026 by GeoPT and Transolver author Haixu Wu, builds physics-native foundation models. It disclosed over $100M in financing within four months, with no shipped product or named paying customer yet.
Executive Summary
TensorWorld, legally Beijing Wuyuan Technology Co., Ltd., is a Beijing-based startup founded on 1 July 2026 that is building physics-native foundation models. Its founder and chief executive, Haixu Wu, is the first author of GeoPT and a co-author of the Transolver line of neural solvers; the other co-founders bring large-scale AI infrastructure, model-based reinforcement learning and 3D representation learning. The company states its aim as constructing physical reality inside a tensor world: a model should learn how geometry, material properties, forces and boundary conditions determine how a system evolves, and should predict the consequences of actions in a measurable and verifiable way. It explicitly frames its goal as physical fidelity rather than visual plausibility, and lists robotics training, industrial design, autonomous driving and scientific computing as intended application areas.
The technical narrative rests on a verifiable academic chain. Autoformer and TimesNet address long-horizon temporal modelling; Uni3D scales 3D representation learning to a billion parameters; Transolver introduces Physics-Token to organise physical fields into learnable units, with later Transolver++ and Transolver-3 work aimed at industrial-grade accuracy; GeoPT, a February 2026 preprint from MIT CSAIL and Tsinghua with Kaiming He and Wojciech Matusik among its authors, uses synthetic dynamics lifted onto off-the-shelf geometry as a self-supervised pretraining signal. GeoPT reports 20 to 60 percent lower labelled-data requirements and twofold faster convergence on industrial-fidelity benchmarks including DrivAerML, NASA-CRM, DTCHull, Car-Crash and Radiosity, with open code and more than one million pretraining samples. Transolver has been reported as adopted in NVIDIA PhysicsNeMo. EfficientZero closes the chain with planning over a learned world model.
Financing arrived with unusual speed. On 5 November 2026 the brand launch on X and several Chinese databases described a single $105M seed round with ten named investors; on 10 and 11 November the company and a wider group of media and investors described two successive rounds, an angel round and an angel-plus round, totalling more than $100M with different lead investors and an additional participant. Neither the individual round sizes nor a valuation were disclosed in the later account, while one database attached a roughly $600M valuation to the earlier account. The company has not disclosed named paying customers, orders, revenue, pricing or a shipped commercial model. This report synthesises public information on the team, technology, capital structure and market context, and rules item by item on the eight groups of conflicting public accounts recorded in DISCREPANCIES.
Company Overview
The legal entity is registered in Haidian District, Beijing, with a registered address on Zhongguancun East Road, a unified social credit code of 91110108MAKHYCW6XR, and a business scope centred on fundamental software development, artificial-intelligence algorithm and application software, AI public data platforms, data processing and intelligent-robot research and development. The registry lists an establishment date of 1 July 2026, an operating term to 30 June 2056, and a change on 14 October 2026 from a domestic limited-liability company to a foreign-invested, non-wholly-foreign-owned limited-liability company. That same change added several institutional shareholders, including vehicles associated with Cathay Capital, HongShan, CDH, LinkX and other investors, alongside four domestic employee-holding partnerships and the two individual founders.
Governance deserves separate attention from the public narrative. Media consistently describe Haixu Wu as founder and chief executive, while the registry lists Wen Zheng as legal representative, manager and financial officer. A 14 October 2026 filing on senior management added Wu as chairman of the board and named Gong Yuan, Li Maoxiang and Chen Han as directors, with Zheng retaining manager, director and financial-officer roles. The registry also records an employee count of zero, which is a social-insurance declaration field and cannot be read as the actual headcount of a research team. Registered capital is recorded as 125 with the currency field in renminbi, but the unit is not stated in the aggregator snapshot and should be confirmed against the official registry extract.
Wu was born in 1998, entered Tsinghua's School of Software in 2016 with a second degree in pure and applied mathematics, spent a research period at Harvard in 2019, and completed his doctorate at Tsinghua under Mingsheng Long. He received Tsinghua's top student honour in 2022, a ByteDance scholarship in 2024 and a place on the Stanford and Elsevier list of the top two percent of scientists in 2025, and began postdoctoral work at MIT CSAIL in 2025 with Wojciech Matusik. Public accounts of his Google Scholar citations range from about 29,600 to more than 30,000. Wen Zheng holds a Stanford computer-science doctorate under Stanley Osher, previously served as vice president of AI at Kuaishou and head of its Y-tech AI lab, and is credited with experience in hundred-million-user products and computing clusters of one hundred thousand accelerators. Weirui Ye earned a doctorate at Tsinghua's Institute for Interdisciplinary Information Sciences, completed postdoctoral work at MIT, visited Pieter Abbeel's lab at UC Berkeley, authored EfficientZero in 2021 and was named to the HongShan and MIT Technology Review AI25 list in 2026. Junsheng Zhou is a doctoral candidate at Tsinghua's School of Software working on 3D spatial intelligence, a Baidu Scholarship recipient and a contributor to Uni3D, an ICLR 2024 Spotlight paper. The team's published collaborations connect Tsinghua, MIT, Stanford and UC Berkeley.
Product and Technology
TensorWorld positions its intended product as a physics-native foundation model, described in Chinese coverage as a physics-native base model or physics-native simulator. Its stated criterion is that predictions should remain aligned with the real world when geometry, material properties, boundary conditions or external forces change. This is a different objective from generating visually convincing video, and it places the company closer to the neural-simulation and engineering-surrogate tradition than to video-generation world models. The company describes three layers of capability: understanding structure, predicting evolution and reasoning over actions.
The most concrete public technical evidence is GeoPT, arXiv:2602.20399, whose authors are Haixu Wu, Minghao Guo, Zongyi Li, Zhiyang Dou, Mingsheng Long, Kaiming He and Wojciech Matusik. The paper observes that neural simulators are constrained by the cost of solver-generated labels, while static geometry alone provides no dynamics supervision and can cause negative transfer. Its lifted geometric pretraining injects synthetic velocity fields into readily available geometry and derives a dynamics-aware self-supervisory signal without physics labels. Using ShapeNet-V1 geometries for cars, aircraft and ships, the authors generated more than one million samples, about five terabytes, reportedly in three days on an 80-core CPU, with a single sample taking about 0.2 seconds. A Transolver backbone was used at three scales from 3M to 15M parameters, with depth scaling tests from 8 to 32 layers. On DrivAerML automotive aerodynamics, NASA-CRM aircraft loads, DTCHull hydrodynamics, Car-Crash maximum-stress simulation and a Radiosity transfer problem, the method reduced labelled-data requirements by 20 to 60 percent and doubled convergence speed, with performance improving as model size and pretraining diversity increased. Code is published at github.com/Physics-Scaling/GeoPT.
Transolver supplies the representation layer. Coverage from investors dates the Physics-Token concept to 2023 and the Transolver paper to 2024, with later Transolver++ and Transolver-3 variants described as reaching industrial-grade simulation accuracy; the architecture is reported as incorporated into NVIDIA PhysicsNeMo. Autoformer and TimesNet cover temporal dynamics, Uni3D covers 3D spatial structure, and EfficientZero covers acting on a learned world model. Together these form the time, space, physics and planning progression that the company and its investors use to argue for a complete physical-base-model stack. Two caveats matter for diligence: these are published academic results from multiple institutions, not a disclosed TensorWorld product benchmark; and the 20 to 60 percent saving and twofold convergence gain are measured on specific fine-tuning benchmarks, so they do not by themselves establish general-purpose physical reasoning, long-horizon rollout stability or sim-to-real transfer for robot policies.
Applications and Target Customers
The company names four application families. Robotics teams need models that can rehearse tasks which are expensive, dangerous or difficult to capture on real hardware, and that can anticipate how an object moves or deforms under force before a robot acts. Industrial engineering teams need to evaluate different designs under load, flow, heat transfer and deformation before building physical prototypes, reducing the number of physical trials in a development cycle. Autonomous driving is named in Chinese coverage as a target domain, reflected in the presence of NIO Capital among the investors. Scientific computing and virtual experimentation are described as a longer-horizon use case, where a model would let researchers explore a wider parameter space and compare predictions with measurements.
The intended buyer is therefore the model and simulation layer behind robot training and industrial R and D: robot companies, industrial design departments, research institutions and possibly automotive engineering groups. Chinese business press frames the commercial test concretely, asking whether simulation results can actually guide real engineering decisions, how much time and cost they save relative to existing practice, and whether customers will pay repeatedly. At the time of writing, no named paying customer, order, contract value, deployment, revenue or pricing has been disclosed.
One investor-side claim deserves careful handling. Cathay Capital says it engaged with Wu before the company was formed, travelled to Boston twice during incubation and introduced industrial software companies including Dassault Systemes to examine potential engineering applications. An English-language report on the financing notes that the announcements did not identify commercial contracts or deployed products arising from those discussions. These are exploratory conversations, not customer evidence. Similarly, widely quoted figures that robot task success reaches 89.4 percent in simulation but about 12.4 percent in real homes are attributed by Chinese press to a Stanford HAI report; the underlying report and its task definitions were not located in this round of research, so the pair is recorded here as a media-cited illustration of the sim-to-real gap rather than a verified statistic.
Business Model and Ecosystem
No commercial model has been announced. Public materials describe a research programme and a brand, with the model still described as under development in the 5 November brand announcement. Third-party investment summaries speculate about base-model licensing, simulation services and a closed loop between embodied data and evaluation, but these are analyst projections and are labelled MEDIA and UNVERIFIED in this report. The company has not said whether its first product will prioritise robot training or industrial R and D, has not published an API, a model card, a benchmark suite or a licence, and has not described a pricing unit.
What is observable is the ecosystem around the company. Its capital table combines multi-currency venture funds with strategic investors whose own operations touch consumer services, industrial internet and autonomous driving. Ant Group appears in both financing accounts, moving from a follower in the earlier description to a co-lead in the later one, and has separate world-model exposure through its LingBot family at Lingbo and through an earlier strategic investment in Inverse Matrix. Meituan Strategic Investment enters as a new co-lead in the later account; NIO Capital, CDH VGC, IDG Capital, Fortune Capital, MSA Capital, LinkX Capital, Cathay Capital and HongShan complete the named roster. A single investment brief also names Landscape Capital as exclusive financial adviser while listing it among angel-plus participants, a dual role that should be confirmed.
The strategic value of this roster is access to scenario data, computing capacity and early design partners rather than booked revenue. It also creates a governance question typical of very young, heavily capitalised research ventures: a large first financing round at an early technical stage sets a high bar for subsequent validation, and investors with internal world-model programmes may evaluate the company partly as an option on a research direction. The company's ability to convert academic credibility into a repeatable product and a defensible data or evaluation loop will determine whether the ecosystem functions as distribution or merely as endorsement.
Competitive Landscape
The physical-AI and world-model field covered on this site contains several distinct approaches. Vinci AI and NP Co. pursue pretrained physics and industrial foundation models; Physical Superintelligence frames its work as a virtual physicist; World Labs works on spatial intelligence and world models; Qingyan Tech builds a differential-geometry-based physics engine; 1011.AI builds a general physics engine; XIRRA proposes a large physics model; Veeda AI and Sudo Tech work on robot foundation world models and simulation-based training; Qianjue builds predictive world models for embodied decision-making. TensorWorld's own stated differentiator is that it starts from physical laws and representation learning for physical fields, with pretraining designed to make physical laws an internal property of the model. Each of these companies addresses a different layer of the stack, and public evidence for all of them is at an early stage, so a ranking by technical merit would not be well founded.
The more immediate competition is for capital and research talent in a cooling market. A November 2026 industry survey by QbitAI reported that leading funds were willing to consider only a handful of world-model projects, that the valuation premium attached to the term world model was shrinking, and that financing announcements typically lag deal closings by about three months. CLS venture data, cited in that survey and in the Kechuangban Daily article on TensorWorld, counted five physical-AI and world-model financing events in October. Public-market comparisons such as Unitree's share-price volatility after its August listing and AMD's announced all-stock acquisition of World Labs at about $8.2B in September were discussed by investors as repricing anchors. Within that environment, QbitAI listed TensorWorld among the few teams still raising smoothly, alongside Inverse Matrix, Natural Will, LiberAI, Lingbo and Yuanli Wuxian, noting the common pattern of strong academic credentials and published technical evidence.
TensorWorld's relative strengths are a first-author research record with open code and reproducible benchmark claims, a co-founder mix spanning physical modelling, 3D representation, embodied learning and hyperscale infrastructure, and a financing round large enough to fund substantial compute. Its relative constraints are equally clear: it was four months old at the time of the financing, has no disclosed product benchmark or customer, and its technical claims are currently academic results rather than a deployed system. The market is also re-pricing from narrative to evidence, which raises the cost of a long research-only phase.
Financing and Capital Structure
Two financing accounts are in circulation. The earlier one, dated 5 November 2026 and tied to the brand launch on X plus Chinese company databases, describes a single seed round of $105M, about RMB 705M, with ten investors: Cathay Capital, HongShan, Ant Group, MSA Capital, Meituan, IDG Capital, CDH VGC, LinkX Capital, Fortune Capital and NIO Capital. One database attached a valuation of roughly $600M and described the same date as the launch of a TensorWorld physics foundation model. The later account, disclosed around 10 and 11 November by the company, its investors and multiple business media, describes two rounds completed within one month, an angel round co-led by Cathay Capital and HongShan with IDG Capital, CDH VGC, LinkX Capital and Ant Group participating, followed by an angel-plus round co-led by Ant Group, MSA Capital and Meituan Strategic Investment with Fortune Capital, NIO Capital, Landscape Capital and existing shareholders participating. That account says the total exceeded $100M, does not give individual round sizes, and does not disclose a valuation.
Neither account has been reconciled by a primary company statement that names both structures. The 14 October 2026 registry change to a foreign-invested entity and the addition of institutional shareholders are directionally consistent with offshore or dual-currency capital entering the cap table, but the registry extract does not itself state round sizes, valuations or closing dates, so any link between the two is an inference. CLS venture data additionally publishes a 71.09 percent model-estimated probability that the company raises again within two years; that is an algorithmic forecast, not a fact about the company, and is excluded from the factual record.
Whichever account is used, the financing is large for a company four months old and unusual in the density of strategic investors. It buys a long research runway and access to computing capacity, which matters for pretraining physics models. It also places valuation and expectation ahead of product validation. The observable next signals are a formal clarification of round structure, the first product release with an independent benchmark, and the first named commercial contract.
DISCREPANCIES
1. Round structure and number of closings. Account A: the 5 November brand announcement and several databases describe one $105M seed round. Account B: the 10 and 11 November company and investor disclosures describe two rounds, angel and angel-plus, completed within a month. Ruling: record both accounts without merging them, use the two-round description as the later company-side account, and mark the exact number of closings as unconfirmed. Source grade: OFFICIAL and DATABASE versus OFFICIAL and MEDIA.
2. Amount and valuation. Account A: $105M, about RMB 705M, with a valuation of roughly $600M. Account B: total above $100M, no valuation, no per-round figures. Ruling: use "over $100M" as the common floor, flag $105M as a single-source precise figure and the $600M valuation as an unverified database claim. Source grade: DATABASE versus OFFICIAL.
3. Disclosure timeline. Account A: the brand launched on X on 5 November and Tianyancha-based feeds reported the financing the same day. Account B: an investment brief dated 10 November calls that day the official announcement, while Kechuangban Daily, China Business Network, China Economic Net and investor posts concentrated on 11 November. Ruling: record 5 November as the brand and first financing disclosure, 10 to 11 November as the second disclosure wave, and note that announcements commonly lag closings by about three months. Source grade: OFFICIAL plus MEDIA and DATABASE.
4. Investor roster and adviser role. Account A: ten named investors, no Landscape Capital. Account B: adds Landscape Capital among angel-plus participants and lists separate leads and followers for each round. A single investment brief also calls Landscape Capital the exclusive financial adviser. Ruling: report the later two-round roster as the company-side list, flag Landscape Capital's dual participant and adviser role as a single-source claim. Source grade: OFFICIAL versus MEDIA.
5. Product status. Account A: the 5 November announcement says the company is developing a physics foundation model and launches the TensorWorld brand. Account B: a database report says the TensorWorld physics foundation model was formally released the same day, while Kechuangban Daily calls a physics-native base model the core product yet states that no named paying customers, orders, revenue or pricing have been disclosed. Ruling: record a brand launch with a model under development, and treat "released product" as an unverified database characterisation. Source grade: OFFICIAL versus DATABASE and MEDIA.
6. Corporate roles and registry data. Account A: media uniformly call Haixu Wu founder and CEO. Account B: the registry lists Wen Zheng as legal representative, manager and financial officer, with Wu added as chairman on 14 October 2026; the same snapshot gives an employee count of zero and registered capital of 125 with an unstated unit. Ruling: report Wu as founder and CEO and Zheng as legal representative and manager, treat the employee count as a social-insurance field rather than headcount, and mark the capital unit as unconfirmed. Source grade: MEDIA versus REGISTRY aggregator.
7. Technical dates and adoption claims. Account A: investor materials date Physics-Token to 2023 and describe Transolver++ and Transolver-3 as reaching industrial-grade accuracy, with Transolver adopted in NVIDIA PhysicsNeMo. Account B: the GeoPT paper cites Transolver as Wu et al. 2024 and Kechuangban Daily gives 2024 for the Transolver publication; one outlet also credits GeoPT with a best-paper award at an ICLR 2026 workshop. Ruling: record the concept as proposed in 2023 and the paper as published in 2024, mark PhysicsNeMo adoption as company and investor reporting pending primary verification, and flag the workshop award as a single-source claim. Source grade: OFFICIAL and MEDIA versus PAPER.
8. Customer validation and illustrative statistics. Account A: Cathay Capital describes incubation trips to Boston and introductions to industrial software companies including Dassault Systemes; Chinese press cites a Stanford HAI report for 89.4 percent simulated versus 12.4 percent real-home task success. Account B: an English report states that the announcements identified no commercial contracts or deployed products from those discussions, and this research did not locate the underlying HAI report. Ruling: record the introductions as exploratory, not customer evidence, and mark the success-rate pair as a media-cited illustration requiring primary confirmation. Source grade: OFFICIAL and MEDIA, unverified.
Risks and Due Diligence Gaps
The first gap is the distance between published research and a company product. GeoPT, Transolver, Autoformer, TimesNet, Uni3D and EfficientZero are verifiable academic contributions from teams at MIT, Tsinghua, Stanford and UC Berkeley, with open code and benchmark numbers for several of them. TensorWorld has not published a model card, a product benchmark, an inference cost figure, a licence or an API. Its 20 to 60 percent data saving and twofold convergence improvement are fine-tuning results on specific surrogate-simulation tasks; they do not demonstrate long-horizon physical rollouts, contact-rich manipulation, sim-to-real transfer or general action planning.
The second gap is commercialisation. No named paying customer, order, contract, revenue, pricing unit or deployment has been disclosed, and the company has not said whether the first product will serve robot training or industrial R and D. A base-model licensing or simulation-service business remains an analyst projection. The relevant diligence questions are whether an industrial customer will accept a learned predictor in an engineering sign-off workflow, what accuracy and uncertainty guarantees are required, and how the model is validated against physical tests.
The third gap is the financing record itself. The single-seed and two-round accounts differ in structure, amount precision, investor roster and valuation, and no primary statement reconciles them. Registry changes on 14 October 2026 are consistent with new institutional capital but do not establish round sizes or closings. A model-generated 71.09 percent refinancing probability is not evidence. Diligence should rely on signed transaction documents, the official registry extract and a company statement that names each round, its lead, its amount and its closing date.
The fourth gap is capital efficiency and team capacity. A financing round above $100M at four months of age enables expensive pretraining, but it also concentrates expectation on a research roadmap whose timeline is undisclosed. Several co-founders hold concurrent academic descriptions in coverage, including a doctoral candidate and a postdoctoral researcher, and the full-time composition, key-person retention and equity arrangements are not public. Computing commitments, data licensing and the cost of generating industrial-fidelity labels will determine how long the runway actually lasts.
The fifth gap is the technical boundary of the claimed paradigm. Long-horizon error accumulation, generalisation across unfamiliar geometries and materials, contact and friction modelling, deformable bodies and fluids, and the absence of agreed evaluation standards are open problems acknowledged in Chinese coverage of the sector. GeoPT's cross-domain result on Radiosity is encouraging but narrow. A credible product claim will need independent baselines, out-of-distribution tests, uncertainty estimates and physical validation, ideally published with enough detail for reproduction.
The sixth gap is information transparency and governance. The CEO and the legal representative are different people, the entity changed to foreign-invested status shortly before the financing disclosure, and several personal and institutional data points such as citation counts, titles and awards come from investor or media profiles rather than primary records. As a private company with no audited financials and no product, TensorWorld should be tracked through primary disclosures, the official registry and peer-reviewed publications rather than through financing headlines alone.
Overall Assessment
techScore is 8.5. The founding team has an unusually strong and checkable research record, with GeoPT providing open code, more than one million pretraining samples, five benchmarks and a scaling trend, and with Transolver reported in an industrial framework. Deductions apply because the results are academic, cross-task generalisation is unproven, and no TensorWorld product benchmark exists.
bizScore is 3.5. The intended application areas are concrete and the investor base offers plausible access to scenarios, but there is no named customer, order, revenue, pricing or go-to-market motion, and the business model is only speculated on by third parties.
fundScore is 9.0. A financing above $100M within four months, with leading venture funds and several strategic investors across consumer, industrial and automotive domains, is exceptional for the stage. Deductions apply for the unresolved single-round versus two-round accounts, undisclosed valuation and absent closing details.
productScore is 3.0. A brand and a research programme have been announced; the model is described as under development, with no release, benchmark, documentation, deployment or service-level evidence. A single database characterisation of a released model is treated as unverified.
marketScore is 7.0. The company is repeatedly named among the few world-model teams still able to raise in a cooling market, and its published research gives it visibility. Deductions apply for being four months old, having no customers, and facing a market that now prices evidence over narrative.
The composite view is 6.2. TensorWorld is a research-led physics-native foundation model startup whose main present assets are a co-founder team with first-author credentials, an open and reproducible pretraining result, and an unusually large early financing. Its main present liabilities are the absence of a shipped product, the absence of any commercial validation, and a public record with eight distinct groups of conflicting accounts. The most informative next signals are a primary reconciliation of the round structure, the first model release with independent baselines, and the first named paying customer.
References
- TensorWorld official site tensorworld.com and the 5 November 2026 brand announcement on X: positioning, the "build physical reality in a tensor world" framing and brand launch.
- Kechuangban Daily and CLS, 11 November 2026: two-round financing, investor roster, founder and co-founder profiles, absence of named paying customers, market cooling data.
- China Business Network, 11 November 2026: company founding date, physics-driven paradigm, three-layer capability description and industry analysis.
- China Economic Net and China Daily, 11 November 2026: technical roadmap narrative, Autoformer, TimesNet, Uni3D, GeoPT, Transolver and EfficientZero, and the cited sim-to-real figures.
- CDH VGC investor post, November 2026: physics-native framing, Physics-Token, Transolver and GeoPT stack, and the investor's thesis.
- NIO Capital investor post, November 2026: founder biography, co-founder biographies, EfficientZero and Uni3D details, and the physical-fidelity criterion.
- Hanlan Capital investor post, November 2026: Physics-Token chronology, Transolver variants, GeoPT with Kaiming He, and sector context.
- GeoPT project page physics-scaling.github.io/GeoPT and arXiv:2602.20399, February 2026: lifted geometric pretraining, one million samples, five benchmarks, 20 to 60 percent data saving, twofold convergence, open code.
- QbitAI, February 2026 GeoPT coverage and the 8 November 2026 industry survey on world-model financing: author affiliations, training-cost details, five October financing events and repricing context.
- Zhangtongshe, Securities Star and Tianyancha-based feeds, Spatial Daily English report, TokenDH, Jushenhome and the qibook registry snapshot, 5 to 11 November 2026: $105M seed account, ten-investor roster, valuation claim, brand launch details and registry fields.
This report is compiled from public sources and does not constitute investment advice. Data is current as of 12 November 2026. TensorWorld is a private, unlisted company with no audited financials and no disclosed product or paying customer; the round structure, amount, valuation, investor roster, product status, corporate roles, technical dates and customer validation carry eight groups of conflicting public accounts (see DISCREPANCIES). Uncertainties are flagged throughout the text; please rely on the company's primary announcements, the official company registry, peer-reviewed papers and signed transaction documents.
Related papers
Related tools
Related cases
Related reports
- NVIDIA Deep Dive: Building the Full-Stack Physical-AI Infrastructure with PhysicsNeMo, Warp, Cosmos World Foundation Models and Omniverse
- World Labs: Fei-Fei Li's Spatial-Intelligence Unicorn, Acquired by AMD for $8.2B
- Vinci AI: Turning Physics Simulation from a Checkpoint into a Continuous Constraint
- Qingyan Tech Deep Dive: A Geometric-Physics-Engine Company Incubated from Tsinghua and Yanqi Lake Institute, Building a World Model from Differential Geometry with a Model-Data Resonance Flywheel, Closing a Nine-Figure A Round in August 2026
This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.