---
title: "A Young Beijing Startup Raised Over $100 Million to Teach AI How the Physical World Works"
date: 2026-11-20
category: Capital & Markets
site: NeuroAI
canonical: https://neuroai.site/a/na-capital-tensorworld-physics-seed
language: en
---

# A Young Beijing Startup Raised Over $100 Million to Teach AI How the Physical World Works

> TensorWorld (物原科技) says it raised more than US$100 million within a month for a 'physical foundation model' that predicts how objects move and deform — a well-funded bet on a central gap in embodied AI.

A robot reaches for a glass and knocks it over. In simulation it never does: the virtual glass behaves, the virtual floor is predictable, and the task succeeds every time. That gap between a tidy digital world and a messy real one is what a young Beijing startup is trying to close.

In early November 2026, TensorWorld (物原科技) said it had raised more than US$100 million (≈ RMB 710M / HK$770M) across two early-stage rounds in a single month. It is a large sum for a company founded only in July, and it reflects a specific bet: that the next bottleneck in artificial intelligence is not language or images, but understanding physics.

## What "Physical Foundation Model" Means

TensorWorld describes its goal as building a physical foundation model (物理基础模型) — a system that learns objective laws such as gravity, friction, collision and fluid behaviour, then uses them to predict how a scene will evolve. The company describes its goal as constructing physical reality in a tensor world.

This is deliberately contrasted with the dominant world-model (世界模型) approach, which trains on visual data and chases photographic realism. TensorWorld argues that a model can produce convincing-looking frames while still having no reliable sense of what will happen next. Its stated sequence of capabilities is to understand structure, predict evolution and then plan actions.

The practical motivation is familiar to anyone working with robots. A policy that performs well in simulation often collapses when moved into a real home or factory, because the real world contains variations that the simulator never represented. Independently building physical rules by hand is also expensive, especially for fluids and soft materials. The company is betting that a learned physical model can close that gap more cheaply than more data collection.

Its technical lineage is public. Co-founder Wu Haixu (吴海旭) contributed to Transolver, a neural solver published in 2024 for physical simulation of engineering structures; the work has been incorporated into Nvidia's PhysicsNeMo framework. Another thread, GeoPT, uses large-scale geometric pretraining to reduce the need for expensive physical labels. The company says a technique that injects random velocity fields into static geometry can cut simulation data needs by up to 60% at comparable accuracy — a claim from TensorWorld itself that has not been independently published.

## Who Wrote the Cheques

The investor list is unusually broad for a seed-stage company. The first round was co-led by Cathay Capital (凯辉基金) and Sequoia China (红杉中国), with IDG Capital (IDG资本), CDH VGC (鼎晖VGC), Xinglian Capital (星连资本) and Ant Group (蚂蚁集团) participating. A second round was co-led by Ant Group, Hoyen Capital (和玉资本) and Meituan's strategic investment arm (美团战投), with Dachen Caizhi (达晨财智), NIO Capital (蔚来资本), Jinghe Capital (景和资本) and several existing investors following on.

The mix is the point. Ant Group brings industrial and healthcare scenarios, Meituan brings consumer-scale operations, and NIO Capital brings autonomous driving. For a company claiming its model generalises across settings, a cap table spanning those sectors is either a vote of confidence or a demanding set of future tests.

Reports consistently describe the total as "over US$100 million (≈ RMB 710M / HK$770M)" across the two rounds, and say neither round's individual amount was disclosed. A funding database later put the seed figure at US$105 million (≈ RMB 745M / HK$809M) and estimated a valuation near US$600 million (≈ RMB 4.26B / HK$4.62B). Those two precise numbers appear in a single dataset rather than in the authoritative news coverage, so they should be treated as estimates, not confirmed terms.

## The People Behind the Bet

Wu, the founder and chief executive, took a bachelor's and a doctorate in software engineering at Tsinghua University (清华大学), added a mathematics second degree, spent time at Harvard and, from 2025, worked as a postdoctoral researcher at MIT's Computer Science and Artificial Intelligence Laboratory. His research covers scientific machine learning and physical foundation models.

The co-founders reinforce that academic-industrial blend. Zheng Wen (郑文) holds a Stanford computer science doctorate and previously served as vice-president of AI technology at Kuaishou (快手) and head of its AI lab. Ye Weirui (叶葳蕤) completed a doctorate at Tsinghua's Institute for Interdisciplinary Information Sciences and a postdoc at MIT, working on reinforcement learning, world models and robotics. Zhou Junsheng (周俊昇) is a Tsinghua doctoral student who contributed to the Uni3D three-dimensional representation model.

Founded in July 2026 and headquartered in Beijing's Haidian district, the company adopted the English brand TensorWorld in early November, targeting robotics training, industrial design and scientific computing.

## Why This Is a Bet, Not a Result

The funding arrives while the wider world-model field is cooling in some quarters. Long-horizon prediction still accumulates errors, the distance between simulation and reality remains stubborn, and the sector lacks agreed evaluation standards. A physical approach may help with all three, or it may simply move the same unsolved problems into a different representation.

No independent benchmark reviewed here confirms that the model outperforms alternatives, and no reproducible results had been published at the time of writing. A large cheque bought the team time; it did not settle the technical question.

## Key takeaways

- TensorWorld (物原科技), founded in July 2026 in Beijing, says it raised more than US$100 million (≈ RMB 710M / HK$770M) across two rounds in one month.

- The rounds were co-led by Cathay Capital, Sequoia China, Ant Group and others, with Meituan and NIO Capital also participating.

- The company is building a physical foundation model (物理基础模型) intended to predict motion, deformation and forces, rather than merely generate realistic-looking images.

- Co-founder Wu Haixu contributed to Transolver, now part of Nvidia's PhysicsNeMo framework.

- A single funding database puts the exact sum at US$105 million and the valuation near US$600 million; authoritative reports give only "over US$100 million" and no confirmed valuation.

## Honest limitations

The financial figures come from company disclosures carried by Chinese financial media, not from an audited filing. The "over US$100 million" figure is consistent across several reputable outlets; the more precise US$105 million and US$600 million valuation appear to originate from a commercial database and are corroborated by encyclopedia entries rather than by independent reporting. I have therefore kept the rounded figure in the main text.

The 60% data-saving estimate is the company's own statement, with no peer-reviewed paper attached in the sources reviewed. Technical claims about Transolver and GeoPT are drawn from coverage of published research, but I have not read the full papers here, and TensorWorld's new model itself has not been independently evaluated.

Finally, a round this large for a four-month-old company reflects investor expectations, not evidence that the route works.

## What readers can do now

- **Separate the route from the result.** A well-funded, well-connected team is a reason to watch a lab, not a reason to assume its approach is correct.

- **Watch for a reproducible benchmark.** If the physical-model thesis is real, it should show up as published comparisons on standard tasks, not only as investor decks.

- **Follow the customers, not the cap table.** The first credible factory or robotics deployment will tell you more than the list of investors ever will.

---

Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-capital-tensorworld-physics-seed
Free to quote with attribution and a link to the original.
