Fysics AI launches physics-based world model to rival OpenAI Sora, Meta V-JEPA
Synopsis
Shanghai start-up Fysics AI, founded by a former Nvidia senior manager, has launched Fysiverse — a world model that hard-codes real-world physics laws rather than learning them from data, directly challenging the architectural foundations of OpenAI's Sora and Meta's V-JEPA in the race to power robotics and autonomous driving.
Key Takeaways
Fysics AI announced Fysiverse on 25 June 2026 , describing it as a 'new-generation physics-based world model' that adheres to real-world physical laws.
The company was founded by Zhang Lihua , a former senior manager at Nvidia , and is headquartered in Shanghai .
Fysiverse claims to address 'physical illusions, reasoning failures, and breakdowns in non-standard scenarios' — weaknesses identified in existing world models.
OpenAI 's Sora and Meta 's V-JEPA represent the two dominant competing paradigms: video-based generation and self-supervised learning respectively.
World models are used to generate synthetic content and train robots and self-driving vehicles , making architectural choices in this space commercially and strategically significant.
Shanghai-based Fysics AI has unveiled Fysiverse, a physics-based world model that embeds real-world physical laws directly into its architecture — a deliberate departure from the data-driven and self-supervised paradigms championed by OpenAI and Meta Platforms. The announcement was made via the company's WeChat account on Wednesday, 25 June 2026, positioning the start-up as a challenger to established Western AI road maps in the fast-growing world model sector.
A New Paradigm in World Modelling
According to the company, Fysiverse is a 'new-generation physics-based world model that adheres to real-world physical laws'. The start-up claims it 'represents a new paradigm' capable of resolving persistent weaknesses in existing world models, including 'physical illusions, reasoning failures, and breakdowns in non-standard scenarios' — limitations that have dogged both video-generation and self-supervised approaches.
Fysics AI was founded by Zhang Lihua, a former senior manager at Nvidia, lending the venture credibility in hardware-aware AI design. The company's approach contrasts sharply with rivals by hard-coding physics constraints rather than inferring them from data.
The Competitive Backdrop
The world model sector — critical for generating synthetic content and training robots and self-driving vehicles — is currently structured around three dominant paradigms. The first is video-based generation, exemplified by OpenAI's Sora, which the firm has described as 'a promising path towards building general purpose simulators of the physical world' by scaling video generation models.
The second paradigm, associated with Meta's V-JEPA series (video joint embedding predictive architecture), allows a model to construct its own internal world rules without prior physics knowledge, using what Meta describes as a 'self-supervised learning approach'. Fysiverse targets both as insufficient for reliable physical reasoning. Other notable players in the space include World Labs, co-founded by Stanford professor and former Google AI chief Fei-Fei Li.
Why It Matters
World models underpin a broad range of next-generation applications, from autonomous driving to humanoid robotics — sectors where physical accuracy is non-negotiable. A model that genuinely encodes physics laws, rather than approximating them statistically, could reduce training data requirements and improve generalisation in edge cases that data-driven models routinely fail.
For the physical AI industry — a term encompassing AI systems that interact with the real world — Fysiverse's claims, if validated, would represent a meaningful architectural shift. The approach echoes Nvidia's own PhysX engine philosophy, which prioritises physics simulation fidelity, a lineage that Zhang Lihua's background makes strategically coherent.
What's Next
Independent benchmarking of Fysiverse against Sora and V-JEPA will be the critical test of whether the physics-first approach delivers on its claims in real-world robotics and simulation tasks. The broader race to define the dominant world model paradigm is accelerating, with significant capital and geopolitical stakes attached to whichever architecture becomes the industry standard for training embodied AI systems.
Point of View
You need less proprietary training data. Zhang Lihua's Nvidia pedigree is notable: the company's PhysX simulation stack has long argued that physics fidelity cannot be approximated away, and that thesis is now being applied to foundation model architecture. What mainstream coverage underplays is that the world model race is not just a product competition — it is an infrastructure battle for the default simulation layer that will train the next generation of embodied AI, from warehouse robots to autonomous vehicles. Whichever paradigm wins will exert platform-level leverage over the entire physical AI supply chain, making Fysiverse's architectural claims worth watching closely even if commercial validation remains pending.
NationPress
25 Jun 2026
Frequently Asked Questions
What is Fysiverse and who made it?
Fysiverse is a physics-based world model developed by Shanghai -based start-up Fysics AI , founded by former Nvidia senior manager Zhang Lihua . It was announced on 25 June 2026 and is designed to simulate reality by embedding real-world physical laws directly into its architecture, rather than inferring them from data.
How is Fysiverse different from OpenAI Sora and Meta V-JEPA?
Fysiverse hard-codes physics constraints into the model, whereas OpenAI 's Sora learns physical behaviour by scaling on massive video datasets and Meta 's V-JEPA uses self-supervised learning to construct internal world rules without physics knowledge. Fysics AI argues its approach avoids 'physical illusions, reasoning failures, and breakdowns in non-standard scenarios' that afflict the other two paradigms.
What are world models used for in AI?
World models are used to generate synthetic content and to train robots and self-driving vehicles by simulating how the physical world behaves. They are a foundational layer for physical AI — systems that must interact reliably with real environments — making architectural accuracy critical for safety and performance.
Why does Fysics AI's physics-based approach matter for robotics?
In robotics and autonomous driving, physical accuracy in edge cases is non-negotiable — a model that hallucinates physics can cause real-world failures. A physics-first architecture could reduce dependence on vast training datasets while improving reliability in unusual scenarios, potentially lowering the barrier to entry for companies without access to large proprietary data pools.
Who else is competing in the world model space?
The world model sector includes OpenAI ( Sora ), Meta ( V-JEPA ), and World Labs , co-founded by Fei-Fei Li , among others. Fysics AI enters as a Chinese start-up proposing a third architectural paradigm, adding competitive pressure from outside the US hyperscaler ecosystem.
