Author | Huálín Wǔwáng

Editor | Jingyu

The aftershock of this year’s “Robot Spring Festival Gala” still echoes in public opinion. The robot products actively appearing on the Spring Festival Gala stage have led many to mistakenly believe they might enter thousands of households overnight.

Just as popular as the Gala is the investor frenzy for the “embodied intelligence” trend.

According to the latest Bloomberg news, “AI godmother” Fei-Fei Li’s World Labs has just secured $1 billion in funding, with investors including Autodesk, NVIDIA, and other industry giants.

This massive funding clearly indicates the giants’ confidence in “embodied intelligence” and “world models.”

01

From $1 billion to $5 billion, the “AI godmother” valuation skyrockets

Rewinding to 2024, World Labs just completed a $230 million funding round, with a valuation of $1 billion. At that time, the public’s perception of the company was still at the level of “Li Fei-Fei’s new startup.”

In just a year and a half, the valuation has increased fivefold. Such growth speed is extremely rare even in the AI field. More noteworthy is the lineup of investors: Autodesk invested $200 million, and giants like Nvidia, AMD, a16z, Fidelity, and others entered the scene.

“This is not an investment based on the ‘AI godmother’s’ personal halo, but a revaluation of World Labs and its focus on world models by capital,” a close-to-deal investor revealed.

Behind the skyrocketing valuation is the accelerated product development of World Labs.

World API interface | Image source: World Labs

In January this year, the company officially launched “World API,” providing developers and robot companies with API access to large-scale world models. Industry insiders call it the “GPT-2 moment for 3D spatial reasoning”—a milestone in productization.

More importantly, World Labs has found a clear path to commercialization.

Autodesk’s chief scientist revealed that the two companies will deeply integrate at the level of professional creative tools: “Customers might first use World Labs’ world model to sketch office layouts, then delve into specific details like table design within Autodesk’s technology.”

This B2B cooperation model gives investors a clear view of monetization prospects.

Unlike OpenAI’s strategy targeting C-end users, World Labs has chosen a more pragmatic enterprise service route.

02

Has the “GPT moment” for world models arrived?

If ChatGPT marked the “iPhone moment” for large language models, when will world models have their moment of glory?

From a technical perspective, world models already have the foundational conditions for breakthroughs.

World Labs’ core product Marble can create spatially consistent, high-fidelity 3D worlds from images, videos, or text, supporting output in standard 3D formats like USD. This means the generated content can be directly imported into game engines or design software, unlike other video generation models that can only output “black box” results.

World Labs’ Marble in Labs | Image source: World Labs

Nvidia CEO Jensen Huang’s judgment is even more direct: “Physical AI based on world foundational models is as fundamental as large language models are to generative AI.” This analogy suggests that world models could become the underlying operating system for next-generation AI applications.

However, there remains a gap between ideals and reality. Fei-Fei Li repeatedly emphasizes that data and evaluation are the fundamental constraints for scaling embodied intelligence. Unlike large language models trained on massive text data, world models require high-quality 3D spatial data and physical interaction data, which are much more costly and difficult to annotate.

Collaborations between World Labs and companies like Guanglun Intelligence aim to address this key bottleneck of “scalable evaluation.” Building a reliable simulation environment to assess a robot’s spatial intelligence is far more challenging than generating beautiful 3D scenes.

03

$1.3 billion influx, the world model track is hot

World Labs’ $1 billion funding is not an isolated case. Since early 2026, over $1.3 billion has flowed into startups focusing on world models.

The main players in this race include: AMI Labs (founded by Yann LeCun, seeking a €3 billion valuation), World Labs (founded by Fei-Fei Li, valued at $5 billion), and Google DeepMind’s Genie 3 project. Each represents different technological approaches and commercialization strategies.

The frenzy of capital reflects a consensus: AI is entering a new development stage. If the past few years were dominated by “language intelligence,” future competition will shift toward “spatial intelligence” and “embodied intelligence.”

Google’s earlier Genie 3 can produce directly operable 3D spaces, causing stock prices of engines like Unity to plummet | Image source: Google

However, some analysts warn: “The growth from $1 billion to $5 billion valuation reflects fierce competition among frontier AI deals and highlights how quickly valuations are changing in AI— even when companies haven’t yet achieved large-scale commercialization.”

Concerns about valuation bubbles are not unfounded.

Currently, World Labs’ main product is still at the API stage, far from large-scale commercial application. The key question is: how will World Labs translate its scientific ambitions into clear business adoption?

04

From lab to industry, how far is the road?

Despite record-breaking funding, World Labs still faces significant challenges.

Technical challenges. For world models to truly be applied to robotics and autonomous driving, they must generate visually realistic scenes and ensure the accuracy of physical laws. A slight error in physical modeling could cause a robot to crash or autonomous vehicles to have accidents in the real world.

Cost is another hurdle. Training world models requires enormous computing power and data. Nvidia’s Cosmos has trained on 9 quadrillion tokens from 20 million hours of real-world data, a cost only a few giants can bear. How will World Labs balance cost control and performance enhancement?

Ecosystem challenges may be the biggest. Unlike large language models with rich downstream applications, the ecosystem for world models is still in its infancy. Developers need time to learn new API interfaces, and enterprise clients need time to verify ROI, which could take longer than expected.

However, Autodesk’s $200 million investment sends a positive signal—at least in professional design fields, world models have found willing customers. This could be a crucial breakthrough for World Labs’ commercialization.

Ultimately, the real gamble behind this $1 billion funding is not Fei-Fei Li’s personal reputation but the broader trend of AI expanding into the physical world. Regardless of whether World Labs can realize its $5 billion valuation target, the development of world models is already irreversible.

As AI begins to understand space, simulate physics, and predict the future, we are one step closer to true “Artificial General Intelligence.”

But this step may be more expensive and longer than we imagine.