Z Potentials has learned that Sand.ai recently completed two consecutive financing rounds, raising over USD 100 million, with a stellar lineup of investors. Leading and participating institutions include Look Capital, Lollapalooza Capital (Wang Huiwen's family office), Jiukun Capital, Matrix Partners China, MSA Capital, Innovation Works, Xianghe Capital, Source Code Capital, CASSTAR, Hongtai Fund, Jinshi Capital, HyT Capital, Yunhui Capital, IDG Capital, Baidu Ventures, and other top-tier institutions. Xinghan Capital served as the financial advisor for these financing rounds.

Matrix Partners China stated: "Matrix continues to focus on paradigm-level technological innovation and has made follow-on investments across multiple rounds. Sand.ai possesses full-stack capabilities in both pre-training and post-training, as well as solid infrastructure engineering expertise. It is one of the earliest teams globally to invest heavily in multimodal autoregressive and MoE architectures. We have witnessed the team's continuous exploration in the technology domain and their achievement of multiple milestones. We look forward to seeing Sand.ai continue to deliver breakthroughs and surprises in the future."

Lollapalooza Capital (Wang Huiwen's family office) commented: "Video generation is a key investment area for us. The development of video generation will create multiple major innovation opportunities, with several successful startups emerging. Chinese teams have very strong global competitiveness in the video generation field. Sand.ai has built a strong closed loop across foundational models, systems engineering, product experience, and commercialization. Over the past period, Sand.ai has demonstrated strong growth trajectory. We believe such a high-growth team will go even further in the future."

Catherine Zeng, Founding Partner of MSA Capital, said: "MSA Capital continues to invest in and empower globally disruptive and original innovation enterprises. As a lead investor in Sand.ai's first round of this financing series, MSA Capital firmly believes in the company's full-stack model and product capabilities, as well as its frontier exploration mindset. The company's next-generation model holds deeper potential in understanding physical laws, causal logic, and long-term narrative. At the same time, it has quickly validated its products, demonstrating that the team not only understands the frontiers of technology but also knows how to bring it to market. Starting from video generation, Sand.ai is steadily moving toward an interactive, evolvable world model. For this reason, we believe Sand.ai has the potential to become a leader in the next paradigm shift of AI."

Behind the simultaneous backing of multiple top-tier institutions lies a question: is this a valuation of the company's current achievements, or an early vote on the future technology paradigm? The answer may lie in the startup's technology choices and its critical judgments about the future.

In the era of large models, the choice of technical path reflects a company's sharpness in judging the future. In the video generation field, Sand.ai is one of the few companies willing to place bold bets and actively explore the fundamentals when the direction is not yet clear. This has made its technical route significantly ahead of the industry — at a time when Diffusion was still the consensus in video generation, it was among the first to shift its research focus to the autoregressive architecture, becoming one of the earliest definers of this direction. At that time, Sand.ai founder Cao Yue's judgment was that video is not just pixel generation, but a compression of space-time and physical laws. Compared to Diffusion, autoregression holds greater potential in real-time interaction, long-term prediction, and world understanding.

This judgment has also been effectively validated. The Magi-1 autoregressive video world model released by the company as early as early 2025 achieved an absolute lead on the Physics-IQ benchmark proposed by Google-DeepMind for evaluating physical authenticity, even surpassing Nvidia's newly launched flagship world model Cosmos-3-Super, and far outperforming other pure diffusion models such as Sora 2.

From content generation to world understanding, autoregression became an important starting point for Sand.ai's bet on the future.

But the real world has never existed in a single modality. Human perception of the environment essentially comes from the synchronized integration of multiple types of information — vision, sound, motion, and space. Therefore, training models on video pixels alone will always yield limited information.

With this in mind, Sand.ai launched a native audio-video co-generation model at the end of September 2025, incorporating audio signals into a unified modeling framework — making it one of the first teams in China to achieve audio-video co-generation. They found that when the model jointly models the two modalities, audio helps the visual generation produce more realistic details, and visuals likewise assist in audio generation.

Behind this, Sand.ai's judgment is that only through high-dimensional, multimodal joint modeling can the world laws compressed by the model come closer to a true representation of the physical world, thereby avoiding the "cognitive gap" caused by traditional single-modality generation.

By 2026, as model scale continued to expand, new challenges began to emerge. Video world models are simultaneously constrained by performance, speed, and cost. How to break through this long-standing "impossible triangle" has become the next battleground.

Sand.ai's latest technical blog post outlined its progress in this direction — to address scalability bottlenecks, the company shifted from traditional Dense to MoE (Mixture of Experts), with multiple innovations in specific architectural engineering.

First, compared to dense models where all network parameters participate in computation, the MoE architecture can dynamically activate only a subset of expert networks based on the content being generated. This allows for continuous expansion of model size, improving model capability while significantly reducing both training and inference costs. Beyond that, to address the challenges of applying MoE in video models, the team introduced a new routing mechanism to optimize communication efficiency, improve expert granularity, and enhance training stability. These innovations enable the model to achieve a better balance among performance, speed, and cost.

Another noteworthy choice is that Sand.ai adopted a single-stream unified architecture, rather than the multi-stream architecture commonly seen in the industry. The core idea is to uniformly map different modalities — text, image, video, audio, and more — into token sequences and model them through a single Transformer.

With the single-stream architecture and the MoE dynamic routing mechanism, different expert networks automatically learn parameter division and modality collaboration based on the input content. This means the model no longer relies on manually preset fusion rules, but can autonomously discover the correlation structures among different modalities during the training process.

To push efficiency to the extreme and tackle the computational challenges of multimodal and long-sequence processing, Sand.ai has continuously invested in underlying infrastructure R&D, systematically optimizing for long-sequence and heterogeneous attention scenarios. For example, the team introduced innovative operators such as Magi Attention, which significantly improve training and inference efficiency while preserving modeling capability, and reduce the computational overhead of long-context processing. These foundational capabilities are critical to determining how far the model can scale and how complex the tasks it can handle.

Looking at Sand.ai's technology evolution path — from its early, decisive bet on the autoregressive route, to its firm commitment to multimodal joint modeling, and now to its innovations in underlying architecture and operators — the company's technical choices have consistently anchored on one ultimate objective: driving the model beyond the surface of "content generation" to genuinely distill an understanding and simulation of the real world's operating laws. This ability to deeply compress and reconstruct the state of the real world is exactly the critical variable that will shape the ultimate competition in global world models.

In Cao Yue's view, video generation has never been the endpoint.

Over the past few years, discussions around world models have intensified, but a unified definition has yet to emerge. Many describe it as "predicting the next state." Cao Yue agrees that prediction is at the core of world models, but he remains wary of "human attempts to define what the hidden state is." History has repeatedly shown that every attempt to deconstruct the world with human priors essentially underestimates its complexity.

This lesson has already been fully demonstrated in the history of large language models. On the path to LLMs, countless efforts attempted to explicitly model representations of words, sentences, paragraphs, and even entire article structures — they were intricate, elegant, and aligned with human intuition about "understanding language," and were at times proven "efficient" at certain stages. But on the path to true scale, all of them, without exception, were outperformed by the simplest approach: predicting the next token. In the end, no one was able to define "the state of language" for the model.

So Cao Yue offered his real judgment: what should truly be predicted is not any human-defined state, but the one thing the world gives you for free and comes with its own supervisory signal — observation itself. This leads to his further conclusion: directly modeling raw data to build a world model may not be the most locally efficient approach, but it is highly likely to be the most scalable one.

And among all raw observations, what comes closest to the real world?

Cao Yue's answer is video.

The evolution of video models is, in essence, a continuous process of getting closer to the real world. In the beginning, they could only generate single images; then they learned temporal continuity; audio-video synchronization added the auditory dimension; multi-shot generation introduced spatial relationships; future prediction established causal associations; and real-time interaction brought closed-loop feedback. Each capability improvement was not the result of humans stuffing in another "state variable," but rather of allowing the model to develop its own understanding of space, time, sound, and causality from more complete observations. As these dimensions are modeled together, video models will eventually evolve into true world models.

Of course, from a more rational perspective, the ultimate ambition of world models is grand, and no one can reach it in a single step. At this stage, video models have already validated their commercial value in markets such as short videos, short drama production, and content creation — turning the technical momentum of "accumulating steps day by day" into tangible returns. Commercialization and technological evolution are not mutually exclusive. Real-world demand generates cash flow on one hand, and on the other, continuously produces new user feedback and data, fueling further model iteration.

Just as next-token prediction proved to be the winning path for reasoning, Cao Yue believes that next-frame prediction will be the same path for embodiment — rejecting the layering of a human-defined state on top of observations, and letting the model optimize itself.

As world models gradually move from technical concepts to engineering reality, a more fundamental question begins to surface: what kind of team can actually make this happen?

Compared to language models, the complexity of video models — and even world models — is elevated across the entire system. Whether in terms of architecture complexity, data supply, or computational consumption, this path is destined for only a few.

Globally, there are fewer than five video foundation model teams with true tier-one capabilities. The nature of competition has accordingly shifted.

In the early stage, it was about single-point capabilities — generation quality, resolution, or speed. But as we enter the world model phase, competition has moved to systematic capabilities such as data pipelines, model architecture, training efficiency, and product feedback loops. At this stage, the deciding factor is the organization itself.

Sand.ai founder Cao Yue was formerly co-founder of Lightyear Beyond, head of research at the Beijing Academy of Artificial Intelligence (BAAI), and a principal researcher at Microsoft Research Asia. His career spans both fundamental research and engineering implementation, representing world-class technical strength. One of his representative works, Swin Transformer, has become a foundational component of vision Transformer architectures and received the Best Paper Award (Marr Prize) at ICCV 2021 — one of the highest honors in computer vision.

In terms of academic influence, his papers have received nearly 90,000 citations, representing a research-driven technical background. This background determines that the team's architectural choices tend to focus more on "fundamental problems" rather than short-term engineering optimizations.

Algorithm lead Zhang Zheng has an equally strong background. A former researcher at Microsoft Research Asia (MSRA), a gold medalist in the ACM Asia Regional Competition, and a core author of Swin Transformer, he shared the Best Paper Award (Marr Prize) with Cao Yue at ICCV 2021. His total Google Scholar citations exceed 70,000, placing him firmly in the academic-driven technical profile.

On the product side, Operations and Growth Lead Wang Jia was one of the original seven members of the Douyin founding team, having served as Operations Director throughout Douyin's journey from 0 to 1. She was also the operations lead for the C-end product at Minimax. Meanwhile, VidMuse Product Lead Zhang Zihe (Zake) previously led product strategy and experience design for the PC version of CapCut from 0 to 1, was responsible for optimizing the camera imaging experience for OnePlus phones, and has long been active in the Bilibili ecosystem as a video content creator, bringing a genuine creator's perspective and product understanding.

This combination gives the team two concurrent capabilities: on one hand, understanding how the model "learns the world"; on the other, understanding how content is "used." This structure itself is a scarce resource. It requires the team to be able to define both the technical boundaries and the product form.

The scarcity of this capability endowment is also reflected in Sand.ai's shareholder structure.

If you look at Sand.ai's cap table, a distinctive feature emerges: it is not composed of a single type of capital, but rather an overlapping combination of multiple types of long-termist funds. There are industrial investors with years of experience in the tech industry, US-dollar funds focused on frontier technologies, institutions that have long backed hard tech and scientists, investors with deep technical backgrounds from tech companies, and a number of individual investors who are serial entrepreneurs themselves.

The logic behind these different pools of money is not entirely the same. Some value long-term technological potential and are less concerned with short-term returns; others better understand the development rhythm of underlying algorithms and are willing to wait; still others are investing primarily because of the team itself.

So this shareholder lineup brings Sand.ai not just capital, but also a cognitive network that covers different perspectives and experiences. For a company that aims to build long-term technology, this combination is more valuable than capital alone.

When a cap table brings together these different types of capital simultaneously, the bet is often not on a single product, but on a technological paradigm that could reshape industry structure. This opportunity clearly belongs to only a very few — and it represents a massive beta.