Source | BohuFN

Physical AI stands poised to be one of the defining keywords of 2025.

At the recently concluded CES Conference, Jen-Hsun Huang boldly declared, "The next frontier in AI is Physical AI, harboring trillions of dollars in potential opportunities."

Previously, Huang has repeatedly emphasized that "the new wave of AI is Physical AI."

As the name implies, Physical AI represents the fusion of physics and AI. In layman's terms, the output generated by AI must adhere to the laws of physics.

For instance, consider text-to-image or text-to-video models. Without considering physics, the generated content might lack details such as gravity and optics. By incorporating physical knowledge, the content becomes more realistic.

At this CES Conference, centered around NVIDIA's newly launched Cosmos platform, "Physical AI" not only transcends the limitations of cloud computing but also imbues AI with the capabilities of reasoning, planning, and action.

After the buzz created by ChatGPT's emergence in late 2022 and the plethora of AI "toy-level products" in 2023, the AI field in 2024 appears somewhat "calmer." Upon closer inspection, it feels like something is missing—that initial "revolutionary" spark, resembling more of a patchwork of existing achievements.

Increasingly, startups are focusing their efforts on the application level, no longer chasing the dream of AGI.

Looking back at NVIDIA's CES Conference, it felt akin to watching a Steve Jobs Apple event—a revolutionary technological breakthrough. It's safe to say that the advent of Physical AI will bring transformative changes to fields such as robotics and autonomous driving.

Furthermore, this could signal a "turning point" in NVIDIA's future business roadmap.

01 From Generative AI to Physical AI: NVIDIA's Bold New Year Move

On January 7, Beijing time, CES 2025 in Las Vegas became a stage where technology and luxury intertwined. Under the spotlight, NVIDIA's founder and CEO, Jen-Hsun Huang, donning his iconic "flashy leather jacket," unveiled a suite of new NVIDIA products: the launch of the new GeForce RTX 50 series graphics cards, delivering a leap in performance while reducing prices; emphasizing the goal of creating a giant chip utilizing 72 Blackwell GPUs or 144 chips, surpassing the capabilities of the world's fastest supercomputers...

Yet, the most captivating announcement was NVIDIA's launch of the first generative world foundation model, Cosmos.

This is a set of models specifically designed for physical interaction, simulating factory and driving environments. It encompasses components such as autoregressive models, diffusion foundations, advanced labelers, and AI-accelerated data pipelines, capable of generating physics-based videos from inputs like text, images, videos, and combinations of robot sensors or motion data.

The introduction of this model signifies that AI has evolved beyond mere perception and generation to encompass reasoning, planning, and action, as if endowing AI with a "brain" that can operate seamlessly within the physical rules of the real world. It marks a new pinnacle for AI technology towards reasoning, planning, and action, granting intelligent agents the ability to comprehend the physical world and interact dynamically with the real environment.

Physical AI represents a new stage of AI defined by NVIDIA, primarily encompassing two levels:

One is as a simulation tool: integrating the Physical AI model into autonomous machines to achieve perception, understanding, and execution of complex operations in the real world.

The other is generating data that conforms to physical laws for model training: creating and outputting more data for extensive model training, breaking through the current bottleneck of scarce real-world data.

NVIDIA's Cosmos model, released this time, focuses on the latter, emphasizing physical dynamics and human interaction, enabling AI to advance from mere perception and generation to reasoning, planning, and action, giving AI a "brain" that can operate flexibly within the physical rules of the real world.

02 The Next Frontier for Physical AI

Just as large language models revolutionized generative AI, Physical AI has become the "key" to unlocking a new era in fields such as autonomous driving and robotics.

Firstly, the challenge of "boarding" large models will be effectively addressed.

Currently, the application of large models in the automotive sector is primarily evident in two areas: intelligent cockpits and autonomous driving. The former naturally aligns with large model technology as current intelligent cockpits emphasize entertainment and interaction functions, which harmonize well with the language processing capabilities of large models. The difficulty lies in the latter.

For autonomous driving, achieving efficient and safe vehicle control in complex and dynamic traffic environments has become a core challenge. Existing autonomous driving systems generally lack multi-agent collaboration capabilities, efficient decision-making and explanation abilities, and struggle to effectively comprehend the behaviors and intentions of surrounding traffic participants in complex traffic scenarios.

Secondly, there's the issue of data. In the realm of autonomous driving, large models need to be "fed" vast amounts of real-world data for training to make them more human-like. Thus, making these data better serve the training of large models is another hurdle currently faced by many automakers.

Jen-Hsun Huang stated, "The world foundation model serves as the cornerstone for driving autonomous vehicle development, but not all developers possess the expertise and resources required to train their own models." The Cosmos model, based on 20 million hours of video training, focuses on physical dynamics and human interaction. Developers can utilize Cosmos to validate intelligent driving program logic in simulation environments and obtain data that is challenging to acquire in the real world for continuous training.

According to Huang, automakers currently collaborating with NVIDIA in the automotive sector include Tesla, BYD, Jaguar Land Rover, Li Auto, Mercedes-Benz, Toyota, Rivian, Xiaomi Automobile, Volvo, Lucid, and Zeekr, among others, as well as numerous L4 autonomous driving companies.

However, it's worth noting that XPeng Motors and NIO, two of NVIDIA's earliest and deeply cooperative partners in the field of automotive chips, were "absent" from the "Best Partner" segment of the CES special event in 2025. This coincidence aligns with the only two automotive enterprises that officially released their self-developed intelligent driving chips for the L3 and even L4 eras last year.

Secondly, humanoid robots are accelerating towards their "ChatGPT moment."

In recent years, inspired by large text models, leading global tech giants such as Google, OpenAI, and Microsoft have favored embodied intelligence. Last year, when its invested AI robot startup Figure AI launched Figure 02, it garnered significant market attention. Figure 02 integrates OpenAI's GPT-4 multimodal large model in its brain, enabling it to better understand and respond to complex instructions.

However, at the 2024 World Robot Conference, Professor Jia Jiaya, Chair Professor at The Hong Kong University of Science and Technology and Founder of Simo Group, pointed out a significant flaw in embodied intelligence—it lacks a perceivable personality or "machine personality," and one cannot discern its emotions such as sadness or happiness.

In the long run, both AI and robots need to leverage application innovations from implementation and scenarios.

With NVIDIA's support, humanoid robots are accelerating towards their ChatGPT moment. During this "tech spring festival gala," Jen-Hsun Huang demonstrated how they train robots to work like humans using the world's real models. Huang joked that this was his "steel legion."

In Huang's vision, the next high-volume robot product manufactured within a robot factory might be a humanoid robot, and the robot that can most seamlessly adapt to the world is also a humanoid robot.

Huang revealed that many leading robot and automotive companies have already become the first users of Cosmos, including 1X, Agile Robots, Agility, Uber, and more.

In the near future, robots will no longer be a distant concept but will increasingly integrate into our daily lives.

03 NVIDIA's Ambitions Extend Beyond Chip Manufacturing

Behind the leap from generative AI to physical AI lies NVIDIA's future growth trajectory.

Before ChatGPT ignited the large model market, NVIDIA's financial performance was steady but far from the current frenetic growth. The rise of large models and the prevalence of AI technology instantaneously propelled it to new heights. Financial reports reveal that from fiscal years 2021 to 2024, NVIDIA's revenue soared from $16.7 billion with a net profit of $4.3 billion to $60.9 billion in revenue and $29.7 billion in net profit.

From the perspective of business income composition, NVIDIA's growth core stems from its data center business, generating nearly 90% of its revenue. According to Guofu Data, 98% of the GPUs used in global data centers originate from NVIDIA, and 92% are utilized in the field of generative AI. Among them, the demand for graphics processing units (GPUs) has driven NVIDIA's market value to soar, once peaking at over $3 trillion.

However, most analysts believe that NVIDIA's fate in the coming years is almost predetermined. Competition in the semiconductor industry is fiercely intense, and in recent years, pioneers in the field of artificial intelligence such as Tesla and OpenAI have also been actively exploring independent and autonomous computing solutions. Although due to the long investment and construction cycle, it remains challenging to shake NVIDIA's market position in the short to medium term, data center investments tend to be cyclical, with both highs and lows. Therefore, NVIDIA urgently needs to tap into another substantial market.

Currently, NVIDIA's focus areas undoubtedly lie in autonomous driving and robotics. At this stage, these two fields have already embarked on a fast track of rapid growth, demonstrating considerable market potential. According to Goldman Sachs' predictions, by 2030, the value of the global autonomous driving industry could exceed $100 billion. And Statista data shows that the global robotics market was worth 728.886 billion yuan in 2021 and is expected to reach 1,820.541 billion yuan by 2028, with a compound annual growth rate of 13.97%.

Previously, NVIDIA has repeatedly mentioned in financial reports that the automotive business is seen as a new growth point in the future, with autonomous vehicles and electric vehicles being pivotal to this transportation revolution. During this keynote speech, Jen-Hsun Huang also specifically emphasized the automotive segment and set a new goal for the automotive business to achieve $5 billion in revenue by fiscal year 2026. This implies that the revenue target for the next two years must at least quadruple to meet the KPI.

It's worth mentioning that while NVIDIA undoubtedly stands a realistic chance of securing a share of the smart car sector with its deep expertise in the fields of artificial intelligence computing power and display computing, it would be fanciful for Huang to expect to dominate the entire automotive and robotics fields as he did in the GPU sector.

According to data released by the China Passenger Car Association on January 2, 2025, China's new energy passenger vehicles maintained a global market share of 69.6% in the first 11 months of 2024. In other words, NVIDIA, which is currently deeply entangled in China's antitrust investigation and attempts to bypass China—with nearly 70% of the global market share—to create a unique business ecosystem is clearly a pipe dream.

Although NVIDIA's intelligent driving chips still occupy the top spot globally due to their computing power advantages. According to Gasgoo data, from January to August 2024, NVIDIA's Drive Orin-X intelligent driving chip ranked first with 1,092,650 installations and a 37.2% market share; Tesla's FSD and Huawei's Ascend 610 ranked second and third, respectively. However, with the gradual maturity of domestic automotive chips such as Huawei's Ascend and Horizon Robotics, as well as the launch of intelligent driving chips by new forces such as NIO, XPeng, and Li Auto, it has become nearly impossible for NVIDIA to "rule the roost."

For robots (embodied robots or humanoid robots), with the support of artificial intelligence, their greatest value lies not just in acrobatic rolls or jumps but in integration with industries, such as automotive robots, mining robots, logistics robots, delivery robots, and more; this also signifies that as the global manufacturing hub and the world's largest producer and exporter of electromechanical products, the large-scale application of AI + robots can only flourish in China.

The most obvious example is the global layout of "Lighthouse Factories." According to data released by the World Economic Forum in October 2024, the total number of global "Lighthouse Factories" has increased to 172. Among them, the number of factories in China reached 74, accounting for 43%.

Therefore, for the Physical AI mentioned by Jen-Hsun Huang, which represents the field where artificial intelligence applications truly take root and create value, China, with its profound manufacturing foundation and industrial scale, will undoubtedly be the largest beneficiary and the largest single market for Physical AI.

Just as Huang predicted, "Physical AI will revolutionize the $50 trillion manufacturing and logistics industries. From cars and trucks to factories and warehouses, everything that moves will be a robot embodied by AI." But the biggest winner will undoubtedly be China.

Reference Sources:

1. Science and Technology Innovation Board Daily: Jen-Hsun Huang, Royal Flush!

2. Science and Technology Innovation Board Daily: NVIDIA Opens Source World Foundation Model, Humanoid Robots Accelerate Towards "ChatGPT Moment"

3. Xinhua News Agency: NVIDIA Under Investigation!

4. Electric Planet: Frontier | NIO and XPeng Exit NVIDIA's Circle of Friends

5. Huashang Taolue: The Global Chase Against NVIDIA

The cover image and illustrations featured in this article are subject to the copyright owner's rights. Should the copyright holder deem their content inappropriate for public viewing or unsuitable for free use, kindly contact us promptly, and we will take immediate steps to rectify the situation.