The Physical AI Era Begins: $10B+ in Investments Signals a Shift Toward Real-World AI Applications

In March 2026, AMI Labs, founded under the leadership of Yann LeCun, a Turing Award winner and former Chief AI Scientist at Meta, announced the completion of a $1.03 billion seed funding round.

Almost at the same time:

• World Labs, founded by Fei-Fei Li, has completed a new funding round of approximately $1 billion.
• Google DeepMind releases Genie 3 world model
• Tesla continues to advance the deployment of Optimus humanoid robots in factories.

These events are not isolated; together, they point to a clearer trend: AI is moving from "understanding the digital world" to "understanding and interacting with the physical world."

If 2024 was the era of large language model expansion and 2025 the period of practical exploration for agents, then in 2026, Silicon Valley’s core narrative is shifting to a more fundamental question: Can AI truly understand how the world works and accomplish tasks in reality?

This is not merely a shift in technological direction—it signifies that the industrial value chain is being rewritten. Over the past two years, the main battlefield of AI competition focused primarily on a few high-barrier areas such as models, computing power, and data centers; but as AI begins to truly enter the physical world, competition is no longer confined to the model layer—it is simultaneously expanding to include hardware itself, system integration, data collection, simulation environments, supply chain collaboration, and real-world deployment. In other words, Physical AI brings not just a single-point breakthrough, but a complete restructuring of an entire infrastructure system.

Precisely for this reason, this wave of change may represent not just a new technological trend but a rare structural opportunity window for the Chinese-speaking world—particularly for Chinese entrepreneurs, engineers, and investors. Unlike the previous cycle, which was primarily driven by large model training resources and super capital, Physical AI inherently demands a blend of skills: understanding algorithms while also mastering engineering, enabling system-level coordination while deeply engaging in manufacturing, supply chains, and industry applications. Teams that combine technical depth, hardware collaboration capabilities, and a broad perspective on both Chinese and U.S. industries are better positioned to seize key roles in this new cycle.

In other words, Physical AI is not just a new story being told in Silicon Valley—it may also be the most important ticket for Chinese stakeholders in the next global shift in technological infrastructure.

Over the past three years, large language models (LLMs) have nearly dominated the trajectory of AI development, with their core paradigm based on next-token prediction using massive text datasets. However, the limitations of this paradigm are becoming increasingly apparent: it can "describe" the physical world but lacks actionable understanding; it lacks the ability to model causality and physical constraints; and it performs poorly in continuous decision-making and long-term tasks.

Therefore, a school of thought led by Yann LeCun has begun advocating an alternative path: the World Model—predicting “states” rather than “text.” The core difference lies in the fact that LLMs take text as their learning object and language as their output format, remaining fundamentally within the realm of “cognition and expression”; in contrast, World Models model the states of the physical world, directly enabling a closed loop of “perception—decision—action.”

This is not just LeCun’s personal judgment. In Q1 2026, key advancements in the world models direction emerged nearly simultaneously: AMI Labs, with JEPA as its core architecture, explicitly bet on a long-term strategy of “research first, product later”; World Labs entered the field through “spatial intelligence,” aiming to enable AI to truly understand relationships, occlusions, and physical constraints in three-dimensional worlds; and Google DeepMind advanced real-time interactive dynamic environment generation with Genie 3, applying it to agent training.

Three companies follow different paths, but they all point to the same trend: the next leap in AI is not just about generating better text, but about more accurately modeling the world and taking actions within it.

The world model addresses the "brain" problem—how AI understands the physical world. But the other half of the battlefield for Physical AI is equally intense: who will build the "body"?

In 2026, the humanoid robotics sector has fully transitioned from “lab demos” to “factory-scale production.” Key figures:

Tesla Optimus Gen 3: Over 1,000 units have been deployed at the Gigafactory Texas and Fremont facilities, performing part handling and assembly tasks. This represents the largest deployment of humanoid robots in a factory in human history. Tesla is building a dedicated facility at Giga Texas with an annual production capacity of 10 million units, targeting a unit cost of just $20,000—down from the industry average of $50,000 to $250,000 just two years ago.

Boston Dynamics Atlas: The consumer version of Atlas at CES 2026 stands 6.2 feet tall, has 56 degrees of freedom, and can lift 110 pounds. More notably, its “soul”—Boston Dynamics has partnered with Google DeepMind to integrate cutting-edge foundation models into Atlas. Production capacity for 2026 has already been fully pre-ordered by Hyundai and Google DeepMind, with a factory capable of producing 30,000 units per year currently in planning.

Figure 03: Figure AI raised $1 billion at a $39 billion valuation. During an 11-month trial run at BMW’s Spartanburg plant, Figure 02 participated in the production of over 30,000 BMW X3 vehicles, moved more than 90,000 parts, and operated for a cumulative total of 1,250 hours. Figure 03 builds on this with a comprehensive upgrade, featuring 48+ degrees of freedom and the proprietary Helix AI platform.

Mind Robotics: Announced a $500 million funding round in March, focusing on industrial-scale deployment of AI robots.

But in this hardware race, an underappreciated component is emerging: the dexterous hand.

The legs of humanoid robots solve mobility issues, and the torso addresses load-bearing challenges, but it is the hands that ultimately determine whether a robot can perform tasks in complex environments. For example, in the Tesla Optimus, the hands account for 17% of the total cost—approximately $9,500—making them the most expensive single component.

The challenge with dexterous hands lies in a fundamental contradiction: the fingers are too small to accommodate large motors; small motors lack sufficient torque, requiring high-ratio gearboxes to amplify force; but high-ratio gearboxes introduce inertia distortion, loss of force feedback, and mechanical wear—three issues that physically "poison" the AI learning process.

A new wave of companies is striving to break through this bottleneck. Some are adopting axial flux motor architectures to reduce gear ratios from 288:1 to 15:1, enabling fully backdrivable dexterous hands; others are synchronously designing data-gloves that allow human operation data to be transferred to robotic hardware with zero loss. These seemingly small hardware innovations may be among the most critical infrastructures for the entire Physical AI ecosystem.

Each technological wave produces a "seller of shovels."

In the era of large models, NVIDIA emerged as the biggest beneficiary through its GPUs and CUDA ecosystem; in the era of Physical AI, its role is further evolving—not only providing computing power, but also aiming to build an entire infrastructure for the age of robotics.

At the 2026 GTC conference, NVIDIA unveiled a comprehensive suite of platform capabilities centered on Physical AI: including the Isaac GR00T vision-language-action model for humanoid robots, the Cosmos series for generating large-scale synthetic data, and an end-to-end toolchain covering training, evaluation, and deployment (such as Isaac Lab and OSMO). These capabilities are not isolated tools, but rather progressively form a complete development and operational system.

Several robotics companies, including Boston Dynamics, Caterpillar, Franka Robotics, LG, and NEURA Robotics, have already built their next-generation systems on the NVIDIA platform.

Its strategy is also very clear:

Not directly involved in end products, but becoming the foundational standard for the entire industry.

If Physical AI is a city under construction, NVIDIA is simultaneously providing cement, rebar, and the power grid.

In the world of large language models, the internet provides nearly unlimited text data. But in Physical AI, a more fundamental issue emerges:

Real-world manipulation data is extremely scarce.

This makes data one of the most critical and scarce resources across the entire industry chain.

Currently, the industry is primarily exploring three pathways.

Real-data roadmap. Represented by Physical Intelligence, its π0 model is trained on over 10,000 hours of real-world robotic operation data, covering diverse robot forms and task types, enabling complex manipulations such as folding clothes and assembling cardboard boxes. Its open-source initiative essentially provides the industry with a “manipulation pre-training foundation.”

Synthetic data pathway. Google DeepMind's Genie 3 and NVIDIA's Cosmos aim to generate vast simulated environments through world models, enabling training in virtual worlds before transferring to the real world. The core challenge of this approach is the sim-to-real gap, which is gradually narrowing as simulation fidelity improves.

Human teleoperation pathway. Using devices such as data-gathering gloves to directly map human actions into robotic systems. This approach offers the highest data quality but still faces limitations in cost and scalability.

Tesla is pursuing a hybrid approach: continuously capturing human operational behaviors through factory videos to train Optimus's movement capabilities.

In the long term, the competitive landscape for Physical AI will likely depend not on who has the best model, but on who possesses the most and highest-quality physical-world interaction data. Once the data flywheel starts turning, its barriers will grow exponentially.

Numbers don't lie. Here are the key funding events in the Physical AI space for Q1 2026:

[World Model Layer]

· AMI Labs (LeCun)—$1.03 billion seed round, $3.5 billion valuation

· World Labs (Fei-Fei Li)—$1 billion new round, Autodesk invests $200 million

[Base Model Layer]

· Physical Intelligence — Raising a new $1 billion round, valuation to exceed $11 billion

· RLWRLD — $41 million seed round extension

Humanoid Robot (Complete Unit)

· Figure AI — Raised $1 billion at a $39 billion valuation (2025)

· Mind Robotics — $500 million, industrial-scale deployment

· Galaxea — $434 million, Series B unicorn

· Humanoid — $290 million seed round, direct unicorn

· Generative Bionics — €70 million seed round

[Infrastructure and Tools]

· NVIDIA — Continued investment in the Isaac GR00T/Cosmos platform

· RoboForce — $52 million, Physical AI labor platform

Based solely on the above public data, Q1 has exceeded $6.4 billion. This does not include internal investments from major companies such as Tesla, Hyundai/Boston Dynamics, and Google DeepMind.

The flow of capital indicates one thing: Physical AI has moved beyond the "proof of concept" stage and entered the "infrastructure building" stage. Investors are no longer asking, "Can robots work?"—they’re asking, "Whose infrastructure will enable robots to scale the fastest?"

Of course, Silicon Valley has never lacked bubbles. Amid the frenzy surrounding Physical AI, several calm questions deserve consideration:

Demo ≠ Deployment. As consensus among industry insiders at Davos 2026 stated: the gap between a compelling demo and a system that can run flawlessly 10,000 times in a row is far greater than marketing suggests. Figure 02 did participate in the production of 30,000 vehicles at a BMW factory, but it performed relatively standardized part handling, not delicate assembly.

Sim-to-real remains a tough challenge. While the fidelity of world models is improving, the long-tail complexity of the physical world—lighting variations, material differences, and unexpected collisions—still poses the greatest obstacle for synthetic data approaches.

The business model has not yet been proven. LeCun himself said that AMI Labs is focusing solely on research in its first year. World Labs is experimenting with a freemium model, and Physical Intelligence has open-sourced its core model. Currently, these companies generate almost no revenue; investors are betting on a paradigm shift and market dominance in 3 to 5 years.

The gray rhino of security and regulation. When thousands of robots with autonomous decision-making capabilities enter factories and even homes, who is liable for accidents? Currently, there is almost no regulatory framework for Physical AI worldwide.

But these very issues indicate that we are at the early stage of a technological inflection point, not at the peak of a bubble. Every true paradigm shift—whether the internet, smartphones, or cloud computing—went through an early phase where demos outperformed actual products. The key distinction is whether the underlying technology is genuinely advancing, not just the PowerPoint presentations.

From LeCun’s JEPA architecture, to Genie 3’s real-time world generation, to π0’s 68-task generalization capability, to Optimus’s deployment at the scale of 1,000 factories—Q1 2026 advancements are tangible engineering breakthroughs, not theoretical fantasies.

Physical AI is not a new赛道; it is more like one of the ultimate forms of AI.

As AI transitions from “understanding the world” to “entering the world,” what gets rewritten is not only the boundaries of model capabilities, but also the division of labor and distribution of value. Future competition will not only occur within model parameters and computing clusters, but also in robotic bodies, dexterous hands, data collection, simulation systems, industry applications, and supply chain organization capabilities.

This is also why this round is particularly important for Chinese speakers.

Over the past two decades, one of the deepest strengths among Chinese professionals has never been a single-dimensional technological label, but rather the ability to truly integrate cutting-edge technology, engineering execution, hardware manufacturing, and cross-regional industrial collaboration. Whether entrepreneurs, engineers, investors, or industry resource organizers, anyone who can seize this shift from digital intelligence to physical intelligence has the opportunity not just to participate in the trend, but to become part of it at key levels.

In 2026, Physical AI may still be far from mature; but precisely because it is in its early stages, the window has just opened. For Chinese participants, this may not be another cycle of follow-the-leader involvement, but rather a new starting point with greater opportunities to deeply engage at the infrastructure, platform, and key component levels.

This article is from the WeChat official account "Silicon Rabbit" (ID: gh_1faae33d0655), authored by Silicon Rabbit.