Jensen Huang Claims Physical AI Could Be Ten Times the Scale of Digital AI
Nvidia CEO Jensen Huang stated that physical AI could represent an industry opportunity ten times larger than digital AI. The physical AI he refers to encompasses robotics, autonomous vehicles, drones, and intelligent machines capable of perceiving, reasoning, and acting in real-world environments.
Huang noted that Nvidia's current annual revenue from robotics, autonomous driving, and related physical AI businesses is nearing $10 billion, and he expects this business to grow to $100 billion over the next decade; this is a tenfold growth target for Nvidia's own physical AI revenue, not a precise comparison of realized revenue or the overall digital AI market size.
Nvidia achieved $96.2 billion in revenue for the second quarter of fiscal year 2027, ending July 26, 2026, representing an 18% quarter-over-quarter increase. In its earnings report, the company highlighted "physical AI rollout" alongside frontier model labs and open model ecosystems, stating that they collectively accelerate the construction of AI infrastructure. Currently, physical AI accounts for only a small portion of Nvidia's total revenue, but it is viewed by management as the next phase of growth.
Nvidia sees autonomous driving as the first scalable entry point for physical AI. Huang believes the automotive market is large enough, with vehicle forms and sensor architectures relatively standardized and commercial value sufficiently clear, allowing for a faster formation of the data, chips, software, and mass production cycle compared to general humanoid robots.
Its technological path does not rely solely on robotics chips but rather on a "three-computer" system: AI factories responsible for training world models, simulation systems like Omniverse generating and testing virtual environments, and onboard or robotic computers executing real-time perception and control. Nvidia covers training, testing, and deployment simultaneously through GPUs, data center networks, simulation software, autonomous driving computing platforms, and edge chips.
The market mechanism involves buyers such as automotive manufacturers, robotics companies, industrial automation firms, warehousing and logistics operators, drone manufacturers, and manufacturing plants; they purchase not just chips but also training computing power, simulation tools, onboard computing, sensor data processing, and safety verification capabilities. Funding will flow towards GPUs, edge AI chips, industrial robots, machine vision, digital twins, synthetic data, and safety certification; traditional low-automation manufacturers, robotics firms lacking data loops, and suppliers relying solely on low-cost hardware assembly will face pressure from technological upgrades and capital expenditures.
Source: Public Information
ABAB AI Insight
Jensen Huang's assessment of physical AI continues Nvidia's platform expansion path over the past two decades. The company first transformed GPUs from graphics chips into general-purpose parallel computing platforms using CUDA, then became a core vendor for generative AI with deep learning training infrastructure; in the autonomous driving field, it has invested over a decade, accumulating vehicle computing, sensor processing, and simulation capabilities through the Drive platform in collaboration with automakers. The current nearly $10 billion physical AI business is not a bet starting from zero, but rather a unification of existing automotive, robotics, and edge computing revenues into the "physical AI" narrative, pushing towards a $100 billion business goal.
The capital path extends from cloud training chips to the long-term computing needs of every movable machine. Digital AI primarily consumes data center GPUs and cloud inference resources; physical AI requires not only training but also sensors, real-time control chips, simulation environments, map or spatial data, functional safety systems, engineering integration, and continuous software updates. If an autonomous vehicle or industrial robot enters mass production, Nvidia can sell GPUs at the training end, Omniverse-related software and computing power at the simulation end, and Drive or edge computing platforms at the deployment end; the capital expenditures of the same customer are split into multiple layers and span multiple years.
Historical analogies should look at smartphones rather than chatbots. Smartphones did not win with a single processor but were formed by mobile chips, operating systems, app stores, developer tools, sensors, communication networks, and supply chains; similarly, physical AI will require standardized software and hardware before large-scale deployment can occur. Tesla drives autonomous driving with FSD data loops, Amazon promotes automation with warehousing robots and logistics networks, and industrial robot companies like ABB and Fanuc have accumulated factory scenarios. Nvidia is currently in an expansion phase selling general computing and simulation foundations to these industries, rather than directly replacing automakers or robot manufacturers.
Essentially, this is a reconstruction of the industrial chain. Digital AI changes the marginal costs of knowledge work, content production, and software services; physical AI alters the unit labor and unit downtime costs in tangible industries such as manufacturing, logistics, transportation, agriculture, construction, and energy. The asset scale of the real economy far exceeds that of the software subscription market, thus the potential market could be larger; however, the deployment cycle is also longer, as each mistake could lead to collisions, production halts, personnel injuries, or product liabilities. Nvidia is not competing for one-time robot sales profits but for the pricing power of all training and execution nodes in a data-driven, simulated, and automated physical world.
ABAB News · Cognitive Laws
Digital AI sells efficiency, physical AI rewrites labor costs
Being able to reason on a screen does not equate to being responsible in reality
The value of robots lies not in hardware, but in data loops