AI News · industry
Physical AI in 2026: The State of Humanoid Robots
China now ships roughly 90% of the world's humanoid robots, while US labs lead the AI brains that run them. Here is the sourced global map, and why the real story is the software, not the hardware.
Physical AI, the effort to give robots a general-purpose "brain" that can see, understand plain language, and act, is having its breakout year in 2026. Two facts define the moment. China now ships roughly 90% of the world's humanoid robots while US companies account for under 5%, per the Information Technology and Innovation Foundation. Yet the highest-value layer, the vision-language-action foundation models that actually control these machines, is still led by American labs. The humanoid robot market reached about $6.24 billion in 2026, up from $4.89 billion in 2025, on its way to a projected $165 billion by 2034, according to Fortune Business Insights. This post maps who ships, who builds the brains, and whether the money makes sense.
What is "physical AI," and why is 2026 its breakout year?
Physical AI is the application of foundation-model techniques to the physical world. Instead of scripting every joint movement, engineers train one large model on a huge variety of physical experience and let it learn common sense about how objects behave and how tasks decompose into motions. The technical term is a vision-language-action model, or VLA: feed it camera images plus a plain-English instruction, and it outputs motor commands directly.
2026 is the breakout year for three reasons that arrived together. First, the brains got good enough to generalize across tasks and bodies, not just replay one scripted demo. Second, hardware got cheap and abundant, with China's Ministry of Industry and Information Technology projecting national humanoid output to exceed 100,000 units this year. Third, capital poured in: global robotics venture funding surpassed $55.8 billion by mid-2026, already breaking prior full-year records.
Who is actually shipping humanoid robots in 2026?
By units, China is not just ahead, it is the overwhelming majority of the market. ITIF, citing Visual Capitalist shipment data, reports that Chinese firms accounted for about 90% of global humanoid shipments in 2025 while US firms sold under 5%. The per-company gap is stark: major Chinese makers sold roughly 12,868 humanoids in 2025, versus about 450 for the leading US names.
| Company | Country | 2025 humanoid units sold | Notable in 2026 |
|---|---|---|---|
| Unitree | China | ~5,500 (G1) | H2 with NVIDIA Jetson Thor; sub-$6K R1 developer unit |
| AgiBot (Zhiyuan) | China | ~5,168 (A2) | Passed 15,000 cumulative units by June 2026 |
| UBTech | China | ~1,000 (Walker) | Walker S2 mass production for auto plants |
| Figure AI | US | ~150 | Figure 03 crossed 1,000 units built at its BotQ line |
| Agility Robotics | US | ~150 | Digit logging real warehouse operating hours |
| Tesla | US | ~150 | Optimus, driven by FSD-derived networks |
Source: unit figures from ITIF, July 2026; 2026 milestones from AI Business Weekly.
The gap has become a policy issue. In late July 2026 the United States banned Chinese-made humanoid and quadruped robots on national-security grounds, a signal of how strategic this category has become. The historical irony, ITIF notes, is that the US invented industrial robots but ceded production leadership years ago to engineering powerhouses in Germany, Japan, and Switzerland, and now to China.
Why the value is the "brain," not the body
Here is the reframe that matters for anyone tracking AI: the humanoid body is becoming cheap, commoditized hardware, and the durable value is the software brain that runs it. Crucially, that brain is deliberately being designed to outlive whatever body it is poured into, a cross-embodiment bet that mirrors the software-side argument for why you should not build your business on a single AI model. The same open-versus-closed, US-versus-China split we mapped for the best LLMs of 2026 is now playing out in robotics, and the US leads this layer.
| Robot brain | Lab / country | License | What it is |
|---|---|---|---|
| Isaac GR00T 1.7 | NVIDIA, US | Open (Apache 2.0) | Called by NVIDIA "the first open, commercially usable VLA model for generalized humanoid robot skills," paired with Cosmos world models and Jetson Thor chips |
| Gemini Robotics 2 / ER 2 | Google DeepMind, US | Closed | A VLA plus an embodied-reasoning model that plans multi-step physical tasks |
| π0.5 (pi-0.5) | Physical Intelligence, US | Research/open | A VLA "flow" model built on a vision-language backbone with a dedicated action expert, designed to transfer across robot bodies |
| Skild Brain | Skild AI, US | Closed | An "omni-bodied" model trained on trillions of simulated experiences plus billions of internet videos of humans |
| GO-1 | AgiBot, China | Open dataset | A VLA trained on the open AgiBot World manipulation dataset |
Model versions current as of August 2026; this field reorders quickly.
NVIDIA is playing the shrewdest hand: rather than build the winning robot, it sells everyone the tools to try, releasing Isaac GR00T 1.7 under Apache 2.0 so any maker can fine-tune it. Google DeepMind brought its frontier lab to bear with Gemini Robotics 2, extending the same multimodal Gemini into a robot controller. Startups like Physical Intelligence and Skild AI sell the brain and nothing else, unbolted from any particular body. A clear pattern: NVIDIA, Google, Figure, Tesla, and AgiBot are each building intelligence meant to be reusable, not a one-off for a single machine. (For a plain-English primer on how VLA models work, this technical explainer is a good starting point.)
Is physical AI a bubble? The honest numbers
Being current means being honest about scope, and the honest read is that finance has run ahead of the business. Annual investment in humanoid robotics reached about $4.3 billion in 2025, against roughly $440 million in actual humanoid sales that year, per Bank of America and IDC figures cited in a widely shared market analysis. Of the roughly 18,000 humanoids shipped in 2025, more than 85% reportedly went to performances, education, data collection, and technology validation rather than productive factory or logistics work.
Valuations tell the same story. Figure AI alone is valued near $39 billion, and Figure, Apptronik, and Agility combined are worth roughly $47 billion, about 107 times the entire industry's estimated 2025 revenue. The long-range forecasts are enormous but distant: Goldman Sachs projects a $38 billion market by 2035, while Morgan Stanley models a market above $5 trillion by 2050. Small changes in adoption timing or robot lifespan swing those 25-year numbers wildly.
The reality check is useful, not dismissive. A few deployments have moved past theater: Figure completed an 11-month production assignment at BMW, Agility's Digit has logged substantial operating hours, and Unitree has delivered thousands of machines. The market becomes credible when customers publish fleet economics (uptime, intervention frequency, cost per completed task) rather than suppliers publishing demonstrations. The robots will likely survive the correction. Many of today's valuations will not.
What physical AI shares with software AI agents
Physical AI is far-field for most businesses, and it will not answer your Instagram DMs next quarter. But it is worth noting that roboticists independently converged on the same bets that already run in production software agents. Three stand out.
First, a general brain decoupled from the body. Robotics labs are building one reusable intelligence that transfers across machines, the same principle behind Entagl's four agents that share one brain: what one agent learns, the others use, instead of gluing together disconnected point tools. Second, fleets that learn together, so every deployed unit feeds experience back into a shared model. Third, and most important, human-in-the-loop governance. The lesson from the "reality check" above is that a demo is not a deployment, and the same governance problem, keeping an autonomous system safe, auditable, and correctable, is exactly the challenge we covered for software agent protocols and guardrails. AI acts; humans govern. That principle holds whether the agent has a body or lives in your inbox.
FAQ
Who makes the most humanoid robots in 2026?
By shipments, Chinese manufacturers dominate, accounting for roughly 90% of global humanoid robots sold in 2025, led by Unitree, AgiBot, and UBTech, per ITIF. US firms like Figure, Agility, and Tesla produce widely cited platforms but sold under 5% of global units. China's Ministry of Industry and Information Technology projects national output to exceed 100,000 humanoid units in 2026.
What is a vision-language-action (VLA) model?
A VLA is a robotics foundation model that takes in camera images and a plain-language instruction and outputs motor commands directly, without hand-scripted motions. Examples in 2026 include NVIDIA's open Isaac GR00T 1.7, Google DeepMind's Gemini Robotics, and Physical Intelligence's π0.5. The idea is to train one model on broad physical experience so it can generalize to new tasks and even new robot bodies.
Are humanoid robots a good investment right now?
The technology is real and improving, but the finances are stretched. In 2025, investment (about $4.3 billion) ran roughly ten times actual humanoid sales (about $440 million), and most shipped robots went to demos and research rather than paying production work. Analysts expect a valuation correction even as adoption grows, so the durable question is which makers can show real fleet economics, not just impressive demonstrations.
How is physical AI related to the AI agents businesses use today?
They share an architecture and a philosophy, not a use case. Both rely on a general model that perceives, reasons, and acts, both improve as a fleet shares experience, and both need human-in-the-loop guardrails to be trusted in production. The difference is the interface: a humanoid acts in the physical world, while a business AI agent acts across chat, voice, and other digital channels.
Physical AI proves a broader point: the value of AI is the coordinated intelligence, not any single body or tool. If you want that closed-loop intelligence working across your customer conversations, ads, and calls, book a 30-minute demo and we will map it to your business.
Sources: ITIF, Fortune Business Insights, China Daily / Xinhua, AI Business Weekly, Forbes, NVIDIA Developer, Google DeepMind, and market analysis via New Market Pitch. Model versions and figures are current as of August 2026 and will change; reverify before citing.