The pitch is getting louder. SoftBank's Masayoshi Son told CNBC this week that physical AI is where he expects the next trillion-dollar company to come from. Wedbush's Dan Ives called humanoid robots "the golden goose for physical AI." Goldman Sachs has just sextupled its 2035 market forecast to $38 billion. In South Korea, four robot-adjacent large caps are up an average 155 percent year to date. Almost every part of this story is bullish. One part is not, and it sits squarely on top of the volume leaderboard.
Start with what's actually changed. The training pipeline for humanoids has moved decisively to sim-to-real transfer: a robot's behaviour is learned inside a physics simulator, then deployed onto hardware with no further tuning. Agility Robotics' Digit runs on a sub-million-parameter network trained inside NVIDIA's Isaac Sim, and is now moving totes for Toyota in Ontario under a Robots-as-a-Service contract priced around $30 an hour. Figure AI's Figure 02 spent eleven months at a BMW plant in South Carolina, helping assemble more than 30,000 cars. Boston Dynamics has reportedly sold out its 2026 Atlas production already, with first units bound for Hyundai. Unit prices have fallen from roughly $200,000 in 2023 to between $40,000 and $60,000 today, a slope steep enough to change the spreadsheet for any operations director in a high-wage market.
Then there's the volume picture, which is where the story turns. Chinese manufacturers accounted for 85 percent of humanoid installations last year, by Barclays' count, and produce robots at roughly half the cost of Western competitors. Unitree Robotics, based in Hangzhou, shipped around 5,500 units in 2025 and is aiming for up to 20,000 this year. On June 1, NVIDIA announced that its first publicly available humanoid research platform pairs Unitree's H2 chassis with NVIDIA's own Jetson Thor hardware and Isaac GR00T software, and is being shipped to academic labs including Stanford and UC San Diego. As industrial partnerships go, this is the marquee deal of the year so far.
It is also the deal that exposes the asterisk. Under Article 7 of China's National Intelligence Law, every Chinese company is required to support and cooperate with state intelligence work, regardless of where its servers sit or what its privacy policy says. That is the operative law of Unitree's home jurisdiction; it is not something a contract can override. Congressional concern predates this week's announcement: 24 House members called for an investigation last May, the Homeland Security Subcommittee held a dedicated hearing in March, and Senators Rick Scott and Tom Cotton introduced the Blocking CCP Spy Tech Act of 2026, which names Unitree specifically. Security researchers have already documented a backdoor (CVE-2025-2894) in Unitree's Go1 robot dogs that streamed sensor data to servers in China, and separate research has identified undisclosed telemetry from the G1 humanoid. Unitree denies wrongdoing and says newer hardware is more secure. No independent audit has confirmed this.
NVIDIA's response is technical: software updates will route through its Blackwell chip, where code is authenticated before execution, and similar reference platforms are planned with US, European, and South Korean partners that NVIDIA declined to name. It's a reasonable mitigation strategy. It also concedes that there is something specific to mitigate.
The interesting tension is not whether humanoids will scale, because the commercial deployments already exist. It's whether the West can buy into the cost curve without buying into the jurisdictional risk. BYD, fresh off announcing its own humanoid programme, plans to sell through its dealer network. Generalist AI just raised $400 million to build robot-agnostic foundation models. Tesla, 1X Technologies and Figure are building US production lines. Several of those, however, depend on Chinese-made components and on training data harvested by an industry of "robot puppeteers" earning a thousand dollars an hour at firms like Encord. The supply chain is genuinely global, and the law that applies to one major node of it is not.
For investors, this is a story worth taking on its own terms. The market is real, the engineering progress is concrete, and the cost compression is documented. But the volume leader operates under a statute no purchase order can rewrite, and most of the standards bodies that would normally adjudicate safety (ANSI for force output, ISO 25785-1 for bipedal walking) have not yet caught up with what is now shipping. The trillion-dollar pitch is plausible. So is the asterisk.