The photo used in the cover was uploaded from https://electrek.co/2026/09/07/xpeng-iron-humanoid-robot-production-line/.
Physical AI Digest is a weekly briefing produced by Klaudia from Physical AI Company xBerry - a tech company based in Poland building tools at the intersection of Physical AI and operations.
On September 7, XPeng's IRON walked off a production line in Guangzhou on its own. Not out of a lab. Not out of a hand-assembled prototype batch. Off a line - an industrialized production facility running at more than 80% automation, built on quality systems transferred directly from XPeng's electric vehicle manufacturing. The same standards that govern how thousands of EVs roll off a line every day now govern how a humanoid robot does.
Stats:
| Value | Description |
|---|---|
| >80% | Automation level of XPeng's IRON production line, built on EV-grade quality systems |
| $3.5B | Figure AI's compute contract with Nscale: GPU capacity secured as a strategic resource |
| $6B | SoftBank's negotiated valuation for a majority stake in 1X Technologies |
| Dec 2026 | Schaeffler's first NEURA Robotics deployment: humanoids enter CAPEX planning |
The Production Line That Changes the Category
XPeng's official announcement describes IRON as the world's first advanced general-purpose humanoid robot to walk off a production line - and the claim holds on two dimensions simultaneously. First, the robot itself: IRON is not a task-specific manipulator or a wheeled logistics platform. It is a general-purpose humanoid designed for environments built for people. Second, the production method: the line runs at more than 80% automation, with quality control systems imported directly from XPeng's EV manufacturing operations.
That second point is the one that matters most for the industry. XPeng does not build humanoids the way a robotics startup builds prototypes - by hand, in small batches, with custom tooling for each unit. It builds them the way it builds electric vehicles: with standardized processes, automated inspection, and production metrics that track defect rates per thousand units rather than per single robot.
XPeng plans mass production by end-2026 and commercial deliveries in China and internationally in 2027. The delivery timeline is secondary. The production system is primary. A humanoid robot that is manufactured the same way a car is manufactured has a cost curve that follows the same logic as a car - and that cost curve is deflationary over time, not inflationary.
The moment that defines a manufacturing category is not when the first prototype works. It is when the first production line runs. XPeng's IRON line on September 7 is that moment for humanoid robots: not a demo, not a hand-built batch - a line with more than 80% automation, automotive quality standards, and a robot that walked off it autonomously. Every competitor now has a production benchmark to match.
Figure AI's $3.5 Billion Bet on Compute as Infrastructure
On September 3, Figure AI signed a $3.5 billion compute contract with Nscale - securing GPU capacity over a multi-year term to power VLA model training at a scale that its operational deployments now require. Figure 03 is running in BMW factories, accumulating tens of thousands of hours of real-world training data. Processing and learning from that data at the rate it is generated requires a level of compute that cannot be purchased on-demand. It has to be reserved.
The strategic framing is precise: Figure is treating GPU capacity the way power-intensive industries treat energy supply. A steel mill signs long-term electricity contracts because variable pricing exposes it to cost risk that makes planning impossible. Figure signed a long-term compute contract for the same reason. A company that secures compute as a strategic resource rather than an operational cost is signaling that its competitive advantage depends on model iteration speed - and that it intends to be the fastest.
SoftBank Builds a Humanoid Stack: From ABB to 1X
SoftBank is in advanced negotiations to acquire a majority stake in 1X Technologies at approximately $6 billion. The context matters: one year ago, 1X attempted to raise $1 billion at a $10 billion valuation and did not close the full round. SoftBank is now negotiating a majority stake at $6 billion - a 40% discount on the attempted valuation, but control rather than a minority position.
SoftBank already acquired ABB Robotics for $5.4 billion - industrial robots for factories. 1X's NEO targets consumer homes. Combined with SoftBank's ownership of Arm, the emerging structure is a vertical stack: silicon at the foundation, industrial robots for enterprise, consumer humanoids for the home.
SoftBank is betting that NEO works well enough in real homes to retain customers - and that the company that controls the consumer humanoid platform will occupy the same position in Physical AI that Apple occupied in mobile computing.
Atlas at DeepMind, Schaeffler's CAPEX Plan, and the Construction Site
Boston Dynamics confirmed its first commercial Atlas deliveries to Hyundai RMAC and Google DeepMind. DeepMind testing two humanoid platforms simultaneously - Atlas and Apollo 2 - is the clearest possible signal that no single company has yet won the VLA model race.
Schaeffler confirmed its humanoid deployment timeline with NEURA Robotics: first units between December 2026 and June 2027, with a target of 1,000 to 2,000 robots in its global factories by 2032. When a 150-year-old precision manufacturer writes humanoid robots into its CAPEX plan with specific dates and volumes, the technology has crossed from R&D into operations.
LimX Dynamics and ZINOVA demonstrated TRON 2 on a real construction site - mounting formwork, laying rebar, using standard tools. A humanoid that works on a construction site has demonstrated something a factory robot cannot: the ability to operate in unstructured environments designed for humans, not for machines.
What to Watch Next
- XPeng IRON mass production ramp: XPeng targets mass production by end-2026 - the first production metric for a general-purpose humanoid will be units per day, not units per year
- Figure AI Nscale first model improvement cycle: the $3.5 billion compute contract enables faster iteration - watch for the first model update Figure attributes directly to the expanded training capacity
- SoftBank 1X deal close: if majority stake is confirmed, NEO becomes the consumer humanoid platform of a company with the capital to subsidize early adoption
- Schaeffler December 2026 first deployment data: precision bearing manufacturing with micrometer tolerances is among the most demanding quality environments for any robot
- DeepMind VLA platform decision: whichever platform DeepMind commits to for primary research will be identified as the leading whole-body intelligence reference architecture
FAQ
Q: Why does XPeng's automotive-grade production system matter for the humanoid industry beyond XPeng itself?
Because it establishes the production benchmark the rest of the industry now has to match. Before September 7, humanoid robots were manufactured in small hand-assembled batches or early-stage production runs with significant manual intervention. XPeng's IRON line at more than 80% automation with EV-quality control systems demonstrates that the manufacturing problem is solved at one company - and that solution defines what competitive production looks like going forward. Any competitor that cannot reach comparable automation levels will face a structural cost disadvantage that compounds as XPeng scales.
Q: What does Figure AI's $3.5 billion compute contract with Nscale actually buy, and why not just use cloud on demand?
The $3.5 billion contract buys reserved GPU capacity over a multi-year term. The reason not to use cloud on-demand is the same reason a power-intensive manufacturer does not buy electricity on spot markets: variable pricing creates planning uncertainty, and bursting to full capacity on demand during peak periods is either impossible or prohibitively expensive. Figure's commercial deployments generate continuous real-world training data. Processing that data fast enough to produce model improvements faster than competitors requires sustained compute at scale, not burst capacity. A company that waits for on-demand GPU availability during high-demand periods loses iteration cycles.
Q: Why would SoftBank buy a majority stake in 1X at $6 billion when 1X could not raise $1 billion at $10 billion valuation one year ago?
Because SoftBank is not buying at a discount - it is buying control at a price it can justify strategically. One year ago, 1X was seeking capital from financial investors evaluating it as a standalone consumer humanoid bet. SoftBank's calculation is different: it already owns ABB Robotics for the industrial layer and Arm for the silicon layer. 1X at $6 billion for majority control gives SoftBank the consumer layer of a vertical stack it is building deliberately. Control at $6 billion is worth more to SoftBank than a minority position at $10 billion would have been.
Top comments (0)