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.
Two events this week, separated by two days, tell the same story from different angles. Figure AI officially retired its F.02 humanoid after a year on the BMW production line - the first time a robot has left service with a verifiable work record, not because it failed, but because its successor is ready. And Unitree Robotics priced its IPO on the Shanghai Stock Exchange at approximately $9 billion, making Physical AI publicly tradeable for the first time at this scale. These are not coincidences. They are two faces of the same transition: an industry that has moved from asking for patience to asking for a position.
Stats:
| Value | Description |
|---|---|
| 90,000 | Sheet metal elements loaded by Figure F.02 before retirement |
| $9B | Unitree IPO valuation on the Shanghai Stock Exchange |
| $14B | Skild AI valuation after 7 months and one funding round |
| $11B+ | New Physical AI capital or public valuation created in the week of August 12–14 |
The First Humanoid With a Retirement Record
Figure AI officially retired F.02 after nearly a year on the BMW Spartanburg production line. The numbers: over 30,000 BMW X3 assembled, over 90,000 sheet metal elements loaded at 99%+ accuracy. F.02 is not being replaced because it failed. It is being replaced because Figure 03 - produced at BotQ at one robot per hour and already past 1,000 units - is ready.
Retiring a robot based on a successor being ready is a product cycle signal, not a failure signal. Consumer electronics, automotive, semiconductors: every mature industry retires products when successors are ready. Physical AI had never done this before. F.02's retirement is the first time a humanoid has left service with a resume rather than a write-off.
Boston Dynamics confirmed first deliveries of Atlas to Hyundai RMAC and Google DeepMind this week, with the entire 2026 production already committed. Two customers, two different models of what they want from the same hardware: Hyundai RMAC is building operational scale; DeepMind is collecting training data for Gemini Robotics. The same robot used simultaneously to scale production and to scale intelligence, in the same year it first shipped.
When an industry retires products on a cycle rather than abandoning pilots, operational data becomes the primary competitive asset. Every F.02 hour is training data for F.03. The companies without deployed products have no equivalent to iterate on.
The combination of Figure's production data and DeepMind's foundation model research builds an iteration loop that most competitors cannot replicate without their own deployments. The gap between companies with operational data and companies without it is not measured in months. It is measured in model generations.
Physical AI Has a Stock Ticker Now
Unitree Robotics priced its IPO on the Shanghai Stock Exchange at approximately $9 billion, becoming the first humanoid robotics company to go public at this scale. The listing opens Physical AI to retail investors and index funds that previously had no access to the category outside private venture capital.
The structural consequence is precise: public markets impose quarterly operational transparency that private companies do not face. Unitree will now report metrics - unit shipments, revenue, margin - that the broader Physical AI industry has not been required to disclose. The first earnings call will be the most-watched data release in robotics in years.
For the competitive landscape, the IPO matters beyond Unitree. Pension funds, sovereign wealth funds, and retail investors can now take positions in Physical AI through a public vehicle. Once one company is public, the pressure on competitors to match that capital access increases. Masayoshi Son said this week that Physical AI and robotics will produce the next trillion-dollar company. In a week that created over $11 billion in new capital or public valuation, the debate is less about whether such a company will exist and more about which company it will be.
Two Theories of How Physical AI Intelligence Scales
The most significant non-IPO capital event of the week: Skild AI closed a $1.4 billion round, tripling its valuation to $14 billion in seven months. Skild is not building a humanoid robot. It is building a horizontal AI control platform: one foundation model designed to operate across any robot hardware, not optimized for a single platform.
Skild's thesis is that Physical AI will converge like cloud computing: the industry will settle on one or two AI infrastructure providers that all hardware platforms run on top of, rather than each manufacturer maintaining a proprietary control stack. $14 billion in seven months is investors pricing that thesis at a premium.
Apptronik closed $520 million in Series A Extension funding to accelerate Apollo 2 production. Apptronik's model is the opposite of Skild's: deep integration between a specific hardware platform and a specific AI research partner, with Robot Park providing the iteration infrastructure. The bet is that the integration between hardware data and model capability produces advantages that cannot be licensed from a third party.
GrayMatter Robotics adds a third data point: 30 million square feet of production floor, 20 industries, 12x productivity versus skilled human labor in finishing operations - grinding, painting, sealing. GrayMatter does not build humanoids. It builds narrow Physical AI systems with domain-specific depth that general platforms cannot match in specialized niches.
Three companies. Three architectures. All raising significant capital in the same week. The market does not yet know which theory is correct and is funding all three simultaneously. The resolution will come from customer behavior, not from analyst models.
What to Watch Next
- Unitree first earnings disclosure: the first public financial report from a humanoid company will set the operational data reference for every Physical AI valuation conversation through end of 2026
- Skild AI first named deployment: a customer announcement would confirm whether the horizontal OS thesis is translating from research to production
- Figure 03 monthly output at BotQ: whether the 1-robot-per-hour rate is scaling or holding flat determines the production credibility of the F.02 succession narrative
- Atlas at Hyundai RMAC: the first performance metrics from Atlas in a production environment, separated from DeepMind's research use
- GrayMatter revenue disclosure: 30 million square feet and 12x productivity is the claim - revenue would confirm whether the market is paying for it at scale
FAQ
Q: What does it mean that Figure AI "retired" F.02?
Retiring a product because a successor is ready is what mature product businesses do. F.02 was replaced because F.03 is ready, not because F.02 failed. The significance is that the category now has its first example of a robot leaving service with a documented operational record: 30,000 BMW X3 assembled, 90,000 sheet metal elements loaded at 99%+ accuracy. Every future robot will be evaluated against that baseline. The industry now has a standard for what "a robot that completed its deployment" actually looks like - and that standard is a data record, not a spec sheet.
Q: Why does Unitree's IPO matter beyond Unitree itself?
Before the IPO, Physical AI investment was exclusively a private market asset class. Pension funds, index funds, and most institutional investors had no vehicle for Physical AI exposure. Unitree's $9 billion listing creates a public entry point and sets a valuation benchmark that every private Physical AI company is now implicitly compared against. The more consequential change is transparency: public markets require quarterly disclosure of shipment volumes, revenue, and margin. This will be the first regular stream of operational data from a humanoid manufacturer that the industry has never had to report - and investors, customers, and competitors will all read it.
Q: Skild AI at $14B versus Apptronik's vertically integrated model - which thesis wins?
Both are defensible and neither is obviously wrong. Skild's horizontal OS bet mirrors the way operating systems won in personal computing and cloud: the control layer is sticky, platform-agnostic, and benefits from network effects across many hardware deployments. Apptronik's vertical integration bet mirrors the way Apple won in smartphones: tight hardware-software integration produces performance that horizontal platforms cannot replicate at the same quality level. The Physical AI industry is early enough that both models can succeed in different market segments. The clearest resolution will come from customer behavior: if hardware manufacturers license Skild at scale, the horizontal thesis is working; if customers pay a premium for the Apptronik-DeepMind integrated stack, the vertical thesis is working. In August 2026, investors are funding both answers simultaneously.
Top comments (0)