From One a Day to One an Hour: Figure, Apptronik, and the Pi0.7 Moment Mark a Turn for Humanoid Robotics

Three signals from the last six weeks deserve to be read together. Figure says its F.03 humanoid has moved from a build cadence of one robot a day to one an hour inside its production facility. Apptronik closed out its Series A at $935 million, putting Apollo on Mercedes-Benz, Jabil, and GXO floors at a reported $5 billion-plus valuation. And Physical Intelligence quietly published Pi0.7, the latest iteration of its general-purpose robotics foundation model. None of them is, on its own, a market-moving headline. Stacked together, they describe a different industry than the one investors were modeling at the start of 2026.

For most of the last decade, humanoid robotics has been a category powered more by demo videos than by units. That stopped being a fair characterization sometime around the first quarter of this year. Counterpoint Research now estimates roughly 16,000 humanoid robots were installed globally in 2025, with cumulative installations projected to clear 100,000 by 2027. ABI Research forecasts 115,000 units shipping in 2027 alone. Goldman Sachs continues to anchor its long-range humanoid TAM at $38 billion by 2035, with shipment estimates revised fourfold higher than its 2023 model. The move that matters in 2026 is not whether the curve exists - it is which companies actually walk down the manufacturing learning curve fast enough to capture it.

Figure's production cadence is the cleanest signal. Public statements and trade reporting around the F.03 line describe a transition over roughly two quarters from a one-robot-per-day proof-of-concept to roughly one robot per hour - implying production capacity in the low thousands per year if held steady, well before the long-promised RoboFab capacity of "hundreds to 10,000-plus per year" comes online. Figure has already validated the F.02 generation with BMW, where the company reported more than 90,000 parts placed and roughly 30,000 X3 vehicles touched over an 11-month deployment window. The story for investors is no longer whether the robots can do the work; it is whether each subsequent unit can be produced for less than the gross margin contribution of its first deployment year of labor displacement.

Apptronik tells a parallel but differently structured story. The Texas-based Apollo developer disclosed in February that its Series A had grown to $935 million after a $520 million extension led by B Capital, Google, Mercedes-Benz, PEAK6, AT&T Ventures, John Deere, and Qatar Investment Authority. Apollo is on the floor at Mercedes-Benz, Jabil, and GXO. Apptronik's near-term commercial focus is logistics and manufacturing, with retail and healthcare flagged as adjacent expansions. Apptronik is also where Google DeepMind's Gemini Robotics work has been most visibly integrated - the partnership announced in December 2024 has deepened, and the resulting stack puts a frontier language-and-vision model directly behind Apollo's planning loop. That hardware-plus-foundation-model integration is the specific architectural pattern more humanoid programs are now copying.

The third signal is upstream. Physical Intelligence's Pi0.7, published April 16, is being positioned by the lab as a "steerable" model with emergent compositional generalization - the ability to combine learned skills into tasks it was never explicitly trained on. NVIDIA's GR00T N1, released earlier in the year, is targeting a similar prize from the humanoid-control side, with a roughly 7-billion-parameter dual-loop architecture optimized for full-body control. Open-source contenders like OpenVLA and SmolVLA show that the cost curve to fine-tune a working policy for a specific robot has fallen to dozens of demonstrations and a single GPU. The same dynamic that made software AI investable in 2023 - a foundation-model layer that absorbs the marginal cost of new task acquisition - is now visibly arriving in robotics.

None of this changes the awkward truth that production economics still bind. Most humanoids in the field today have two-to-four-hour battery runtimes, unit costs in the multiple hundreds of thousands of dollars, and reliability profiles that require human supervision in any non-trivial environment. Tesla's Optimus story compresses that reality. CEO Elon Musk confirmed on Q1 2026 earnings that Optimus 3 production begins at Fremont in late July or August, after a four-month conversion of the Model S/X line, with output that will be "quite slow" and 10,000 unique parts that have never been mass-produced. By Musk's own admission in January 2026, Tesla had zero Optimus units doing useful factory work after a year of trying. China's BYD, Agibot, and Unitree are publicly targeting tens of thousands of annual units, but verified deployments at scale remain concentrated in their own captive supply chains.

The investable structure that emerges from this is layered. The OEM platform layer - Figure, Apptronik, Agility, 1X, NEURA, plus Chinese counterparts - is consolidating around a small number of credibly funded names. NEURA's reported €1 billion round backed by Tether at a roughly €4 billion valuation, and Apptronik's path from a $415 million initial Series A to a $935 million extension in less than a year, both illustrate how decisively late-stage capital is now picking winners. The foundation-model layer - Physical Intelligence, Skild AI, NVIDIA GR00T, Google's robotics stack, Generalist - is where the leverage on data and compute compounds, but where commercial revenue today is mostly research-grade. And the component layer - actuators, tactile sensing, on-device compute, batteries - is where deployment-grade reliability gets won or lost, and where Asian supply chain incumbents currently set the pace.

For research purposes, the cleanest near-term watch list has four items: 1) realized 2026 unit production from Figure, Apptronik, Tesla, Agility, and Unitree, against publicly stated targets; 2) attach rate of foundation-model fine-tuning programs (GR00T, Pi, OpenVLA derivatives) inside named OEMs; 3) bill-of-materials trajectory toward the sub-$50,000 unit cost that Goldman's 2024 revision implicitly assumes; and 4) the IPO and secondary signals around the named platform leaders, which is where private-market valuations will get tested against public comparables.

Educational takeaway. The AI Robotics work AdValorem Research publishes is built around exactly this kind of layered framework. The current research cycle treats humanoid platforms, physical AI foundation models, and the hardware supply chain as three distinct lenses, and tracks the milestones above on a rolling basis. As pilots convert to production and as foundation models start absorbing the marginal cost of new tasks, the analytical question stops being "do humanoids work?" and starts being "which layer captures the cash flow when they do?" That is the thread our AI Robotics topic series will keep pulling on through the rest of 2026.