The US Physical AI Wave and What It Means for a Spanish Factory

Between 2025 and 2026 the conversation about artificial intelligence moved from the screen to the machine. This analysis sets out what is happening in the United States with physical AI, with a source and a date for every figure, separates what is already proven on a real plant floor from what is still an announcement, contrasts it with European and Spanish figures and with the regulatory calendar that applies here, and ends with the only actionable part for an industrial SME: which decisions make sense in the next twelve months and which do not.

What is happening in the United States

It pays to pin down the term before reading the figures. Physical AI is AI that perceives, reasons and acts on the real world. In a factory it starts by capturing the data of the equipment that already exists properly. The official NVIDIA glossary (consulted in September 2026) frames it from the system side: perceive, understand, reason, and perform or orchestrate complex actions in the physical world. It is the same idea seen from two different altitudes, and the difference between them explains much of what follows.

The US wave is not a single sector. It is four layers moving at different speeds, and they should not be mixed when reading a headline:

• Compute and world models. A world model is a model that learns the dynamics of the environment: given a state and an action, what the next state is and what is expected to be observed. The NVIDIA paper accompanying the Cosmos platform (arXiv 2501.03575, 7 January 2025) defines a world foundation model as a general-purpose world model that can be fine-tuned into customized world models. The platform was presented at CES on 6 January 2025 and its third generation, Cosmos 3, was announced at GTC Taipei on 31 May 2026, with a mixture-of-transformers architecture and the weights of the Super and Nano variants published on Hugging Face under the OpenMDW-1.1 licence (NVIDIA Newsroom).
• Policy models. A policy is the function that turns what the system perceives into the action it executes. Here the competition is among labs training generalist policies: Physical Intelligence, with a public model cadence running from pi-0 in October 2024 to pi-0.7 on 16 April 2026 (company blog, consulted in September 2026); Skild AI, whose co-founder and CEO, Deepak Pathak, states that their model can control robots it has never trained on (The Robot Report, 15 January 2026); and Google DeepMind, which separated action from reasoning in Gemini Robotics 1.5 and ER 1.5 (25 September 2025) and added whole-body control and an on-device variant in Gemini Robotics 2 (30 July 2026).
• Hardware integrators. Figure AI, Agility Robotics, Boston Dynamics and Tesla. It is the most visible layer and the one concentrating the capital, and also where the distance between demonstration and deployment is widest.
• Industrial users. BMW, Amazon and Foxconn in the United States, Siemens in Europe. This is the layer that decides whether any of this leaves the lab, because it is the only one that brings the real context: material variability, tolerances, shifts and responsibility for the safety of people.

NVIDIA is also the meeting point of the four layers. It presented Isaac GR00T N1 at GTC on 18 March 2025 as an open foundation model for humanoids, alongside the open physics engine Newton and with partners such as Google DeepMind, Boston Dynamics, 1X, Agility, NEURA and Intrinsic (NVIDIA Newsroom). In March 2026 its CEO, Jensen Huang, argued that physical AI has arrived and that every industrial company will also become a robotics company (NVIDIA Newsroom, 16 March 2026). And on 31 May 2026 the company released an open humanoid robot reference design for academic research, combining a Unitree H2 Plus chassis, Sharpa Wave hands and Jetson AGX Thor T5000 compute, with availability planned through Unitree in late 2026 and users such as Ai2, ETH Zurich, the Stanford Robotics Center and UC San Diego.

Capital follows. Four deals are enough to give the order of magnitude, and all four are public:

Those figures read better against a contrast. Agility Robotics is today the only humanoid manufacturer with public accounts, because of its listing: it reported $1.8 million in net revenue and $140 million in operating loss for 2025, with more than 65,000 operating hours for Digit across nine customer sites and more than $300 million in multi-year orders for version 5 (The Robot Report, 7 September 2026). The distance between valuation and actual revenue in the sector is therefore several orders of magnitude. At the other end of the horizon, Morgan Stanley Research titles its humanoid robot market analysis with a figure of $5 trillion by 2050 (report dated 29 April 2025). A twenty-five year forecast is not a budgeting basis: it is a signal of where capital is pointing.

What is proven and what is announced

The costliest confusion when following this wave is treating a documented deployment on an active line, a two week proof of concept and an announcement with no schedule as equivalent. The table below separates the three, actor by actor. Every figure for hours, parts and success rates comes from the vendors themselves or from the users, and none is audited by an independent third party.

Three readings hold on that table. First: the demonstrated tasks are internal logistics and loading and unloading, not process work. Unstacking totes, sequencing parts for ordered delivery or loading sheet metal into tooling are millimetre-tolerance, second-scale jobs, not tenth-of-a-second ones. Second: the time horizon of the trials is short. Ten months on an active line is the longest deployment documented in the table, and the vendor is the one declaring it. Third: whoever deploys already had an instrumented plant before the robot arrived. That is the point that matters to a Spanish factory, and it orders the rest of this analysis.

Why the conversation moved from software to the machine

This is not a change of market mood. There are three chained technical reasons, and they are worth understanding because they determine which part of all this transfers to a plant.

First: the architecture transferred. In 2023, the RT-2 work from Google DeepMind (arXiv 2307.15818, 28 July 2023) defined the vision-language-action model, that is, a policy that takes images and a natural language instruction and emits low level actions directly, expressing the actions as text tokens and incorporating them directly into the training set. That made it possible to reuse in robotics the same pretrained vision and language models that already worked on screen. From there the family grows: pi-0.5 from Physical Intelligence generalises to the open world using co-training on heterogeneous tasks (arXiv 2504.16054, 22 April 2025) and GR00T N1 from NVIDIA adopts a dual system architecture trained on a heterogeneous mixture of real robot trajectories, human videos and synthetically generated datasets (arXiv 2503.14734, 18 March 2025).

Second: physics imposed its own architecture. The dual system is not an aesthetic choice, it is a consequence of latency. Figure describes Helix as a unified vision-language-action policy in which the reactive system runs at 200 hertz and semantic reasoning at 7 to 9 hertz, trained on around 500 hours of teleoperation (Figure AI, 20 February 2025). Translated into times: the reactive loop decides every 5 milliseconds and the layer that understands the scene revises the plan roughly every 110 to 140 milliseconds. That split is recognisable in any plant. Common cycle times in periodic industrial communication run from 10 milliseconds to 0.5 milliseconds, according to the 5G-ACIA and ZVEI white paper on integrating industrial Ethernet networks with 5G networks (November 2019). In other words, the reactive loop of a humanoid lives in the same time range as the control loop of a machine, and reasoning lives two orders of magnitude above it. A system trying to close the fast loop through a remote platform does not fail on bandwidth: it fails on the variance of the delay.

Third: data stopped being a wall and became the work. Open X-Embodiment (arXiv 2310.08864, 13 October 2023) aggregated data from 22 different robots collected through a collaboration between 21 institutions, demonstrating 527 skills, and showed positive transfer across morphologies. Teleoperation here means real robot trajectories guided by people, recorded with what was observed and what was done at every instant. That observation and action pair is exactly what a conventional plant history does not have: there are process variables, but no record of what an operator changed or what resulted. With purely observational data you can learn what usually happens, not what happens if you intervene.

One limit remains that none of the three reasons solves, and it is why these demonstrations are slow to scale. The gap between simulation and reality is defined in the literature as the abstractions and approximations that inevitably introduce discrepancies between simulated and real environments (arXiv 2510.20808, 23 October 2025, published in the Annual Review of Control, Robotics, and Autonomous Systems 2026). It is not only a matter of visual appearance: friction, inertias, backlash, contact deformation and sensor delay weigh more. The canonical technique for crossing it, domain randomization, means training models on simulated images that transfer to real images by randomizing rendering (Tobin et al., arXiv 1703.06907, 20 March 2017). It works when reality falls inside the trained family of parameters, and centring that family requires measured data from the real machine. A photorealistic simulator with the wrong friction still fails.

The European and Spanish contrast

The starting point matters more than the speed. In 2024, 542,000 industrial robots were installed worldwide and the operational fleet reached 4,664,000 units, according to the International Federation of Robotics (World Robotics 2025, 25 September 2025). The split is uneven: Asia took 74% of new installations, Europe 16% and the Americas 9%. Europe fell 8% to 85,006 units, and Germany, the leading market on the continent, fell 5% to 26,982.

Spain installed 5,086 industrial robots in 2024, up 1%, and overtook France as the third largest European market behind Germany and Italy, which recorded 8,783 units (IFR and VDMA Services, World Robotics 2025 executive summary, September 2025). Average world density in manufacturing was 177 robots per 10,000 employees, with Asia at 204, Europe at 148 and the Americas at 131 (same summary). At the top end, the Republic of Korea records 1,220 robots per 10,000 employees, followed by Singapore with 818, Germany with 449, Japan with 446 and the United States with 307 (IFR, 8 April 2026). The executive summary does not publish the Spanish density in that same breakdown, so no figure for Spain is given here: giving one without a source would be inventing it.

AI adoption draws a similar map. 20.0% of EU enterprises with ten or more employees used AI in 2025, up from 13.5% in 2024, with Denmark at 42.0%, Finland at 37.8% and Sweden at 35.0%, and Romania at 5.2% at the low end (Eurostat, 11 December 2025). The Spanish figure comes from another source: the statistics office INE puts companies with ten or more employees using AI at 21.1% (22 October 2025) and Fundación Cotec notes that industry stops at 17.5% (same date). In other words, industry adopts less AI than the economy as a whole.

The figure that orders the decision most, however, is a different one. According to Metalindustria (22 May 2026), reporting on the third Barometer of industrial digitalisation and automation in Spain presented at Advanced Factories 2026, only 3.3% of factories describe themselves as fully digitalised, and the same barometer puts plants with full system integration at 22.4%. With that starting point, the relevant question for a Spanish plant is not whether to buy a humanoid: it is whether the data from the machines it already owns leaves the equipment with time, unit and context. There is organised movement in that direction: the Spanish humanoid robotics and physical AI ecosystem has grouped into the ÁNIMA association, driven by AFM Cluster and AER Automation, which together represent more than a thousand companies (Automática e Instrumentación, 4 September 2026).

The European regulatory framework and its dates

A European factory does not compete under the same framework as a US one, and that changes the investment calendar. Four dates govern the planning:

The AI Act entered into force on 1 August 2024 and rolls out its obligations in stages: general provisions and prohibitions from 2 February 2025, general purpose AI and governance rules from 2 August 2025, and new prohibitions on synthetic content from 2 December 2026 (European Commission, AI Act Service Desk, consulted in September 2026). The high risk deadlines were postponed by the Digital Omnibus package, and one law firm reading further holds that AI embedded in products covered by the Machinery Regulation would be largely exempt from AI Act obligations (Gibson Dunn, 27 May 2026). That claim is not confirmed in an official source and should not underpin a purchase decision.

As for Machinery Regulation (EU) 2023/1230, adopted on 14 June 2023, there is an industry proposal to postpone requirements, but the February 2026 position withdrew the request concerning the artificial intelligence provisions and is limited to cybersecurity, and the source itself considers the postponement rather unlikely (IBF Solutions, updated 19 March 2026). The operational conclusion is to plan with 20 January 2027 as a firm date.

On top of that sit the technical standards already published, which define how the safety and interoperability of all the above is demonstrated. ISO 10218-1:2025 and ISO 10218-2:2025 replaced the 2011 editions of the safety requirements for industrial robots in February 2025, the first addressed to the robot manufacturer and the second to the cell integrator (ISO). The revision absorbs the content of technical specification ISO/TS 15066, which defined the four collaborative operation modes and the biomechanical limits per body part (The Robot Report, 18 February 2025). For interoperability, companion specification OPC 40010-1, version 1.02, covers the vertical integration of motion device systems with production management systems (OPC Foundation, released 8 September 2025). For time, IEEE 1588-2019 defines precise clock synchronisation in measurement and control systems, and the time-sensitive networking profile for industrial automation IEC/IEEE 60802 was published on 29 June 2026 (IEEE 802.1). And for cybersecurity, the ISA/IEC 62443 series sets the requirements for industrial automation and control systems, with edition 2.0 of IEC 62443-2-1, from August 2024, devoted to the asset owner security program (IEC).

What can be copied and what cannot

The useful question is not whether an industrial SME can do what Figure does. It is which part of the method survives the change of scale. What can be copied, and is also cheap:

• Recording the observation and action pair. What makes teleoperation data valuable is not volume, it is that every instant carries what was observed, what was done and what resulted. A plant can start today recording every intervention (a setpoint change, an adjustment, a forced stop) together with the state that motivated it and the effect measured afterwards. Without that record there is no way to learn from the plant itself.
• The discipline of the timestamp. Synchronising clocks and measuring the deviation is the condition for ordering events causally across different equipment, which is what lets you answer what tripped first when a failure cascade lasts less than a second.
• The split into two speeds. What Helix solves with two systems, a plant solves by deciding where each decision lives: the deterministic loop in the cabinet, supervision at the edge, shift optimisation on the platform and training in the cloud, never on the critical path of a fast decision.
• Evaluation on unseen cases. The test of a learned system is not the demonstration, it is its performance on a held-out set of cases that were not in the training data, with a safety envelope that does not depend on the model. That protocol can be required of any vendor, on any project and at any scale.
• Calibrating simulation with real data. If a twin is built or a line is simulated, its parameters must be identified with measured, synchronised data from the machine. Without that, the resulting gap is blamed on the simulator when it is an identification problem.

And what cannot be copied, with the figure that explains it:

• The scale of capital. A $1,400 million round or a $39,000 million valuation cannot be replicated with an industrial investment plan. Copying the spend without the return thesis is the fastest route to an abandoned pilot.
• The volume of demonstration data. Around 500 hours of teleoperation for a single family of tasks, or the aggregation of 22 robots and 21 institutions in Open X-Embodiment, are out of reach for a plant. The advantage of an SME is not volume: it is the specificity of its process and the ability to narrow the task sharply.
• Tolerance for loss. $140 million of operating loss in a single year (The Robot Report, 7 September 2026) is a way of funding learning that an industrial company does not have. Each rung has to pay for itself.
• Vertical integration. BMW can create its own centre of competence for physical AI in production and Amazon can pilot at one facility and deploy across more than 300. An SME cannot amortise an internal capability like that, which is why its decision is about procurement and data governance, not development.

What to decide in the next twelve months

Six decisions worth taking now, all verifiable and none dependent on the humanoid wave maturing:

1. Name a data owner and write the signal dictionary. With asset, unit, range, sampling rate and data quality per signal, and versioned, dated changes. The most common reason a three year history turns out to be useless is not technical: someone renamed or rescaled a tag and it was never recorded.
2. Synchronise the clocks and measure the deviation. Tens of milliseconds are enough for OEE, energy and shift analysis; one millisecond or better is needed to order events causally across equipment, and that is where IEEE 1588-2019 comes in.
3. Replace polling with change of state capture where downtime is counted. An eight second micro-stop disappears entirely if machine state is polled every thirty seconds, and with it the largest availability loss on many lines.
4. Write a single definition of OEE. Of planned time, of nominal speed per product and of what counts as a reject. Without it the indicator is not debatable, it is merely arguable, and meetings are spent negotiating the denominator.
5. Close one supervised loop and measure it. A bounded action, with an identified owner, a tested rollback path, a declared maximum age for the data that justifies it and an effect measured against a baseline. One done properly teaches more than five done halfway.
6. Put the regulatory calendar into the investment plan. 20 January 2027 for the Machinery Regulation and 2 August 2028 for AI embedded in a regulated product are design dates, not compliance-after-the-fact dates.

And five decisions not worth taking yet:

• Buying a humanoid for a plant that still writes downtime on paper. The system needs data with time and context in order to work and for anyone to notice when it degrades. Without that, the pilot works in the demonstration and dies in production with no diagnosis.
• Budgeting on a 2050 market forecast. It is useful for understanding where capital is heading, not for sizing a 2027 investment.
• Closing a millisecond loop through a remote platform. With industrial cycles of 10 to 0.5 milliseconds (5G-ACIA and ZVEI, November 2019), the problem is not bandwidth but the variance of the delay, and the symptom is wrongly blamed on the algorithm.
• Replacing the machine fleet to get AI-ready equipment. The existing fleet already perceives and already acts. What is missing is getting the data out with the protocols it already speaks.
• Signing a pilot with no written success criterion. With no held-out set, no declared baseline and no stopping condition, the pilot can neither be cancelled nor renewed on any rational basis.

The ladder as a decision instrument

The Physical AI Ladder orders all of the above into five levels defined by what the plant can do with its data, not by the technology it buys. It serves two purposes: placing the plant today, and knowing what must be solved before investing in the next rung.

Read against the US wave, the ladder says something concrete. Everything demonstrated at level 4 consumes what the three previous levels produce: level 1 produces reliable perception with time and context, level 2 produces the ground truth to evaluate against, and level 3 produces the action and effect pairs without which a model cannot be corrected nor its degradation detected. That is why jumping straight from level 2 to level 4 is the costliest failure mode seen in industrial projects: the demonstration works and the deployment does not, and with no record of action and effect there is no way to know why.

Captia covers levels 1 to 3: capturing data from existing equipment with Captia Connect, real-time indicators and alerts with Captia.ai, and the supervised closed loop through rules, workflows, Energy and Service. Level 4 is the state of the art of the sector and this piece describes it as such, in the third person. The foundations, by contrast, pay for themselves: the base supporting the only in-house result we publish, over 30% energy savings in energy-intensive companies, is exactly the connected and contextualised data of level 1.

Frequently asked questions

For the full framework, the reference guide is what physical AI in manufacturing is and why it starts with data, and the architecture on which levels 1 to 3 are built is described in the industrial data platform. If you want to place your plant on the ladder using your own equipment, tell us what you have installed.