Robotics: where does the value go in the physical AI era?

What the PC and the smartphone teach us

Two value migrations offer a lens on physical AI. On the PC, the IBM PC's open architecture (1981) allowed third-party cloning (Compaq and others): hardware commoditized while value concentrated with the processor and OS suppliers, Microsoft and Intel, the "Wintel" alliance. On the smartphone, the pattern repeated, worse for manufacturers: in 2023 Apple captured roughly 85 percent of the industry's global profit with only 21 percent of unit volume (Counterpoint Research, Bloomberg), the rest fighting over Android margins built on Qualcomm chips and ARM architecture. The software-defined vehicle is on a similar path: per McKinsey, the software content of a mid-range sedan moves from about 10 percent of value (1,220 dollars) in 2018 to a projected 30 percent (5,200 dollars) by 2030.

Will robotics follow the same script, with NVIDIA as "robotics' Wintel"? The company combines an edge chip (Jetson Thor, Blackwell architecture, up to 2,070 FP4 teraflops, 128GB of memory), an open foundation model (Isaac GR00T) and a near-standard simulation platform (Isaac Sim). TechCrunch summed it up in January 2026: NVIDIA wants to become "the Android of generalist robotics". Boston Dynamics is integrating Jetson Thor into Atlas, Agility Robotics has announced it for Digit 6, Figure AI is building on it. Unlike the PC, though, the physical diversity of robots makes full standardization harder than a single x86 baseline.

Hardware is commoditizing faster than expected

The price compression has surprised the industry itself. Unitree launched its G1 in May 2024 at 99,000 yuan, about 16,000 dollars (The Robot Report); by 2026 it offered a fixed-base dual-arm platform from 4,290 dollars and bipedal R1 / R1 Air versions between 4,900 and 5,900 dollars. Unitree's own average selling price fell from 593,000 yuan in 2023 to 168,000 yuan over the first nine months of 2025 (IPO prospectus, March 2026), while staying profitable through vertical integration of its motors, reducers and servo drivers. The pressure runs down the chain: per China Robot Industry Alliance data (2024-2025), harmonic-reducer maker Leaderdrive already holds more than 35 percent of units installed in China, priced 40 to 60 percent below Japanese Harmonic Drive equivalents.

The parallel with solar power is direct: Chinese solar module prices have fallen roughly 90 percent since 2010, with a further halving in 2023 and another 25 percent in 2024 amid overcapacity (CSIS, EnergyTrend). What China did to solar margins, it appears to be redoing to robot actuator margins. What remains defensible on hardware is no longer price but proven reliability, repeatable precision, certified safety (CE marking, see our AI Act and standards pillar) and a local service chain.

One distinction matters here: there are two kinds of hardware, and they follow opposite dynamics. Emerging hardware (humanoids, quadrupeds) commoditizes on sticker price, not on reliability: a 16,000-dollar G1 runs two hours on a battery and is not rated for any production duty cycle. Proven industrial robots sell something else entirely: tens of thousands of hours of MTBF, 24/7 operation for eight to twelve years, repeatability within a tenth of a millimetre and spare parts guaranteed for twenty years. That reliability is the product of decades of engineering and factory feedback: it cannot be copied in eighteen months, which is why industrial arm prices erode slowly while humanoid prices collapse. Conflating the two dynamics is the classic mistake of analyses written far from the shop floor. Our humanoid production index tracks this shift maker by maker.

Where is the real differentiator? Not in the model

The VLA is this wave's enabling technology, not its durable differentiator. GR00T N1 ships with open weights, so do OpenVLA and SmolVLA, and training recipes are published continuously: when any competent team can fine-tune a decent model in a few weeks, the model protects no one. As with large language models, performance converges and value moves to what surrounds the model.

What remains defensible in 2026: first fleet data, the deployed robots to data to model to robots loop that only players with machines in the field possess (see the section on the data economy); then industrialized reliability, the gap between a demo and eight auditable hours a day that has always structured industrial robotics; then hardware cost, which explains the spread in our production index between Unitree and Tesla; and distribution, the integrator and service network that incumbent arm makers own and no humanoid maker has built yet.

Onboard compute is a leveler: NVIDIA sells the same Jetson to everyone. The operational takeaway: between two vendors each claiming a VLA, do not compare models, compare fleet data, verifiable pilots and the service network.

Market proof: in October 2025 SoftBank did not buy a model lab, it bought ABB's robotics division for 5.375 billion dollars, that is a distribution network, an installed base and forty years of service, now assembled with Skild AI and AutoStore inside its Robo HD holding (announcement, 2025-10-08).

Can a roboticist actually do software? The structural question

Let's be direct: the cultures are largely incompatible. A roboticist thinks in 5- to 7-year product cycles, validated by machine safety dossiers; an AI team thinks in weekly iterations and continuous deployment. The first sells a CAPEX asset depreciated over a decade; the second builds a recurring SaaS-style relationship, with entirely different sales cycles and commercial organization.

The software-platform precedents from major robot makers call for caution. KUKA Connect, ABB Ability and FANUC Field have existed since the mid-2010s, with real monitoring features. But none of the four majors discloses a material recurring software revenue line distinct from robot sales: ABB's Robotics division generated 2.3 billion dollars in 2024, about 7 percent of the group, overwhelmingly tied to hardware. That is not proof of failure; it is proof software has not yet become, for these players, a profit center comparable to a SaaS valuation multiple.

The talent war widens the gap: in 2026, a senior researcher at a top AI lab (OpenAI, Anthropic) earns 500,000 to more than 2 million dollars a year, while an industrial robotics engineer sits around 145,000 to 185,000 dollars. A multiple of 3 to 6 that few industrial robotics P&Ls can absorb without rethinking the organization.

The counter-example that matters: Figure AI ended its OpenAI agreement in February 2025 to internalize its model, now Helix. Brett Adcock summed it up: "to solve embodied AI at scale, you have to vertically integrate robot AI" (TechCrunch, February 4, 2025). Brain-body integration genuinely matters, but Figure funded it with a raise valuing the company at 39 billion dollars, an entry ticket out of reach for a mid-size roboticist.

The new value chain

Stacking up the layers now in play produces a margin hierarchy, still shifting but already legible:

Two lines deserve emphasis. First, the bottom two layers, operational data and service, are ones an established roboticist can occupy without waiting: no need to rival NVIDIA on chips or the labs on models, only field presence already in hand. Second, the "certified cell" layer is not residual: it is the only one where regulation creates a structural barrier, since CE liability cannot be exercised from Santa Clara or a cloud.

The robot data economy

One fundamental difference separates physical AI from LLMs: the web's text was free and effectively unlimited, robot data has to be manufactured. Every manipulation trajectory costs hardware, operators and time, which makes data itself a manufactured good with its own supply chain. The result is a four-story industry, young but already taking shape.

1. Data factories. Tesla recruits "Data Collection Operators" paid up to 48 dollars an hour to walk more than 7 hours a day in motion-capture suits and VR headsets, with height requirements matched to Optimus (5'7" to 5'11"); more than 50 operators hired according to press reports, with collection costs estimated at up to half a billion dollars. In Shanghai, AgiBot runs a farm of about one hundred teleoperated dual-arm humanoid robots: its AgiBot World dataset exceeds one million trajectories, nearly 3,000 hours covering 217 tasks, 87 skills and more than 3,000 objects, released open source together with its simulation digital twins.

2. Capturing the human gesture. Egocentric video and wearable devices like UMI (Universal Manipulation Interface, Stanford / Columbia, 2024), a handheld gripper that records in-the-wild demonstrations without any robot, produce policies transferable across platforms at a per-demonstration cost far below teleoperation.

3. Synthetic data. Simulation (Isaac Lab, Cosmos at NVIDIA) is becoming a credible substitute: on July 8, 2026 Mistral AI unveiled Robostral Navigate, an 8-billion-parameter navigation model trained 100 percent in simulation (about 400,000 trajectories across 6,000 scenes), running on a single RGB camera.

4. Curation and standards. Hugging Face's LeRobot format is taking hold: the hub grew from about 1,145 robotics datasets at the end of 2024 to more than 58,000 by May 2026 (IEEE Spectrum), and Open X-Embodiment aggregates more than one million trajectories from 22 robot types.

The conclusion for European industrials: in this economy, the process data of a real factory, cycle times, tolerances, failure modes, variability, is the scarcest deposit of all. No teleoperation farm and no simulator can reproduce it. It is a tradable asset vis-a-vis the model labs, provided you capture it and put it under contract now. Many of these building blocks come out of young companies tracked in our startups comparator.

Software gets copied: defending your value

An honest observation before any software strategy: in 2026, code is no longer a defensible asset in itself. Published architectures are reproduced within months. Model weights get distilled: in early 2025, OpenAI and Microsoft claimed that DeepSeek had trained part of its models on ChatGPT outputs; the accusation remains contested and has led to no legal action to date, but the episode demonstrated that a frontier model can be approximated at a fraction of the cost by distilling its outputs. And purely software patents are fragile in Europe: Article 52 of the European Patent Convention excludes "programs for computers" from patentability "as such", with only inventions of technical character, typically the control of an industrial process, escaping the exclusion.

The real defenses, in order of effectiveness:

1. The proprietary data stream. Not a stock: a pipeline. Every deployed robot improves the model; the copier owns yesterday's code, never tomorrow's data. This is structural defense number one, and it connects to the previous section: whoever operates the fleet owns the stream.
2. Iteration speed with real customers. A product cycle that learns weekly in the field mechanically outruns a follower restarting from published code.
3. Certification and liability. The copy carries neither the CE marking, nor the insurance, nor the legal signature of AI Act conformity. European regulation, often experienced as a cost, acts here as a defensive moat.
4. Channel and service. The trust of integrators and OEMs is built over years of uptime and delivered parts; it cannot be cloned.
5. Ecosystem and network effects. Platform, developer community, marketplace: this is the Wandelbots NOVA bet already cited, where value lies in the number of connected players more than in the code.
6. Legal, as backup only. Trade secrets (EU Directive 2016/943 of June 8, 2016 on the protection of undisclosed know-how and business information), dual licensing, patents targeted at the hardware-software coupling (where technical character is established), contracts governing customer data ownership.

The practical consequence for a roboticist: never sell the software as a standalone product, whose price invites copying, but as access to that loop: a subscription tied to service, to maintained certification and to continuous model improvement from fleet data. The summary fits in one sentence: you do not protect a robotics software product, you protect the customer-data-certification loop around it.

Should a roboticist invest now, and for what return?

Yes, but investing does not mean building your own generalist foundation model: the entry ticket is out of reach for nearly every established roboticist. Investing means choosing your box in the table above, not occupying all of them.

Four realistic strategies:

1. Become the best body for other people's brains. An open approach like Unitree's, or partnerships within the GR00T ecosystem: Boston Dynamics, Agility Robotics and Figure AI are already building on Jetson Thor rather than competing with NVIDIA on the chip.
2. Acquire or ally rather than build alone. Even the best-capitalized players buy the competence: Meta acquired Assured Robot Intelligence in May 2026 (team folded into Meta Superintelligence Labs); Mobileye announced in January 2026 the acquisition of Mentee Robotics for 900 million dollars; Amazon had already acquired Fauna Robotics back in March 2025.
3. Verticalize on your application niche. A roboticist who already knows the tolerances and typical failure modes of a sector holds a data asset neither NVIDIA nor a generalist lab possesses.
4. Play the agnostic orchestration layer. European example: Wandelbots (Dresden) unveiled NOVA in late 2024, the first agnostic operating system for robots, compatible at launch with ABB, FANUC, Universal Robots and KUKA, integrated with NVIDIA Omniverse; 126 million dollars raised in Series C.

The ROI of this choice is not a project ROI: it is survival insurance plus an option. Indicators to track: the share of recurring software revenue in turnover (moving from 0 to 10-20 percent changes an industrial firm's valuation, echoing automotive software moving from 10 percent in 2018 to a projected 30 percent by 2030 per McKinsey), margin retention against compression, and the actual data asset built. For the project-by-project math, see our factory pilot ROI pillar.

One counter-example deserves tact: Nokia dominated mobile phones with excellent hardware, roughly 39 to 40 percent global share in 2007-2008, but failed to follow the shift to iOS and Android. Its share fell below 5 percent by 2013, the year it sold its handset business to Microsoft for 5.44 billion euros, with almost all its peak market value gone. Excellent hardware was not enough once value shifted into software.

Our take

1. The robotics value chain is being reshuffled, not settled: NVIDIA holds the strongest platform position (chip, model, simulation), but no robotics OS yet owns the monopoly Windows had on the PC.
2. Hardware alone no longer protects anything: the compression already seen in Chinese humanoids, from 16,000 to 4,290 dollars in under two years, makes any hardware-margin-only strategy untenable.
3. Becoming a generalist foundation-model lab is not realistic for most roboticists: the capital and talent entry ticket, with packages 3 to 6 times the industry's, requires a raise at Figure AI's scale.
4. Locally defensible value rests on two assets: the regulatory liability of the certified cell (CE, AI Act) and operational data from customer contact, neither replicable remotely by a chip or a model.
5. Not deciding is a decision: every year of waiting costs a year of data-asset build-up and competitive lead, in a market where hardware compression is accelerating, not slowing.

D-Fairy helps robotics players work through this strategic positioning question in the face of physical AI.