Physical AI & Quantum - Issue 2

 00 · EXECUTIVE SUMMARY

Five Things This Briefing Will Tell You

Issue 1 identified that the binding constraint on AI was not electricity generation but grid delivery. Issue 2 goes one layer deeper. AI is leaving the screen and entering the world, and the constraints that shape its physical deployment are categorically different from anything the digital build-out faced. The robots are real. The data they need to become generally useful is not. And the compute architecture that could eventually unlock the hardest problems AI cannot yet reach is not an evolution of what we already have — it is something else entirely.

1 — Why data — not hardware — is the binding constraint on physical AI.

2 — What three complications the consensus robotics thesis misses.

3 — Who is quietly winning the physical AI data race — and why it is not who most investors are backing.

4 — Why quantum computing is not the next stage of AI — and what it actually is.

5 — What near-term quantum applications already generate commercial value, and what signals tell you whether the broader timelines are moving toward you or away.

PART ONE: PHYSICAL AI & ROBOTICS

01 · CONSTRAINT DIAGNOSIS

The Binding Constraint Is Data, Not Hardware

Robots are not new. The automotive industry understood decades ago that structured, repeatable physical tasks — welding, painting, precision assembly — could be automated with sufficient investment and engineering. What changed is not the robot. It is the ambition.

Consider what happened at BMW's Spartanburg plant in South Carolina in 2025. Two Figure 02 humanoid robots ran ten-hour shifts for eleven months — loading sheet metal parts onto welding fixtures with a five-millimetre tolerance, in under 37 seconds per cycle. Over 1,250 hours, they loaded more than 90,000 parts, contributing to the production of over 30,000 BMW X3 vehicles. Accuracy stayed above 99%. By any previous standard, this was a landmark.

And yet BMW confirmed at the end of the pilot that there are no Figure AI robots deployed at Spartanburg and no timetable for reintroduction. What the eleven months produced was not a commercial contract. It was data — 1,250 hours of physical interaction data in a controlled industrial environment, which went directly into the design of Figure 03. The deployment was a data collection exercise dressed as a commercial proof point. That distinction is the constraint thesis in miniature.

Compare that to Sorted — a UK company whose AI-powered robots sort mixed recyclables on conveyor belts in material recovery facilities. Their technology works at commercial scale because the environment is fixed, the categories are defined, and the robot never needs to generalize. Sorted is a concrete commercial application. It is also, in the context of where physical AI is trying to go, a relatively simple problem.

The ambition now is for physical AI systems that can navigate unstructured environments, adapt to novelty, handle objects they have never seen, and operate autonomously alongside humans in spaces designed for humans. The gap between what Figure achieved at BMW and what that ambition requires is not a hardware gap. It is a data gap — and it runs deeper than the raw hours of training data collected so far.

 What We Know

DROID — a large-scale robot manipulation dataset produced through a collaboration across 52 buildings and 13 institutions — contains approximately 76,000 trajectories and roughly 350 hours of real-world robot manipulation data. Open X-Embodiment, the most ambitious cross-institutional effort to date, pools over one million trajectories from 22 robot embodiments across 34 labs. Combined with egocentric video datasets like Ego4D, the global open-source physical interaction corpus — the shared research commons accessible to any institution building foundation models — sits at roughly 5,000 to 6,000 hours of diverse real-world data. These are open-source datasets: shared across institutions precisely because no single player has accumulated enough diverse physical interaction data to train general-purpose models alone.

For academic purposes these numbers are not small. For training the kind of generalized physical foundation model that the sector’s ambitions require, they are orders of magnitude too small. Language models train on trillions of tokens. Physical AI researchers have been direct about the gap: language models won because they could draw on data at internet scale. Robotics has no equivalent open-source corpus.

Larger proprietary datasets do exist — but they do not resolve the constraint. Scale AI’s Physical AI Data Engine has collected over 100,000 hours of real-world robotics data: a significant commercial effort that nonetheless remains roughly two orders of magnitude below what researchers believe foundation models will need for reliable generalisation outside controlled environments. Amazon’s warehouse robots and Waymo’s autonomous fleet generate data at volumes that dwarf the open-source corpus — but both are deeply domain-specific. Amazon’s data describes fulfilment centre logistics. Waymo’s describes the surface of public roads. Neither teaches a robot to reposition a frail patient, navigate a kitchen that was not designed for machines, or respond to the unpredictable behaviour of a person in distress. 

The same pattern holds in elder care: Cera, the UK home healthcare company, has built a dataset of over 300 billion health data points through its carer-facing app — the richest longitudinal elderly care dataset in Europe, per Cera’s own published figures. That data is clinical and behavioural: it describes what carers observe and record. It does not capture the physical manipulation, the dexterous caregiving micro-tasks, or the environmental navigation that a humanoid robot operating in a care home would need to learn. The most direct route to that data is teleoperation: human operators piloting robots through tasks in real environments, with every sensor reading, camera frame, and motor command logged as a training pair. Several companies are now deploying robots primarily as teleoperation rigs — paying people to pilot them through domestic and care-adjacent tasks, generating labelled physical interaction data that no simulation or internet video can replicate. The methodology is right. The scale at which it is currently operating does not yet close the gap.

The data constraint is therefore better understood not as simple scarcity — that the data does not exist — but as fragmentation and domain specificity. The data exists, in volume. It is locked inside corporate IP walls, collected in narrow operational contexts, and inaccessible to the players who need diverse, generalised physical interaction data to train foundation models for unstructured environments. That reframing matters for where the investment advantage actually sits.

02 · WHY CONSENSUS GETS IT WRONG

Three Complications the Standard Thesis Misses

The Data Moat Is Being Built — Inside Operating Businesses

The investment consensus focuses on the companies building the robots. This is the wrong layer. The question is not who builds the best robot. The question is who owns the data that trains the next generation of physical AI — and that data is accumulating fastest inside companies that are not primarily robotics companies at all.

Amazon crossed one million deployed robots in July 2025, confirmed in a company press release announcing its DeepFleet AI foundation model for fleet coordination. Those robots, operating across more than 300 fulfilment centers, are generating proprietary physical interaction data at a scale no academic consortium can match. Waymo — having crossed 100 million fully autonomous miles in July 2025 and surpassing 450,000 paid rides per week by December 2025 — has accumulated autonomous driving data at a volume no competitor can replicate without years of operation. Uber, per Financial Times reporting in April 2026, has committed over ten billion dollars to autonomous vehicle deployment through equity stakes and fleet purchase agreements — a bet not on a single technology, but on owning the platform relationship with the rider regardless of which autonomous system does the driving. The more interesting question is where the same pattern is playing out less visibly.

The sectors with the highest structural need for autonomous physical systems — elder care, healthcare, domestic settings — are also the sectors where proprietary care-interaction data is building quietly inside operating companies. Cera, the UK home healthcare company, has assembled a dataset of over 300 billion health data points through its carer-facing app, covering medication adherence, dynamic risk scoring, and real-time care observations across a large deployed base. It has since acquired GenieConnect, the companion robot platform it ran pilots with, and is integrating its dataset with robotics training directly. 

The logic is the same as in logistics and autonomous vehicles: an operating business accumulating physical interaction data as a byproduct of its core operations, at near-zero marginal cost, while venture-funded robotics startups pay to generate equivalent data one trajectory at a time. The constraint is not that no one is building this data. It is that the data being built is fragmented across operating businesses with different formats, different consents, different deployment contexts — and that none of it yet captures the full stack a care humanoid would require: the physical handling of frail bodies, the dexterous micro-tasks, the environmental navigation, the emotional and conversational regulation that makes the difference between a robot performing a task and a robot providing care.

 The Structured-to-Unstructured Gap Is Much Wider Than It Appears

The robots operating at commercial scale today share a defining characteristic: they work in environments built around their limitations. The automotive assembly line. The Amazon fulfilment center. The recycling facility. Fixed layouts, defined tasks, controlled conditions.

The environments where physical AI is most structurally necessary are categorically different. Elder care happens in homes that were not designed for robots, with residents who move unpredictably. Construction sites are among the most variable and hazardous physical environments humans operate in. Rodney Brooks, founder of iRobot, wrote in September 2025 that the industry is more than ten years from the first profitable humanoid deployment. Jensen Huang put the timeline at two to three years. The distance between those two estimates — from two of the most credible figures in the field — is the honest measure of the uncertainty the market is not pricing.

The Tacit Knowledge Problem

The data collection efforts above capture what people do. They do not capture what people know. The tasks physical AI is most needed for — elder care, surgical assistance, skilled construction — are not simply technically complex. They draw on a kind of knowledge that has never been written down, because for most of human history it has not needed to be.

 A skilled care worker reading a patient's discomfort before the patient can articulate it. A surgeon adjusting pressure based on the feel of tissue that no training manual describes. A plasterer knowing from the sound and resistance of the trowel when the mix is ready. This is tacit knowledge — embodied, contextual, accumulated over years of practice — and it is the dominant form of intelligence in the environments where robots are most structurally needed.

The data collection efforts underway — gig workers filming their chores, teleoperation rigs in fulfilment centers — capture explicit, observable action sequences. They do not capture the judgment layered beneath them. A robot can learn to fold a shirt by watching footage. It cannot yet learn to notice that the patient who asked for the shirt is cold because their colour has changed, and that this matters more than the shirt. The gap between task execution and contextual judgment is not merely a function of training data volume. It is a function of the kind of knowledge that is hardest to externalize — and therefore hardest to collect, label, and learn from.

This does not make the problem unsolvable. It makes the timeline longer than the hardware-focused investment consensus assumes, and the eventual moat deeper for whoever finds a way through it.

The Institutional Constraint Will Slow Deployment Where It Matters Most

Germany faces a projected shortfall of seven million skilled workers by 2035. Japan's working-age population has been shrinking for two decades. South Korea has the world's lowest birth rate. The demand for physical automation in care, construction, and healthcare is structural, not cyclical.

What the economic models do not capture is the institutional friction that will slow deployment in exactly these sectors. Liability frameworks for autonomous systems operating near vulnerable humans do not yet exist at scale. The insurance markets that would need to underwrite these deployments are still forming. The social licensing required to deploy autonomous systems in intimate, consequential settings is far from assured. The sectors deploying physical AI today are the sectors where institutional constraints are lowest. The sectors with the highest structural need are the sectors where those constraints are highest. The market is pricing the easy deployments. It is not pricing the friction on the necessary ones.

03 · INVESTMENT IMPLICATIONS

Optionality Over the Robot

The analytical frame from Issue 1 applies directly here. The AI energy build-out is attracting capital at every layer of the stack — chips, data centers, generation capacity — but the binding constraint sits beneath all of it: grid delivery, interconnection queues, transformer lead times. Physical AI is following the same pattern. Boston Dynamics, Figure, Unitree, and a dozen well-capitalized competitors are racing to build the most capable robot. That race is real and the leaders will matter. But the constraints that will determine how fast physical AI can actually scale sit beneath the hardware race — in the data required to train generalized physical intelligence, and in the supply chains every robot depends on regardless of which design wins.

 Rare earth magnet supply chains outside Chinese control are the most concrete near-term hardware constraint. China controls approximately 91% of global rare earth refining, and export restrictions introduced in April 2025 on seven rare earth elements remain in force; a second wave announced in October 2025 covering twelve elements and processing equipment was suspended until November 2026 following US-China trade negotiations. Every robot joint depends on these materials. Lynas Rare Earths achieved the first commercial production of dysprosium oxide outside China in May 2025. Iluka's Eneabba refinery in Australia targets commissioning in 2026. Carester and Solvay have committed over €400 million to Western processing capacity in France. Niron Magnetics in Minnesota is developing iron-nitride magnets requiring no rare earths at all. None of this closes the gap before 2028 at the earliest.

Operating platforms with existing physical data generation — Amazon, Walmart, Waymo — are accumulating training data at near-zero marginal cost. The investment thesis here is not primarily about robotics. It is about recognizing that the companies running physical operations at scale are quietly building the data infrastructure for the next generation of AI capability.

PART TWO: CLASSICAL COMPUTE TO QUANTUM

04 · THE NEXT FRONTIER

A Different Kind of Problem

Every conversation about AI eventually reaches the same wall. The models keep improving. The applications keep expanding. The infrastructure keeps being built. And yet the hardest problems — the ones that would matter most — remain out of reach. Not because the models are too small or the chips too slow, but because classical computing has physical limits it cannot engineer its way through.

The physical AI constraints in Part One illustrate this precisely. The rare earth magnets every robot joint depends on are scarce because the molecular properties that make them uniquely useful occur in specific geological formations concentrated in specific countries. To design a synthetic alternative — a magnet with equivalent properties assembled from more abundant materials — requires understanding quantum mechanical interactions between atoms at a level of precision that classical computers cannot accurately model. The simulation becomes exponentially more complex as the molecule grows, and at some point the approximations stop being close enough to be useful.

More GPU clusters will not solve this. The problem requires a different kind of computation entirely.

Quantum computing is receiving more serious institutional attention right now than at any previous point. In 2025, IonQ completed a $1.075 billion acquisition of Oxford Ionics, a UK trapped-ion quantum computing company cleared under the National Security and Investment Act. This is the largest quantum M&A transaction to date, and a signal that the sector has moved from academic curiosity to strategic asset. IBM has reported cumulative quantum revenue approaching $1 billion since 2017, with enterprise clients now paying for access to quantum systems in anticipation of near-term advantage rather than waiting for proof of it. 

Amadeus, the global travel technology company, partnered with quantum algorithms firm Kvantify to investigate quantum approaches to airline network revenue optimisation — a combinatorial problem so large that classical computers can only approximate a solution. Their joint research won Best Innovation at the AGIFORS Revenue Management conference and was published in the Journal of Revenue and Pricing Management. Amadeus has been transparent that the hardware is not yet at commercial scale — but the research investment is real, and the problem it is preparing for is real.

Serious enterprises do not invest in research that has no institutional return. They invest when they believe the window between now and commercial viability is short enough that the positioning cost today is worth the advantage it buys. The Amadeus commitment, like IBM’s cumulative quantum revenue and IonQ’s acquisition of Oxford Ionics, is evidence of that calculation being made.

Quantum computing does not replace classical compute — it addresses a categorically different class of problems that classical computing cannot approach efficiently regardless of scale, and some of those problems sit directly inside the constraints that physical AI cannot solve without it.

Quantum Is Not What Most People Think It Is

Most people treat quantum as a faster classical computer — a successor that does the same things better. NVIDIA is not going to be displaced by quantum hardware. Neither is Intel. A GPU running a large language model and a quantum processor running a molecular simulation are doing fundamentally different things.

Classical compute is extraordinarily good at optimizing within known solution spaces — pattern recognition, language generation, fraud detection. It is also improving rapidly at prediction within genuinely uncertain territory, where sufficient training data allows it to assign reliable probabilities. Anthropic's Claude, integrated into Salesforce, is operating above 80% accuracy on sales workflow suggestions at major financial institutions — a threshold at which productivity gains become real and user trust compounds. Classical AI's frontier is expanding into the known unknowns as well as the known knowns. Quantum addresses something categorically different — not a harder version of the same problems, but a class of problem that classical computers cannot approach efficiently regardless of scale.

GPU error rates sit at roughly one error in ten million operations. Quantum hardware currently operates at roughly one error in a thousand. That is not a version of the same thing. It is a different machine for different problems, and the division of labour between them is still being worked out. 

 Where the Field Actually Sits

Raw qubit counts are not the right metric. What matters is the combination of qubit count and error rate. A system with fewer qubits at lower error rates can outperform a system with more qubits at higher error rates — which is why comparing companies by qubit count alone is misleading.

Google's Willow chip, published in Nature in December 2024, achieved below-threshold error-correction scaling — meaning the logical error rate fell as the system grew larger. This is what quantum error correction theory predicted had to happen for fault-tolerant computing to be possible, and it had not been demonstrated cleanly before. It is a physics result, not yet a computing result.

Quantinuum's H-Series systems have demonstrated 99.9% two-qubit gate fidelity — among the highest publicly validated in trapped-ion hardware. IBM targets 200 logical qubits by 2029 and 2,000 by 2033. Quantinuum targets fault-tolerant operation by 2029. Both are roadmaps. Microsoft's Majorana 1 announcement in February 2025 — claiming the world's first topological qubit — was accompanied by a Nature editorial note stating the results did not constitute evidence for the quantum states the approach depends on. Treat it as contested until peer-reviewed demonstration appears.

Scott Aaronson, a quantum computer scientist at the University of Texas at Austin is widely regarded as the field's most rigorous independent commentator, has placed the probability of a key error-rate threshold being demonstrated before 2030 at above 90% for at least one research group. He is significantly more skeptical about the commercial framing layered on top of those results. The physics timeline and the commercial timeline are not the same timeline.

Where Quantum Actually Wins: Three Near-Term Applications

The long-horizon debate about fault-tolerant quantum computing is real, but it is not the only question worth asking. Where is quantum already generating commercial value, or close enough to doing so that the positioning decision is today rather than in five years?

Application One: Breaking the Encryption That Protects Everything

All public-key encryption — the system that secures online banking, international wire transfers, diplomatic communications, and most of the internet — rests on a mathematical assumption: that factoring very large numbers into their prime components is computationally impossible in any useful timeframe. A sufficiently powerful quantum computer running Shor's algorithm would not weaken that assumption. It would break it.

That threat is already operational, not forecast. Governments and sophisticated state actors have been systematically collecting and storing encrypted communications since at least the 1990s — material that is unreadable now but will become readable the moment a cryptographically capable quantum computer exists. The data is in storage. The clock is running. The US intelligence community does not treat this as theoretical.

NIST (the US National Institute of Standards and Technology) finalised three post-quantum cryptography standards in August 2024 — FIPS 203, 204, and 205. The NSA’s CNSA 2.0 mandate requires quantum-resistant encryption for software signing by 2030 and operating systems by 2033. The regulation arrived because the intelligence community considers the threat live, not as a precaution against a remote possibility. Every regulated financial institution on earth will need to comply. The procurement window is approximately three years. After that, it is infrastructure, not alpha.

Application Two: Quantum Random Number Generation — The Infrastructure Already Running

Every cryptographic system, every Monte Carlo model, every secure financial transaction runs on random numbers that are not actually random. Classical computers generate pseudo-random numbers — deterministic algorithms that simulate randomness. In high-security infrastructure, a sophisticated adversary who reverse-engineers the seed can predict the sequence. Quantum random number generators produce true randomness derived from quantum mechanical processes: physically guaranteed, not an approximation.

In March 2025, a team from JPMorganChase, Quantinuum, Argonne National Laboratory, Oak Ridge National Laboratory, and the University of Texas at Austin published results in Nature demonstrating the first mathematically certified truly random numbers generated by a quantum computer — using the 56-qubit Quantinuum H2 trapped-ion system accessed remotely over the internet. The randomness expansion protocol they demonstrated is unachievable by classical computation.

ID Quantique, a Swiss firm founded in 2001 as a spin-off of the University of Geneva’s physics department, has sold certified QRNG hardware for over two decades. Their devices are independently verified to AIS31 standard and deployed in banking infrastructure and national state lotteries. Samsung’s Galaxy Quantum series — now in its sixth generation, sold exclusively in South Korea through SK Telecom — integrates ID Quantique’s QRNG chip, demonstrating that the hardware has reached commercial consumer deployment. In 2025, IonQ acquired a controlling stake in ID Quantique. The hardware works. The regulation is in place. The deadline is 2030.

Application Three: Financial Risk Modelling — The Foundation Being Built Now

Monte Carlo simulation is the workhorse of derivatives pricing, credit risk, and Value at Risk calculations. It works by running millions of random simulations and averaging the results — an approximation bounded by compute time. Quantum Amplitude Estimation can, in principle, achieve the same accuracy with an exponentially smaller number of samples.

In May 2024, JPMorganChase, Argonne National Laboratory, and Quantinuum published a paper in Science Advances demonstrating clear evidence of a quantum algorithmic speedup for the Quantum Approximate Optimization Algorithm (QAOA)— with potential applications in financial modelling, logistics, and materials science. The hardware is not at commercial scale yet. What is real is the research foundation and the institutional positioning. Even a 1% improvement in portfolio optimisation is commercially significant at institutional financial services volumes. That is not disruption. It is incremental, compounding edge — and it is where the near-term value sits.

05 · SIGNAL WATCH

What to Track Over the Next 18 Months

These are not investment triggers. They are diagnostic signals — the observable indicators that tell you whether the constraints described in this issue are tightening or easing. 

 06 · WHAT TO DO WITH THE SIGNALS

You are already receiving intelligence about physical AI and quantum computing. Most of it is structured around a question that is genuinely unanswerable right now: which humanoid company wins? This briefing is not about that race. It is about the infrastructure that every runner depends on — and where patient capital has a structural advantage because the timelines are longer than a single reporting period.

You probably don't need new positions. You need a better lens on what you already hold. If you own positions in automation, industrial technology, or AI infrastructure, ask one question about each this week: does this asset sit closer to the data problem or the hardware problem? The hardware problem is getting solved, noisily and expensively, by well-capitalised companies that will be correctly valued by the time most allocators move. The data problem — and beneath it, the tacit knowledge problem — is quieter, more structural, and further from consensus pricing. The companies that will look prescient in ten years may be the ones running the warehouses, the autonomous fleets, and the care facilities — accumulating physical interaction data as a byproduct of operations that were already profitable before anyone used the word humanoid.

On quantum, the investment decision worth making today is not a bet on when fault-tolerant computing arrives. It is a recognition that a regulatory deadline — 2030 for software signing, 2033 for operating systems — is already driving real procurement spending. The infrastructure serving that compliance is commercially available, independently certified, and generating revenue. That is not a future thesis. It is a present one, in a market that has not yet been repriced for the tailwind.

The constraints this issue describes are real. They are also problems the world is in the process of solving — not because the solutions are easy, but because the human impulse toward ingenuity under constraint has not changed.

The rare earth concentration that gives a single country disproportionate leverage over the materials inside every robot joint is the kind of structural problem that historically gets solved — slowly, expensively, and then all at once — by economic incentive meeting scientific curiosity. Researchers working on iron-nitride magnets are not making headlines. They are doing what researchers have always done: finding a way around a wall that everyone assumed was permanent. The same is true of the encryption foundations being rebuilt right now — not because a quantum computer has broken them yet, but because the people responsible for global financial security have decided not to wait to find out.

The tacit knowledge locked inside the world's best surgeons, care workers, and craftspeople is the most valuable dataset no one has yet collected. Unlocking it — making it transferable to machines that could then deploy it at scale in the places that need it most — is one of the genuinely hard and genuinely important problems ahead. It will not be solved quickly. When it is solved, it will matter enormously.

Every structural shift this publication tracks is, at its core, a problem the world is in the process of solving — with ingenuity, with patient capital, and with the same restless human curiosity that has navigated every previous transformation. The investors who understood the structure of the constraint — not just the excitement of the theme — will be the ones still standing when the cycle turns.

 07 · WHAT'S COMING IN ISSUE 3

Issue 1 tracked the constraint inside the data center: grid delivery. Issue 2 has tracked the constraint as AI crosses into the physical world: data scarcity, tacit knowledge, supply chains, and a different computational architecture that addresses problems classical AI cannot reach.

Issue 3 follows the water. Every data center built, every chip manufactured, every cooling system running around the clock draws on a resource the market has not yet properly priced. Water is the constraint hiding in plain sight behind the AI build-out — connecting to food, agriculture, and industrial geography in ways that will reshape regional investment logic over the next decade.

Issue 4 takes a deliberate turn. In a world of infinite digital abundance, the scarce resource becomes something no algorithm can replicate: genuine human presence. Why live events, sports franchises, and physical venues are seeing structural demand that defies every efficiency argument — and what that means for patient capital.

 APPENDIX

Numbers That Should Stop a Room

Use at will to start the debate. Amber rows require source confirmation before distribution.

 THE CONSTRAINT SIGNAL  ·  Issue 2  ·  May 2026  · 

This publication is produced for sophisticated private investors and family offices. Nothing herein constitutes investment advice. www.constraintsignal.com