Physical AI: where the value actually accrues

01The Orien view

One page, stated plainly

Physical AI is the application of learned intelligence to machines that sense and act in the real world. It has crossed from a deep-tech bet into early commercialisation, what the market now calls its "GPT-2.5 moment": the capabilities are real and the scaling laws are visible, but production-grade reliability is not yet here. Capital deployed before that inflection compounds differently from capital deployed after it. That is the reason to have a view now.

Our view rests on one mechanism and one variable, and everything else in this paper hangs off them.

• The mechanism is integration versus recombination. While the interface between intelligence and hardware is immature, vertical integration wins, you must co-design across the boundary, and deployment data compounds for whoever owns the whole robot. As that interface standardises, the stack unbundles and value migrates to whoever holds the scarce asset.
• The swing variable is cross-embodiment transfer quality. If intelligence trained on one body keeps transferring to others, and in 2025–26 it demonstrably does, on a measurable scaling curve, then the stack unbundles, data moats weaken, form factor stops mattering, and the intelligence layer captures the margin. If transfer stalls, integration holds and the full-stack leaders compound. We do not have to guess which; we track it.
• A security and supply-chain overlay bends the whole picture. Because a robot is a sensing-and-data platform, data-exfiltration and IP concerns attach to any deployment in the West, which slows unbundling in exactly the highest-value markets and turns the allied component supply chain into a first-class investment theme, not a footnote.

From those three, our working stance across the sector:

• Selective  Foundation models / "robot brains." The prize is real but the rounds are enormous and the field is consolidating around a handful of names. We do not try to lead here; we look for teams turning general models into durable, specific advantages.
• Conviction  Robotics infrastructure & the component supply chain. Picks-and-shovels with physical, capital-intensive moats, actuation, allied magnets and reducers, data and evaluation tooling, sit closest to Orien's edge and are structurally under-supplied in the West.
• Conviction  Applied autonomy in structured environments. Underwritable on real unit economics today; narrow-task production reliability already clears 99%.
• Contingent  Full-stack humanoids. First-movers have a real, but time-limited, structural advantage. We own the leaders where we can, and we watch the swing variable to know when the advantage decays.
• Performance-first  Defense. The one vertical where the buyer pays for capability over unit economics, and, per Bessemer, the likely source of the category's first $50B+ IPOs.

We treat humanoids, then, as the sector's highest-variance option, the frontier bet whose payoff is enormous if the general-purpose form works, not the base case for where physical AI earns money over the next few years. This is also why our framework insists form factor is second-order: the durable question is which layer captures value as intelligence transfers across bodies, and it is answered the same way whether the winning body turns out to be an arm, a wheeled base, a drone, or a biped.

02Why now

Three curves turned at once: the intelligence became scalable, the hardware became cheap, and the labour it addresses became scarce. Capital followed, faster than most balance sheets have adjusted to.

The intelligence became scalable

For a decade, robot learning was stuck on an island: every robot, task and lab needed its own model, and there was no internet-scale corpus of physical experience to train on. That has changed in a way that rhymes with the language-model story. Between late 2025 and mid-2026, three independent groups published scaling laws for robot policies, showing that performance improves predictably and log-linearly with data, and, crucially, that intelligence trained on one body transfers to others. NVIDIA's EgoScale fit that curve at R²=0.998 across 20,000 hours of human video, with real-robot task completion rising from 30% to 71% as data scaled. Physical Intelligence and Georgia Tech showed the transfer emerges sharply above a diversity threshold, the way abilities emerge in large language models.

The investable reading is subtle but decisive: the question has moved from "does robot learning transfer?" (now settled, yes) to "how much data and diversity buys a target reliability?"

The hardware became cheap, and concentrated

A capable robot is now mostly a commodity-electromechanical problem. Chinese vertical integration has driven a genuine learning curve: Unitree cut the average selling price of its humanoid from ¥593k in 2023 to ¥168k by 2025, and its G1 now retails from roughly $13,500. Bank of America puts a pilot-stage Western humanoid at $90–100k today, with a Chinese bill of materials already near $35k and falling below $17k by 2030. Cost is no longer the moat. But the cheapness is Chinese cheapness, and that distinction, developed in §05, is where much of our thesis lives.

The labour became scarce

The demand pull is structural and demographic, not a technology cycle. Warehousing turnover runs above 100% annually at large operators; manufacturing faces retirements it cannot backfill; the roles most exposed are the unsafe, dull and unfilled ones. This is why the sober institutions now carry very large numbers, though we treat them as directional, and note how much they disagree because they measure different things in different years.

We deliberately show these as a table, not a chart: putting a $38B-by-2035 estimate on the same axis as a $7T-by-2050 estimate would mislead more than it informs. The signal is not the point estimate, it is that every serious house has revised upward, and that the only genuinely current numbers (Morgan Stanley, IFR) are the largest and the hardest.

Capital followed

Funding roughly doubled in 2025 and, on the same definition, H1 2026 alone has already beaten every prior full year. Bessemer's structural point still holds underneath the surge: over five years only 42 robotics companies raised $30M+ rounds against 745 software companies, this is a category catching up from under-investment, not a bubble inflating from over-investment. The caveat is definitional, and we flag it wherever it matters: "robotics" (narrow) and "physical AI" (broad) differ by roughly 2×.

03The stack & the map

Physical AI is not the SaaS stack. Value is contested at distinct layers, each with its own bottleneck, and the same technology reads differently depending on which end-market it points at. We separate the two axes deliberately.

The horizontal layers (where the technology is built)

• Bodies, actuators, reducers, sensors, structure, batteries. Commoditising fast; China-concentrated.
• Brains, vision-language-action models and world models. Capital-intensive arms race.
• Simulation & data, world models as data engines, synthetic environments, teleoperation. The scalable substrate.
• Infrastructure & MLOps, fleet orchestration, evaluation/QA, safety validation. Defensible, software-like.

The end-markets (where the technology is sold)

Each is a different application of the same stack, with a different buyer, form factor, unit economics, security profile and regulatory regime, and therefore a different verdict on where value accrues. This is why the thesis fans out into eight deep-dives rather than issuing one blanket call. The two horizontal layers (foundation models, robotics infrastructure) get their own treatment; the six end-markets, defense, industrial robotics, autonomous vehicles, consumer robotics, climate & agriculture, and health / life sciences, each get theirs.

04The central mechanism

Integration versus recombination

The oldest pattern in technology economics decides this sector. When the interface between two layers is immature and messy, the integrated firm wins, because performance requires co-designing across the boundary. When the interface standardises, the stack modularises and value flows to whoever holds the scarce component. Physical AI is living through exactly that transition, in real time.

Today favours integration. The interface between a learned policy and a physical body is still bespoke. To make a robot work you tune the model to the hardware and the hardware to the model, and every hour of deployment generates data that improves your stack on your body. That is the structural advantage a first-mover like Figure genuinely holds right now: not a patent, but a compounding loop that a modular competitor cannot yet assemble.

But every force in the sector pushes the interface toward standardisation. Hardware is commoditising below $20k. Cross-embodiment transfer is working. Open-source pipelines are maturing. The moment intelligence and body decouple cleanly, the integrated player's great strength, owning everything, becomes a cost disadvantage against best-of-breed recombination, and the deployment-data moat leaks, because data collected on any body starts to benefit models on every body.

The sharpest near-term version of World B is not exotic. It is best-in-class model + commoditised body + systems integrator, a Physical Intelligence-class brain, which already runs across many platforms, dropped onto a body that costs a fraction of a bespoke Western one. The much-discussed "a frontier lab buys a model company and plugs it into someone else's hardware" scenario is one instance of this. Notably, it does not require Tesla to do anything out of character: Tesla will never outsource the brain, because the brain is its thesis, which is also why Tesla's intelligence problem and its manufacturing strength are not easily combined with anyone else's opposite profile.

05The security & supply-chain overlay

The recombination story has a hole in it, and closing the hole is where a lot of our thesis lives. "Commoditised body" quietly assumes the body is a neutral commodity. It is not. A robot is a sensing-and-data platform, cameras, microphones, LIDAR, edge compute, connectivity, and in the West, data-exfiltration and IP concerns attach to any such platform, across every vertical, not just defense.

Decompose the body

The resolution is to stop treating "the body" as one thing. A robot is muscles-and-skeleton (actuators, reducers, screws, structure), a nervous system (sensors, edge compute, connectivity), and a brain (the model). The security concern lives almost entirely in the nervous system and the brain, the parts that perceive and transmit. The cost, and China's dominance, lives in the muscles.

So the only version of a modular robot that survives a Western high-value deployment is: trusted model + Western-controlled sensing / compute / connectivity + allied-or-reviewed actuation. You may buy the muscles wherever is cheapest and allied; you cannot let the eyes and the radio be adversary-controlled. That does two things. It narrows the modular challenger's cost advantage, because the security-sensitive sensing layer is exactly the part you must rebuild rather than buy cheaply. And it makes the security overlay a friction on unbundling, one that bites hardest in the most valuable markets, extending the full-stack incumbent's runway in the West beyond what the raw technology curve implies.

The dependency is quantified, and uncomfortable

This is not a hypothetical. Actuation, the muscles, is over half the bill of materials and the hardest thing to source outside China, and the single cleanest statistic in the sector is McKinsey's: building Tesla's Optimus without Chinese suppliers costs roughly three times more, with the bill of materials rising from about $46k to about $131k.

Two supply chains, and the geography of unbundling

Put security and cost together and the market does not resolve into one global commoditised hardware layer. It splits into two parallel chains: Chinese bodies for China, the Global South, and cost-tolerant, low-sensitivity uses; allied bodies for the West and for anything security-sensitive. That means unbundling proceeds at different speeds in different places, fast where buyers are relaxed about provenance (open-field agriculture, domestic Chinese deployment), slow where they are not (Western defense, in-home consumer, IP-rich factories). Which is why the allied supply chain, actuators, reducers, ball and roller screws, tactile sensors, and the rare-earth magnets beneath them, is a first-class theme rather than a sidebar. The magnet layer is already capitalising fast (MP Materials, Vulcan Elements, rare-earth-free Niron); reducers, screws and frameless motors remain far thinner ex-China, and that gap is the opportunity.

06Humanoids: the case study

Humanoids are where the framework is easiest to see, and where the popular debate is most confused. The popular debate is about form factor, bipeds versus wheels versus specialised arms. We think the swing variable demotes that debate to a second-order question.

Why form factor matters less than everyone thinks. If intelligence transfers across bodies, and it does, then the value does not accrue to whoever guessed the right body. It accrues to whoever accumulates the most useful deployment data and the best model, on any body. Cross-embodiment results mean data collected on an arm improves a humanoid; a form-factor bet is no longer a bet-the-company decision, because the intelligence survives the hardware generation. That is why we read the shipment data below as a story about manufacturing and supply chain, not about who picked the winning shape.

Why humanoids exist at all, the honest case. A purpose-built machine beats a humanoid on any single task's ROI. The humanoid's argument is not per-task efficiency; it is flexibility-weighted return. A humanoid at 80–90% of a specialist's productivity across twenty tasks can beat twenty specialists, because capital amortises across applications, changeover cost falls toward zero, and one platform is maintained instead of fifteen. For the right operator, mixed tasks, brownfield facility, acute labour shortage, sacrificing 10–20% task efficiency to buy that flexibility and those fleet-level economies of scale is rational. So the answer is not "humanoids or specialists." It is coexistence: specialists win high-volume narrow tasks; humanoids win the larger mixed-task, brownfield, labour-substitution universe. The form is second-order; the intelligence and the economics are first-order.

07Subsector verdicts

The framework applied across the eight subsectors. Each row is the headline; each has its own deep-dive that reruns the full analysis. The verdicts differ on purpose, the buyer, the security profile and the unit economics change the answer.

Note the shape of the table: our two conviction calls are the layers that win in both worlds of §04, infrastructure and the supply chain, plus applied autonomy, which pays for itself today. Defense is the deliberate exception to every unit-economics rule in the document, because its buyer is.

08What breaks the thesis

The four risks we take most seriously. Each is a place the compass could be pointing wrong, and each maps to something we can watch.

The reliability gap may not close on schedule

Narrow, structured tasks already reach production-grade reliability, Amazon's Vulcan picking is above 99%, DYNA folded 200,000+ towels at 99.4% over a 24-hour continuous run. But general, long-horizon, open-world tasks are nowhere close, and the last stretch from 80% to 99.9% is non-linear, Bessemer's view is that it needs fundamentally different approaches, not just more data. The most credible 2026 counter-evidence is Physical Intelligence's π*0.6, whose on-robot reinforcement learning cut failure rates by more than half and ran a coffee station for an 18-hour day. The gap is closing in narrow domains via a real method; it is not closing broadly.

Edge inference economics are unproven

A language model serves many users from one data-centre pipeline; a robot must generate world-state every few milliseconds, on its own body, on battery. Whether per-robot inference cost follows the ~1,000×-in-three-years curve that LLM serving did is genuinely unresolved, and it swings long-run gross margins, especially in consumer. The favourable signs are real (NVIDIA's Jetson Thor puts data-centre-class compute at 40–130W for $3,499; small on-device models like SmolVLA run on a laptop; dual-system architectures split a slow cloud brain from a fast on-body controller). But there is no demonstrated cost-decline curve yet. This is the clearest open economic risk in the stack, and a differentiation axis for whoever solves it.

The soft-body wall, and liability

Robots are good at rigid bodies and bad at deformable ones, cloth, cables, food, tissue. Precision insertion collapses to ~50% success; the physics of soft-material contact is computationally brutal. Whoever cracks real-time soft-body manipulation unlocks healthcare, agriculture and domestic robotics. Separately, because learned policies are probabilistic, no one can yet mathematically certify what a robot will do in an edge case, and, as our regulatory work found, no standard yet exists to certify a non-deterministic learned policy. The firm that turns physical-AI safety into an auditable, insurable artifact becomes a gatekeeper; until then, liability is an unpriced overhang.

Capital intensity and the China cut-off

Hardware-heavy companies return to the capital markets repeatedly and dilute early investors; a heavy-hardware bet for an independent fund is often a trap unless the path to a low bill of materials is clear. And the supply-chain dependency runs both ways: the same Chinese components that make bodies cheap can be export-controlled, as the 2025 rare-earth and magnet restrictions showed when they visibly hampered Optimus. Both risks argue for the same posture, favour the layers where intelligence scales faster than capex, and treat allied supply as a feature, not a cost.

09The signpost dashboard

This is the living part. Each row is a dated, falsifiable claim whose movement tells us which world we are in. We update the "current reading" every review; when a reading flips, the thesis above changes with it.

10Method & sources

Data current as of 23 July 2026. Figures are sourced to public material; where a widely-cited number is stale (Goldman 2024, ARK 2024, Citi 2024), we date-stamp it rather than present it as current. Funding totals are definition-sensitive, we state "narrow" (Crunchbase robotics) or "broad" (Dealroom physical AI) at the point of use. This paper draws on Orien's earlier internal research but contains no confidential information on any private company; all company facts are re-sourced to public disclosures, filings, and reporting.