Physical AI’s Toll Booth

Every machine in the physical world is becoming intelligent: the factory arm, the mobile robot, the vehicle, the humanoid. The value is not in the machine but in the layer that lets it perceive a human and act safely, and that enablement layer is entered through a narrow certified control point one company holds today, on a deadline.

Joachim Laqueur, General Partner, 432 Legacy. 432 Legacy pairs a quantitative mapping engine with a founder-first practice. We scope an opportunity from the macro down to the specific companies sitting at the highest points of friction, across stages from early to mature. The analysis below runs across eight Physical AI value chains totalling 1,189 nodes and cross-checked against the field, in direct conversations between a portfolio company and the chip makers, robot makers, and sensor incumbents building this stack.

The two-minute read

Physical AI is the moment every machine in the physical world becomes intelligent: industrial arms, autonomous mobile robots and AGVs, vehicles, humanoids, any system that has to work in a space it shares with people. It is a transition on the order of a trillion dollars by 2035, though no single number captures it. The value in it is not the machine. The machine is becoming a commodity, and the money and the growth are moving to the layers that make it intelligent: AI in manufacturing alone is forecast at 155 billion dollars by 2030, growing several times faster than the hardware beneath it. The deepest of those layers is perception, the intelligence that lets a machine read where a person is, what they are doing, and what they are about to do, and act safely on it. That is the enablement layer, the foundation every intelligent machine has to stand on, and the capital is not pointed at it. In 2025, general-purpose and humanoid robotics took 13.2 billion dollars of venture, up from one hundred million two years earlier, while the enablement layer underneath those machines took roughly 1.6 billion, flat across six years. The way into that layer is a single control point: certified human-aware safety perception, the one node every deployment has to pass through, a software market of only about 1.3 billion by 2030 precisely because it is the hardest thing to build and the last to commoditise. Whoever holds it does not own a niche. They own the layer every intelligent machine depends on, on a regulatory clock that sets the timing.

We mapped eight value chains that decompose how a machine perceives a human and acts safely on that perception. Three findings.

One. The capital map and the human-sensing map point in different directions. The dollars concentrate on the robot, the humanoid, and the sensing chip. The layers that decide whether any of it works around people, the model that reads a human, the data that trains it, and the certification that lets it deploy, draw almost no venture funding.

Two. Across all eight chains, deployment friction beats supply scarcity in every one. The number of nodes where the technology exists but deployment is gated by certification, standards, or validation exceeds the number where physical supply is the constraint, in every chain without exception. Severe supply scarcity that is also ready to deploy: zero nodes, in every chain. The hardware exists. The rails to deploy it under a safety regime do not.

Three. The single scarce thing, certified human-aware safety perception at the grade industrial robots legally require, sits behind a single vendor and a regulatory clock. Veo Robotics’ FreeMove, the only commercially shipping safety-rated 3D speed-and-separation stack, was bought by the logistics company Symbotic in August 2024. The robot makers now have to buy their safety layer from a competitor or co-develop it. The EU Machinery Regulation makes the new safety regime mandatory on 20 January 2027. The deadline is real. The certified supply to meet it is not.

If your mandate touches Physical AI and you can deploy across the stack rather than only into a seed round, the question this answers is simple. Are you funding the robot, or the toll booth every robot has to pass through. The two are not the same trade.

I. The map is not the map

Robotics is one of the fastest-repricing categories in venture, even if it is far from the largest. The frontier AI labs raise on a scale robotics does not approach; robotics drew 20.9 billion dollars in 2025, a fraction of what flows to a handful of model labs. Inside robotics, though, the repricing is violent, and it has a clear centre of gravity. Humanoids took 6.1 billion dollars on their own in 2025, roughly four times the year before. General-purpose robotics went from a rounding error in 2023 to 13.2 billion in 2025. Autonomous vehicles, the prior wave, ran the other way, from 9.7 billion in 2021 down to 1.1 billion. Capital is moving hard toward embodied systems, and toward the most visible object in the frame: the robot itself.

The map of where capital is moving and the map of where the constraints sit are different maps.

Over the past year we built a critical-path map across eight value chains that, taken together, describe one system: a machine perceiving a person and acting safely on that perception in a shared space. Multi-modal perception, on-chip RF sensing in two forms, presence and state analytics, material-handling perception, robotics safety software, safe factory automation, and the SDK layer that ships all of it. One thousand one hundred and eighty-nine nodes. The same filter we apply everywhere, on priority, duration, and severity, surfaces a small head of genuine constraints in each.

When we overlay where the capital is going against where the constraints sit, the two maps barely touch. The capital is on the robot and the model. The constraint is the layer between perception and certified safe action: a narrow, unglamorous, deployment-gated slice that almost no venture dollar is chasing. Value migrates to bottlenecks. In Physical AI the bottleneck has already moved, and the capital has not followed it.

This article is about the gap.

II. The deeper finding is not the robot

If Physical AI were only a hardware race, the investment thesis would be straightforward: back the best robot, the best humanoid, the best sensing chip. The map shows something else.

Across all eight chains, the dominant constraint is deployment, not supply.

In every chain, the largest single cluster of nodes is the same shape: the technology supply is abundant and the deployment is gated anyway, by certification, by a standard, by a validation step, by an institutional rule. In the material-handling chain that cluster is forty-five percent of all nodes. In safe factory automation it is the plurality. In robotics safety software it sits alongside an equally large commodity floor, so that roughly seven in ten nodes are technically solved and the interesting question is entirely about what unlocks the gate. The pattern is not subtle and it does not vary.

The sharper version of the finding is in the cells that stay empty. In every one of the eight chains, the count of nodes that are both severely supply-constrained and ready to deploy is zero. There is no clean supply-only play anywhere in Physical AI’s safety stack. Every node where the technology is scarce is also gated by something that capacity funding cannot move. Where supply is the problem, deployment is also the problem, stacked on top.

The proof sits at the hardest nodes. The single highest-priority node in the material-handling chain is not a chip and not an algorithm. It is the safety-rated laser scanner, certified to IEC 61496-3 and required under ISO 3691-4 for driverless industrial trucks, supplied by a short list of incumbents. The hardware exists, the demand is real, and the deployment is paced by who holds the certification. Capacity exists. The rails do not. That is the Physical AI problem in one node, and it repeats across the corpus.

III. Eight chains, one shape

The chains split into three groups, and each group says the same thing from a different angle.

The sensing silicon is commoditising. Three chains cover the chip and IP layer: on-chip RF sensing across mmWave, ultra-wideband, and Wi-Fi, in both its present-state and forward-looking forms, plus the future multi-modal perception stack. The silicon here is not where the scarcity lives. The 60 GHz radar SDK node sits across four named vendors, Infineon, TI, Calterah, and Acconeer, with a fifth, Asahi Kasei, entering in 2026. At ultra-wideband frequency the field is wider still. What is scarce is one layer up: the per-packet Wi-Fi CSI driver that exposes the raw sensing data, the licensable UWB transceiver IP, the inference libraries that turn radio returns into a decision about a human. Those nodes carry the highest priority scores in the silicon chains, and they are gated, not supply-short. The chip is becoming the commodity. The thing that reads the chip is not.

The analytics layer is already a commodity. The presence-and-state analytics chain is the most commoditised in the corpus: more than three-quarters of its nodes sit at excess capacity. People-counting, occupancy, dashboards, the analytics product itself, are largely solved and largely free. The one severe bottleneck that is also ready to deploy in that entire chain is the qualified HBM memory and CoWoS packaging that feeds AI accelerators, which is to say a lock held by NVIDIA and TSMC, upstream and unrelated to the analytics business. The lesson for an allocator is blunt. Do not fund the dashboard. The dashboard is done.

The safety and certification layer is where everything binds. Four chains, material handling, robotics safety software, safe factory automation, and the SDKs that ship safe robotics, carry the deployment gate that the whole stack runs into. Here the binding constraint is certified human-aware perception at the grade the law requires: speed-and-separation monitoring under ISO 10218 and ISO/TS 15066, perception certified to ISO 13849 PL d and IEC 61496-3 Type 3. No certified machine-learning perception library exists at that grade as a generally available product. The motion-planning and collision-checking SDKs that would feed it are gated. The standards corpus itself is gated, sold as licensed feeds rather than open reference. This is the toll booth. It is a small slice of the total node universe, it is the layer with the fewest commodity nodes and the most stacked constraints, and it is the layer the capital is not funding.

IV. The foundational layer: the model and the data

Underneath the certification gate sits a layer the capital map misses, and it is the one that decides whether a machine can read a person at all.

A robot working near people does not need a better camera. It needs a model that turns whatever the sensor sees, radio returns, depth, video, into a reliable read of where a person is, what they are doing, and what they are about to do, and it needs that read to hold across sensing modalities and across sites it has never seen. That is a foundation model for human-machine interaction, and it is a different object from the language and vision models the labs are building. Our chains place its development, the perception and intent modelling, the cross-modal alignment, the behavioural read of a person in a shared space, among the hardest nodes in the corpus.

The harder constraint is the data the model has to learn from. Teaching a system to keep someone safe around a moving machine takes privacy-compliant, in-domain recordings of exactly that: people near forklifts, people inside collaborative cells, ground-truth on body position captured in the real environment rather than a lab. In the chains those dataset nodes are not commodity and not merely gated. They are severe bottlenecks. Procurable, in-domain industrial human-safety datasets effectively do not exist, and in-situ human ground-truth capture around moving equipment is scarce and blocked. The data that makes a robot human-aware is the rarest input in the stack.

That asymmetry is worth pricing. The robot makers and the large industrial operators own the one thing the model needs, the factories and fleets where the human-machine data lives, and they do not own the model. The companies that can build the model across modalities do not own the data. Whoever closes that loop, the model trained on proprietary human-machine data, holds a position the robot-chasing capital is not bidding on. This is the floor of the toll booth, below the certification and below the silicon, and it is empty of competition for the dollar.

V. The convergence: the chip makers came looking for the software

A value-chain map is a hypothesis until the field confirms it. Ours did, from two directions at once, in the conversations one of our portfolio companies, Algorized, has been having across this stack. Algorized builds the perception layer the map says is scarce: a sensor-agnostic foundation model that runs on the edge and reads mmWave, ultra-wideband, Wi-Fi sensing, and vision through one stack. What is instructive is not the company. It is who is pulling on it, and why.

From the silicon side, the chip makers came looking for the software. Texas Instruments has the radar silicon and wants the application layer that turns radar returns into an industrial safety decision, the inference layer the silicon does not deliver on its own. Asahi Kasei announced a public partnership in May 2026 to put Algorized’s model on its radar and take it from automotive to industrial and elder-care customers, because the silicon vendor has the chip and the channel and not the inference stack. Silicon vendors are reaching for the same application in gesture and gait recognition, where the value sits in the layer above the chip. Four chip makers, the same gap: the silicon is the commodity, the thing that reads it is not, and each of them needed a partner to build the layer their silicon depends on.

From the safety side, the gate is the certification, and even the incumbents cannot get through it alone. KUKA, working toward fenceless bin-picking cells, runs straight into the fact that the only commercially shipping safety-rated 3D perception stack belongs to Veo, owned since 2024 by the logistics company Symbotic, which leaves the robot OEM sourcing its safety layer from a company that acquired it for its own warehouses. A Japanese automotive manufacturer cannot put machine-learning perception on a production robot arm without certified outputs that, as of today, no generally available library provides. A major forklift OEM’s automated-forklift line is exposed at exactly the safety-laser-scanner bottleneck the map flags as the chain’s hardest node. And the dominant safety-sensor incumbent, the firm that holds more constrained positions in the robotics-safety chain than anyone, said the quiet part in plain words: there is no way to get this done inside, and there is a clear need to get it done. When the incumbent says it cannot build the thing in-house, the bottleneck is real.

Two demand vectors, the silicon side and the safety side, converging on one scarce layer. That convergence is the toll booth, and it is why a single perception company finds itself pulled by a chip giant, a robot OEM, and a sensor incumbent in the same quarter.

VI. Where this leaves capital

The way you capture the opportunity follows the shape of the gate. Inside the toll booth sit a small number of positions where the certification gap and the silicon-inference gap compound, and the action that captures each one differs. In some, the move is direct equity into the company that holds the inference layer. In others, it is a channel position alongside the silicon vendor whose chip becomes useless without the software. In a few, the most interesting shape is consolidation. The certified-perception layer is held by a logistics company that acquired it for its own warehouses, which leaves the entire robot-OEM industry buying its safety stack from a competitor. A regulatory deadline in January 2027 forces every one of those OEMs to resolve that dependency. That is the precondition for a wave of acquisition, and the asset that gets acquired is the certified layer, not the robot.

We are operating against several of these positions now. Algorized is the clearest: the certified-perception layer the map identifies as the single scarce position, pulled from the silicon side and the safety side at once, the company both halves of the convergence are reaching for. We are not naming the rest.

The Physical AI thesis that funds the robot is the consensus trade, and it is priced as one. The thesis that funds the toll booth is the one the map supports. The capital is going to the robot. The value is migrating to the enablement layer between perception and safe action, the model that reads the person, the data that trains it, and the certification that ships it. The control point into that layer is narrow, gated by a deadline, and for now nearly empty of competition for the dollar, and it commands a market many times its own size. Stop funding the commodity. Fund the structure that compounds.

Sources

The capital and market figures are sourced and tiered; the chain-level structural findings are drawn from 432 Legacy’s proprietary ARIA value-chain corpus and from confidential field conversations.

•    Robotics and humanoid venture funding 2019-2025: F-Prime Capital / PitchBook, State of Robotics 2025 (Americas, Europe and Israel). Humanoid-specific funding: MIT Technology Review / Crunchbase, 2025.

•    Operational robot stock (4,663,698 units, 2024), annual installations (542,076), sector mix (electronics 24%, automotive 23%, down from a 34% automotive peak in 2016): IFR World Robotics 2025.

•    Unplanned downtime (~$1.4 trillion per year across the Fortune Global 500, ~11% of revenue up from 8% in 2019; $2.3M/hr automotive): Siemens, True Cost of Downtime 2024. Cross-sector ~$125,000/hr: ABB Value of Reliability 2023 (n=3,215 plant leaders).

•    US manufacturing labour: 3.8 million roles needed by 2033, ~1.9 million at risk of going unfilled: Deloitte / The Manufacturing Institute, December 2025.

•    Physical AI theme size (~$900 billion by 2035, range $0.5-1.4T): Barclays, “Decade of the Robot,” February 2026. Adjacent market forecasts (AI in manufacturing $155B by 2030 at 35.3%; embodied AI $23B by 2030 at 39%; functional safety $9.2B by 2030; robotics software $19B by 2030): MarketsandMarkets, analyst consensus. Robotics-safety software slice (~$1.3B by 2030): 432 Legacy synthesis from the robotics-safety-systems market. Human-sensing opportunity aggregate (~$25B by 2030): 432 Legacy / ARIA synthesis summing the sensing-and-perception opportunity across automotive cabin, industrial human-robot collaboration, buildings, healthcare ambient, and public/commercial surfaces; a stitched estimate, not a single analyst figure.

•    Enterprise WLAN concentration (Cisco ~37%, HPE/Aruba ~20%): IDC Worldwide WLAN Tracker, June 2025. mmWave radar leaders (TI, Infineon): Coherent Market Insights / Yole.

•    Standards: ISO/PAS 8800 published December 2024 (ISO); IEEE 802.11bf finalised 2025 (IEEE SA); EU Machinery Regulation 2023/1230 mandatory 20 January 2027 (EUR-Lex); EU driver-monitoring mandate July 2026 (EU GSR 2019/2144); ISO/TS 15066 in force since 2016 (ISO).

•    BMW Spartanburg / Figure 02 pilot (10 months, 30,000+ X3s, retired and succeeded by Figure 03, no public ROI): BMW Group press release; Figure AI. Symbotic acquisition of Veo Robotics, August 2024: public record.

•    The eight Physical AI value chains (1,189 nodes), the bottleneck-and-deployment structure, and the counterparty field evidence: 432 Legacy ARIA corpus and confidential transcripts, vintage May-June 2026.