Infrastructure: the layer that wins whoever wins

01Why this is the conviction call

The master thesis ended on a tension: the science of cross-embodiment transfer points toward a recombination future (World B), while the economics and security of the West still sit in an integration present (World A). Infrastructure is our highest-conviction subsector precisely because it does not require us to resolve that tension. Whoever wins the robot race, a full-stack integrator, a best-of-breed recombination, a Chinese volume champion or an American premium platform, every one of them needs data to train on, simulators to test in, software to run fleets, and physical components to move. Infrastructure is the toll booth on a road whose destination we do not yet need to know.

But "infrastructure" is not one thing, and the discipline in this chapter is in the sorting. It splits into a software layer, data, simulation, evaluation, fleet operations, and a physical layer, actuators, reducers, screws, sensors, magnets. Both are picks-and-shovels; they behave completely differently as investments. The software layer is being partly commoditised from above by a giant giving its tools away; the physical layer is gated by manufacturing know-how and, increasingly, by geopolitics. We are far more constructive on the physical layer, and this chapter explains why.

02The software layer, where NVIDIA sets the ceiling

Capital is pouring into the data-and-simulation layer, and the valuations are large. But most of these numbers describe companies serving autonomous vehicles and general AI as much as robotics, and, more importantly, they sit under a ceiling set by NVIDIA, which is deliberately giving its robotics software away to sell the silicon underneath.

The NVIDIA problem. NVIDIA now offers Cosmos (world models), Isaac (simulation and policy training), GR00T (an open humanoid model) and, reportedly since mid-2026, Omniverse free even for production. It does this because robotics software is a moat around its silicon, not a profit centre in itself, margin accrues in the chip, where corporate gross margins sit north of 70%. For an independent software vendor, this sets a hard ceiling: you cannot easily charge for a layer the ecosystem's gravity centre is giving away. The durable independent software businesses are therefore the ones that own something NVIDIA does not: a standard or data gravity. Foxglove owns MCAP, the de-facto logging format that robotics teams write petabytes into, switching cost by accretion, and layers a paid data platform on top. Applied Intuition is embedded inside the safety-critical development pipelines of most global automakers and classified defence programs, a relationship and regulatory moat that free tooling does not dislodge. These are the exceptions that prove the rule.

The three software sub-layers, sorted

Two adjacent categories deserve a specific warning. Teleoperation data factories, the "warehouse of hundreds of robots and remote operators" model (XDOF, Tutor, Micro1), are real and well-funded, but their core input is $15/hour human demonstration labour, rig costs have collapsed 10–400× (an open-source UMI gripper is ~$500 against a $200k mobile manipulator), and sim plus internet video threaten to substitute the data entirely. Durable margin there accrues only to proprietary deployment loops or genuine platform lock-in, not to renting teleoperators. And merchant edge-AI silicon is a graveyard: Hailo's valuation was cut from $1.2B to under $500M with a forced SPAC and ~50% layoffs, the clearest evidence that standalone inference chips commoditise against NVIDIA's software-moated high-compute tier. Notably, compute is only ~10–15% of a robot's bill of materials, so the "brain silicon" is neither the cost problem nor, for most, the margin opportunity.

03The physical layer, where the durable moats are

Now the part we are most constructive on, and the part most investors under-weight because it is unglamorous. A robot's value is overwhelmingly in its body's actuation, and actuation is gated by manufacturing know-how that took Japanese firms decades to build and that China is now racing to localise. This is capital-intensive, slow-to-replicate, and, critically, a place where an allied supply gap has become a matter of national policy.

The two hardest components have no Western greenfield at all. Precision reducers (harmonic strain-wave and cycloidal) are ~85% controlled by Japan's Harmonic Drive Systems and ~60% by Nabtesco, with China's Leaderdrive and Laifual climbing fast. Planetary roller screws, the linear actuators in every leg and knee, at $1,350–2,700 apiece, roughly fourteen per humanoid, are held above 50% by two private Swiss firms, with China ~80% import-dependent today. Here is the striking fact for an allocator: every single greenfield roller-screw plant announced in 2024–26 is Chinese (Beite, Wuzhou, Shuanglin). There is no Western startup and no Western greenfield. The same is true of independent precision-reducer capacity. That is not a crowded market to avoid, it is an unbuilt one, and it is exactly the allied-supply gap the master thesis argued would be forced open by the security overlay.

Why magnets are the template. Rare-earth permanent magnets are where the West woke up first, because China's 2025 export controls visibly hampered programs including Tesla's Optimus. The response is instructive: the US Department of Defense took a ~15% stake in MP Materials with a guaranteed $110/kg price floor and a ten-year offtake; Vulcan Elements assembled a $1.4B government-and-private package; Niron is scaling a rare-earth-free iron-nitride magnet. This is the model for how an allied component champion gets underwritten, policy converts a capital-intensive, China-dominated input into a de-risked domestic asset. The reducer and roller-screw gaps have not yet had their MP Materials moment. When they do, that is the entry.

04The framework read

Defense inverted the master framework; infrastructure confirms its core claim in the cleanest possible way. Infrastructure is the layer that gets paid in both worlds, but the split between software and physical maps precisely onto the master's distinction between moats that commoditise and moats that compound.

The through-line: value in this layer accrues to manufacturing scale and data gravity, and leaks away from cheap labour and undifferentiated silicon. That is why our conviction is real but selective, the subsector is a conviction call at the level of the right components, not a blanket bet on everything with "infrastructure" in the pitch.

05Orien's verdict

Infrastructure signposts we track

• The reducer/screw "MP moment." Watch for the first credibly-funded allied precision-reducer or roller-screw builder, or a government offtake package extending the magnet template to actuation. That is the buy signal for the deepest moat in the chain.
• Does NVIDIA's free-software strategy hold? If Omniverse-free-for-production and Halos entrench NVIDIA as the convening layer, independent software value stays capped and we stay narrow. If antitrust or OEM vertical integration (Tesla silicon) cracks the ceiling, the independent software thesis re-opens.
• Data gravity vs. data commoditisation. Open datasets (Open X-Embodiment, LeRobot's 58,000+ datasets) are commoditising the data layer, bearish for generic data sellers, bullish for whoever owns the pipeline and the platform. Track whether Foxglove-style data platforms convert logging standards into pricing power.

06What breaks this call

NVIDIA eats more of the stack. The single largest risk to the software half is that NVIDIA's free-tools-to-sell-silicon strategy extends further, into evaluation, fleet ops, data, leaving independents with ever-thinner slices. Our mitigation is to own standards and data gravity, not features; but if NVIDIA absorbs those too, the software thesis narrows to almost nothing.

The physical moat cuts both ways. The reducer and roller-screw gap is real, but it is unbuilt in the West for a reason, the manufacturing barrier that makes it defensible also makes it slow, capital-hungry and hard to execute. Betting on an allied builder is a bet on industrial execution, not software iteration, and the timeline is measured in years. The China incumbents are cheaper and faster; if the security overlay ever softens, the allied premium erodes.

Cost-down can outrun the moat. If humanoid bills of materials fall as fast as the bull case says (China BOM below $17k by 2030), component margins compress across the board, and even the "good" components trend toward commodity economics. The defence is to own the components where know-how, not scale alone, sets the price, reducers and roller screws over motors and structure.

Data commoditisation. If sim and internet-video pre-training (the Skild bet) genuinely substitute for collected robot data, the entire teleop-and-data-factory layer deflates, which is why we only watch it, and would only own the proprietary-deployment-loop version of it.

07Sources