Embodied AI has split into brains and bodies. The brain companies are raising like the body companies, and they are being pulled into building fleets and hardware they never wanted to own. What they need is the thing autos already invented: a Tier-1 who designs the deterministic hardware under the software stack, stands behind it, and never touches the stack. Nora is building that Tier-1 for the motion domain of robots.
The stack has split. Half the well-funded US embodied companies build bodies (Figure, Apptronik, Agility, 1X); the other half build models and run on other people's hardware (Skild, Physical Intelligence, Generalist, Genesis, Dyna). The model companies are a permanent customer class now, not a research phase.
The model companies are being dragged into hardware. Physical Intelligence, which says it builds no hardware, is hiring NPI, manufacturing and build technicians in Fremont. Dyna has a semi-humanoid prototype. Genesis went full-stack because hand fidelity gated data transfer. Skild is sourcing actuators and moving to contract manufacturing in Asia. None of them has a firmware team; all of them are standing up thin reliability and integration groups because nobody else owns the problem. Their supplier options are whole robots from Unitree-class OEMs, which every serious team bypasses and re-characterizes by hand, or bare components with no firmware, no measurements and no safety chain.
They will not buy software; they will buy hardware with data attached. Fifteen years of "operating system for robots" vendors (Rocos, Renovo, MOV.AI, Freedom, Formant, Viam, Apex.AI) produced two acqui-hires, two pivots and zero customers among companies whose product is the robot. The vendors these companies verifiably pay are NVIDIA, Scale AI, Lightwheel, Jabil, and the Chinese hand makers. The pattern is consistent: the buyable object is a physical thing whose behaviour is documented, not a layer of code.
The seam is standardizing, which is what makes a product definable. Across Figure, Boston Dynamics, NVIDIA and the academic stacks, the interface between the learned policy and the hardware has converged on joint-position targets with gains at around 50 Hz into a kilohertz torque loop. What sits below that line is the same for every customer, and it is where the sim-to-real cost lives: a policy trained on a model of the robot fails when the real robot drifts out of that model.
A motion domain controller, the characterized joint modules it drives, and the measurements that put both inside the customer's simulator. For customers going mobile, a delivered locomotion capability under their policy rather than a research problem on their roadmap.
Motor, reducer, encoder and drive with Nora's firmware: FOC and torque loops, safe torque off in the drive, a frozen and versioned contract, and an end-of-line record per unit. Sourced from the Shenzhen and Suzhou supply chains, with a US counterparty and a path to US integration.
A 1 kHz executor with verified timing bounds, the bus master, sensing with hardware timestamps, and an independent safety MCU carrying the functions a learned policy cannot. Designed to ISO 13849 / 10218 patterns and tracking ISO 25785-1 as it moves toward publication.
Per-joint dynamics measured on a rig, exported as simulator assets validated against real trajectories, with randomization ranges from measured spread. Deployed fleets re-identify themselves and flag drift. This is the part of sim-to-real that is measurement rather than research, and it is redone by hand on every body today.
Model-first companies that started on tabletop arms are adding wheeled bases and legs. Most have no wish to become locomotion labs. Nora delivers a whole-body controller and reflex layer, frozen, verified and characterized, that runs under the customer's policy; customers who own their own low-level policy run it on the same controller.
The precedent is the automotive domain controller. When cars consolidated ECUs, the software-strong OEMs kept their stack and bought the deterministic hardware beneath it from Tier-1s who designed the box, wrote its BSP and safety firmware and manufactured it. Those Tier-1s (Bosch, Aptiv, Desay SV) became large companies on modest per-unit margins because attach was universal and the relationship outlived model years. Schaeffler is already moving into integrated joint actuators for humanoids; the actuator, reducer and screw suppliers in the Optimus chain report gross margins above 35% at pre-volume pricing.
The math is attach rate times bill of materials, plus a software line the component suppliers do not have. Actuators, reducers and screws are roughly two-thirds of a humanoid's bill of materials in the broker analyses of the Optimus chain. A Tier-1 that ships the modules, the controller and a per-unit characterization and safety license on top captures the component margin and a recurring software margin the pure component makers cannot charge for, on every unit, every revision. Unit volumes are early but real: thousands of humanoids shipped in 2025, plants announced in the tens of thousands a year, and model-first fleets planned in the hundreds within eighteen months.
The expansion path compounds on the same customers.
Three clocks are running. The model companies are in their first hardware generations, the window in which an Apple buys the modem rather than building it; that window closes when they either hire firmware organisations or, like Figure and 1X, decide the body is the moat. The July 2026 FCC action on foreign-produced mobile robots makes a US counterparty with Asian sourcing and a path to US integration the only shape of supplier that can serve both US and rest-of-world deployments. And the chip vendors are pushing brain-and-cerebellum integration into the compute module, which means the durable layer is the joint-level contract and the characterized modules beneath it, not any particular box; Nora is built around the contract, and the controller is its first physical form.
Nora is a US company. The founder trained at the MIT Media Lab in AI/ML, on the NLP side of the first deep-learning wave. The team has spent five years designing and implementing consensus, parallel execution, zero-knowledge and, most recently, formally verified infrastructure, some of it on bare metal, funded by roughly $15M in open-source grants from foundations such as StarkWare and the StarkNet Foundation; that infrastructure secures several hundred million dollars of economic value under adversarial conditions. Determinism, timing guarantees and verified correctness are what the team has shipped, and this proposal is scoped to the layer those strengths cover completely. Sourcing runs through the Shenzhen and Suzhou supply chains directly.
We are in first conversations with a leading model-first company about a 90-day pilot: characterize two candidate actuators to a shared schema, deliver simulator assets and a conformance spec, then an EVT sample of the controller running their low-level policy against their joints.