A few weeks ago, I attended an event at Stanford with the topic of Humanity and AGI. There were a few companies that used the event as a marketing channel to demonstrate the capabilities of the humanoids they've built and are selling to consumers, some priced at $6,000 a piece. Two humanoids dancing on stage fell off it; one shattered the presenters' TV. When another humanoid almost fell on a person while taking the stairs to get off the stage, I wondered how close we really are to having a robot making an omelet for breakfast. With all the advancements in the AI space, why is it so hard to make a humanoid that can be invited to our homes to give us relief from activities we crave to outsource to robots the most, the house chores? As a strategic finance expert in frontier tech, I saw the performance failures of humanoids at the Stanford event as clear signals that developing robots that can actually be useful must be a very expensive endeavor. I wondered how strategic finance can get involved to help the companies in the robotics space to manage engineering challenges in a financially sound manner.
My investigative journey led me to a key conclusion that the humanoid industry has a structural bias towards spending capital where physics makes data cheap (locomotion, demos) rather than where the economic value actually lives (manipulation, hand dexterity), and closing the gap between demo-ready and commercially-ready humanoid is as much capital-allocation problem as it is an engineering one. Let me take you through the journey.
Robot
Aristotle said in his Politics: “If every tool, when ordered, or even of its own accord, could do the work that befits it... then there would be no need either of apprentices for the master workers or of slaves for the lords.” The idea of having robots supporting the advancement of humanity existed in our imagination since at least 350 BC.
Isaac Asimov defined a robot in his fictional universe as a computerized, mobile machine equipped with a "positronic brain" capable of performing complex tasks and governed strictly by the Three Laws of Robotics:
First Law: A robot may not injure a human being or, through inaction, allow a human being to come to harm.
Second Law: A robot must obey orders given by human beings, except where that conflicts with the First Law.
Third Law: A robot must protect its own existence, except where that conflicts with the First or Second Law.
Practically, I see two versions of robots:
1. Programmed Robot — a machine equipped with mechanical and electronic parts, which can accomplish a prescribed task in a physical world without the need for understanding it.
2. Intelligent Robot — a machine that requires an understanding of the world to accomplish a prescribed task in a physical world using its mechanical and electronic parts.
In both versions, Robot is a physical, tangible object. It is instructed to achieve a task, transforming one state of the physical world into another state. It has hardware and software that govern its operation. It has mechanical parts (body, legs or wheels, arms, hands, pickers, fingers) that interact with the physical environment and physical objects. It has electronic parts that make the mechanical parts move to manipulate the state of reality. However, a Programmed Robot, like the industrial robotic arms that assemble cars, does not understand what piece of a car it attaches to what piece of a car and why; it relies on humans and the operational process design to ensure that the right pieces are put together. Robots that run through Amazon warehouses to move packages have dedicated lanes and don’t need to understand the layout, just the lanes they are assigned to operate in. Programmed Robots don’t need to make judgment calls; they operate with a clear set of instructions in a structured environment. If they encounter situations not covered by their code, they simply stop execution of the task until a human assesses the situation and restarts the operation. Intelligent Robots, like autonomous cars, make judgment calls all the time. They need to understand which part of the road is safe to drive on, what road objects to avoid or fine to drive over, when to yield to another car or a human, when to squeeze between the cars changing lanes. They constantly analyze the physical environment, understand surrounding objects, make predictions, create alternatives for behavior, assess the cost function of each alternative, and make a call on what action to take with safety being the primary objective. For this article, I want to focus on a humanoid as a type of Intelligent Robot that has a lot of promise but also a lot of complications on the path to commercialization.
We entered a new stage of evolution with the large language models (LLMs) powering functionality of ChatGPT, Claude, Replit, and other AI tools. However, understanding the unstructured physical world is extremely complicated, more so than understanding the language world. Per Fei-Fei Li’s essay, A Functional Taxonomy of World Models: “Language models have given machines an extraordinary command of concepts, vocabulary, and reasoning, but the physical world, virtual or real, runs on a different substrate. Where language models learn the statistical structure of text, world models learn the statistical structure of space and time: how light falls on a surface, how a garden looks from an angle no camera has captured, how objects respond to force and follow the laws of physics.” Learning to understand the physical world requires a new codified version of the world, a different kind of data than the one that fed LLMs, a vast amount of that data, and maybe even a different kind of learning. Let’s break the complexity into pieces.
Hardware
Humanoid is not software, it’s a physical object navigating through a physical space and interacting with physical objects. Programmed Robots have been deployed in factories for decades, since the Industrial Revolution. But Intelligent Robots are a fairly new phenomenon; they haven’t gone mainstream yet. Deploying Intelligent Robots into environments designed for humans without radically changing those environments is the holy grail, it will unlock trillions of dollars of value. So the robotics industry and Venture Capital have dedicated a sizable amount of capital to developing humanoids that replicate the functionality of a human body. That would make their adoption easier as it won’t require a major overhaul of the physical infrastructure to support their operation in a world designed for humans. However, that path is not without its challenges:
- Movement.
The human body is a piece of art. Da Vinci’s Vitruvian Man is one of the most recognizable and all-time iconic images of Western civilization. The human body has ~350 joints enabling over 200 skeletal degrees of freedom (DoF), with approximately 80 utilized for everyday whole-body motion. The human hand itself is a miracle of nature with 27 degrees of freedom, capable of performing virtually unlimited sets of tasks that created the world as we know it. Replicating a fully human body and human hand in a humanoid is extremely challenging. Humanoid robots typically are built with 30 to 50+ degrees of freedom. A standard breakdown allocates roughly 12 to 14 DoF for the legs, 2 to 4 for the torso and neck, 6 to 7 per arm, and up to 20+ if using multi-jointed, dexterous hands. Each joint is a mechanical part with sensors that must generate data for training, execute commands when in production mode, withstand extensive hours of use, and be easily serviceable. Humanoids developed for demos might not need to handle the complexity that comes with dexterity. However, for commercial applications, dexterity is one of the key drivers for customers’ willingness to pay and adoption rates. Requirements for dexterity of the body and arms come with major hardware challenges, data requirements, and heavy price tags for the developer and the eventual consumer.
- Power.
Humanoids need to be able to operate without being constantly plugged in. Most humanoid platforms on the market can operate for two to four hours on a single charge. That might be okay for personal use and a task to clean a house, but not sufficient for deployment in industrial settings. And that challenge can’t be just solved by choosing a bigger battery. A humanoid must carry its energy source inside a human-shaped body and balance that mass over two feet. A wheeled design might address that challenge by storing a bigger battery in the base of the humanoid and reducing power demand for balancing the robot. But wheeled humanoids substantially reduce the scope of practical applications. Bipedal, or legged, humanoids, like Tesla Optimus and Figure 03, have a much broader set of use cases making them significantly more commercially attractive. But bipedal design choice creates limitations for the size of a battery and where that battery can be placed without wrecking the robot’s center of gravity. Advanced battery technologies like solid-state cells, silicon anodes, and high-energy lithium-sulfur chemistries could solve the energy problems by vastly increasing energy density, but those improvements must outpace the additional energy demands created by rising complexity.
- Technical limitations.
Use cases for humanoids include deployments for dirty, dangerous, and demeaning jobs (3D). Think of the following examples: dirty - handling unsanitary or toxic materials, dangerous - repairing outside of the space station, demeaning - monotonous sorting cycles that reduce humans to automation machines. There is a high demand to deploy humanoid robots in environments involving extreme physical risk or extreme conditions. Making standard robot joints and wiring harnesses durable enough to survive these conditions adds weight, cost, and power demand. Harsh environments degrade mechanical joints quickly. Dust, water, and extreme heat ruin sensitive electronics. Walking on uneven, slippery, or shifting surfaces and handling steep stairs and tight spaces is very difficult for humanoids. Imagine what happens if a humanoid carrying hazardous material loses power or one of its knee joints malfunctions, making the robot fall due to gravity; that can cause a secondary accident with catastrophic consequences. Underground or heavily shielded facilities can block wireless signals, making real-time control and monitoring impossible. 3D tasks might require special tools that place additional demands on the dexterity of the robotic hand, requiring dozens of miniature actuators, sensors, and cables packed into a tiny space. This intricate wiring is highly prone to failing under rough conditions.
These technical challenges are very expensive to solve. They are highly dependent on the use case, cost of engineering roadmap to solve them, and customer willingness to pay. Enhancing hand dexterity, solving for durability and withstanding harsh conditions, addressing power limitations, and expanding versatility of deployment use cases impacts BOM, maintenance costs, warranty, replacement cycles, and cost-to-serve, i.e. TCO, which is the number the buyer actually cares about. 2-4 hour battery life is not an inconvenience; it's a utilization ceiling that drags down ROI and caps the value a buyer gets from robot’s deployment. In my own experience, I witnessed the consequences of engineering roadmaps designed without finance inputs, setting engineering objectives around demonstration of capabilities at the expense of investing in the actual capabilities, failing to account for full cost impact of hardware architecture in the product pricing, and scaling with negative unit economics. Those are expensive decisions that can be avoided if finance is brought to the table to partner with the C-team ahead of time.
World Model
We don’t know with 100% certainty what the physical world is truly made of (particles, energy), or even whether fundamental reality is not material, but mental or experiential, and what we call matter is how that reality appears. But we do know for sure that physical reality is not made of words. We interact with the world through our senses. We feel it through vision, hearing, smell, taste, and touch. We have spatial awareness to navigate through the physical world without stumbling over objects on our way. We recognize objects and their functional purpose even before we assign a word to that object. We understand how to interact with objects in a manner that accomplishes a task created in our mind. We have the capability to predict the outcome of our interactions with the physical environment and the objects in it before we execute the action. By observing how the world and the object respond to our action, we build a more comprehensive understanding of the range of possibilities for our world interaction model.
And that’s what I have in mind when I say “world model”. It’s not the simulated version of the world like the one you see in the video game. It’s a dynamic interaction model between physical action and the state of an environment. A transformer LLM is trained on next-token prediction: given a sequence of tokens, estimate the probability distribution over the next one. A world model predicts the next state of an environment, conditioned on an action. World models for humanoid robots are generative neural networks that encode an environmental state, predict future physical consequences conditioned on a robot's actions, and simulate physics in a latent space. LLM is text-in, text-out. World Model is action-in, future-world-out.
There are many competing schools of thought on how a robot should model the world, or whether it needs an explicit world model at all. There is a Latent Dynamics Model, which predicts how an environment evolves inside a compact, abstract mathematical space rather than generating full pixel-by-pixel videos. The most prominent example is Meta's V-JEPA 2. There is a Generative Video Prediction Model, which creates realistic, high-definition future video frames of a scene given the robot’s current camera view and an intended movement. Nvidia Cosmos, DeepMind's Genie are amongst the most instructive production examples. There is a World-Action Model, which is not really a separate model, but a fusion approach that blends video generation and continuous motor-control policies into a single architecture, mapping both future visual outcomes and precise joint movements. There is also a camp that argues that you don't need an explicit world model at all. Figure's Helix, Physical Intelligence's π₀ are direct vision-language-action imitation: map observation plus instruction straight to action with no future-prediction step. Which approach is the most efficient and future-proof, only time will tell.
Regardless of what world model architecture is used by the humanoid developer, model training needs data, and data for training intelligent robots is not cheap: text is nearly free and effectively infinite, while physical-interaction data has to be generated through simulation, teleoperation, human video demonstration, or real robot time, all of which are slow and expensive. The most valuable stream, the action, is the one the internet can't provide. There are endless videos of people doing things online, but with no record of the motor commands that produced the motion, making those videos less useful for modeling interaction. It’s not just the data economics that is a bottleneck, it’s also data complexity, variability, duration, and interrelations that real-world deployment demands. Data economics is a critical area that requires the involvement of strategic finance. What data flywheel to build, what data to collect, what data to manufacture, what functionality that data should enable, what customers are going to pay for that functionality, and how that impacts hardware architecture and margins of commercial units — these are finance questions that need answers before data consumes the capital provided by investors.
Data
Let’s dive deeper into data.
The most advanced consumer-facing robotics product, in my opinion, is a robotaxi. Waymo has been taking over the streets of major cities, demonstrating that it is possible for a machine to fulfill a task operating in a physical world sharing it with regular people. By summer of 2024, when Waymo opened its robotaxi to all San Francisco users, Waymo had driven about 20 million fully autonomous miles on public roads vs. tens of billions in simulation. This gap is exactly the economics of physical data. And that was for modeling behavior in a constrained environment — roads designed for vehicles. The surface where autonomous cars can operate is clearly marked (mostly), there are rules of behavior for all participants, there are road signs that apply to all participants, the primary goal is to avoid collision with other vehicles and objects on the road (people, dogs, stationary and moving obstacles). We designed roads for vehicles and placed intelligent robots inside that structured physical environment. Now, if we want intelligent robots to operate in space shared with humans or even come inside our house, that’s a whole different game. That environment is unstructured with more ambiguous rules of behavior. There are a lot more objects that interfere with achieving the task and a lot of variability of those objects. The cost function that drives choice between alternatives is unclear. The interaction model with the world is more physical (touching, manipulating objects rather than avoiding them, as autonomous vehicles do).
Deployment of intelligent robots in the real world has a vastly more complex functional task than the autonomous vehicles. The goal of an intelligent robot is not to avoid the objects, but to interact with and transform them from one state to another. There are many nuances to consider, a few I can think of:
1. What kind of data to collect for training? Many companies collect video data and physical motion data to build connections between mechanical interaction with the objects and visual representation of different states before and after the interaction occurred. This is similar to how a child experiences the world, seeing, interacting, and understanding the connection between the two.
2. What point of view to take? We, humans, experience the world from the ego-centric point of view; we see the world with our eyes and place ourselves at the center of experiencing the world around us. So many companies collect data from the same ego-centric point of view of a robot connecting video feed from the camera with the physical motion of the arms controlled by a human through teleoperation. The idea is to collect data from the ego-centric point of view of the robot connecting its perception with the mechanical motions of the arms/legs. However, humans can also imagine the world as a 3D space outside of their own selves that helps us with predictions of how the world will change in response to our actions. We can create a 3D model of every object we interact with to assist us with achieving the purpose of that interaction.
3. What / how many sensors to build reliance on? Robot needs to understand its surroundings and its placement in those surroundings. It also needs to understand the nature of the objects it is expected to interact with to accomplish the tasks and identify those objects in its environment. Humans have a very good depth perception and very sophisticated body that supplies the brain with sensory experience data. We understand how gentle we need to be with the object or how much power to use for handling a heavy object. We can even manipulate the objects without seeing them, just relying on the sensory inputs. That poses questions about whether we should assign different purposes for the visual perception sensors and tactile sensors? Robot needs to understand how to interact with the objects to accomplish the task without breaking them. That involves building a connection between observation and action, how to manage mechanical parts (hands, joints, grippers, force, tactile sensing) of the robot to control interaction with the objects. Collecting sensory data is the problem that many companies are trying to solve, but that data is dependent on the hardware itself, so if hardware changes, the sensory data that controls the movement of that hardware is likely to change as well. We’ve seen videos of robots breaking eggs and making omelets. But can that robot also pick up salt shakers that take different forms in our kitchens and put the right amount of salt into the egg mixture? Can it measure and put the flour into the mixture to make muffins? That’s an extremely complicated multi-faceted data problem from both visual understanding and sensory experience perspectives that is hard to solve without a meaningful capital infusion.
4. How to convert world interaction data into training data? How to represent physical motion data? What kind of labeling approach to apply to perception, motion, and action data? Children learn to interact with objects before they can label them; they learn to understand the functional purpose of different objects without the need for language. So is it helpful to go through the pain of labeling objects?
5. When to use real data and when to use simulation? Simulation generates trajectories at near-zero marginal cost with perfect labels and supports reinforcement learning where trial-and-error is free. The cost is the sim-to-real gap — contact physics, deformables, and friction are exactly where simulators are weakest and exactly what manipulation depends on in the real world deployment.
6. How to leverage data from the deployment of the robot itself? Establishing a data flywheel to collect data from the deployment of robots that feeds into retraining of the model allowing for continuous improvement is a great idea that came from the software world. However, expectations for capabilities and safety metrics of deploying hardware in environments where they interact with humans are much higher. The robot is expected to execute its functional purpose in a safe manner before it can be commercially deployed, so leveraging deployment data might prove to be difficult for certain robot use cases.
Training a humanoid takes a massive amount of data. And not just one dataset, like text that is used for training LLMs, but several distinct data streams, each answering a different question about the robot's understanding of the world around it and supporting execution of the assigned task. I group these data streams by their purpose:
Vision - “where I am and what's in front of me?”.
This understanding comes from cameras on the head or torso to read the scene, and often from cameras mounted on each wrist for a close-up view of whatever the hands are manipulating. The wrist views matter because they see the contact up close and don't get blocked by the robot's own body. In some situations, it might be helpful to add LiDAR sensors that sharpen depth perception and spatial understanding.
Proprioception - “what shape I am in right now?”
The core signals are joint positions (the angle of every joint) and joint velocities (how fast each is changing), read from encoders on each motor. From those angles plus the known lengths of the links, forward kinematics computes where every part of the body is in space, where the hand is, where the elbow is, how the torso sits relative to the hips. This is entirely internal: it tells the robot's pose without any reference to the outside world or to gravity. A robot lying on its back and a robot standing upright can have identical proprioceptive readings if their joints are bent the same way.
Balance and inertial state - “which way is down, am I tipping, and am I about to fall?"
The anchor sensor is the IMU (inertial measurement unit), which measures linear acceleration and angular velocity — the same principle as a human’s inner ear. From the IMU, plus the proprioceptive pose, plus the foot force readings, the system estimates which direction is truly vertical, where the center of mass sits, and where the center of pressure under the feet falls relative to the support area. This is the only one of these streams that is about the robot's body as a single object oriented in the world, not about any one joint or contact point. Balance is also what couples everything together. When a humanoid reaches for a cup, that movement shifts its center of mass, which forces the legs to compensate, which changes the torso pose, which moves the cameras and the arm's base frame. Manipulation and locomotion aren't separable in a humanoid, because they share one body, one balance problem, and one power budget.
Force and torque - “what is pushing on me, and how hard?"
These come from dedicated load cells, usually six-axis sensors at the wrists and ankles, plus torque estimates at the joints. Where proprioception is not designed to understand contact with the environment, force and torque are entirely about the boundary between the robot and everything it touches. When the gripper presses a button, when a foot lands on the floor, when a hand lifts a heavy box versus an empty one — proprioception can't tell those apart, because the joint angles might be identical, but the force sensors read completely different loads. This is the stream that helps with questions: "how firmly I am grasping," "did I make contact yet," "is this heavier than expected," "is my foot actually bearing weight." It's the sense of effort, not the robot’s position.
Action - “what I am trying to achieve?”
Everything above is input data describing the robot's situation. The action is the output: the actual commands issued to the mechanical parts to execute the task. This is the answer the model is being trained to produce. It's the single most valuable stream and the single most expensive, and it only makes sense paired with all the observations around it — time-synchronized egocentric video, joint positions and velocities, torques, and tactile readings, alongside the command that was actually sent. Actions are collected through teleoperation, motion-capture suits, egocentric human video, or simulation. Natural-language task instructions are paired with the specific physical episode they describe. Instruction to "pick up the mug" is attached to the exact trajectory of movements that did it. This is what lets the robot connect a goal to a physical behavior.
If the above wasn’t enough to signal the complexity of the AI model training for humanoids, I want to mention a few data challenges:
- Time synchronization.
There is a difference in how quickly a model can process incoming information from perception and motion sensors and how quickly motion sensors need to react to a change in the environment. In a nutshell, it’s the difference between mind processing and body reaction. A vision-language model with billions of parameters has to encode several camera images, run them through dozens of transformer layers with the instruction, and decode an action. On the compute a robot can actually carry in its chest, that takes something like 100 to 200 milliseconds. So it can produce maybe 5 to 10 decisions per second. But physics doesn't wait for that long. If a humanoid's foot slips or its ankle torque is slightly wrong, the fall develops over tens of milliseconds. To correct it you need to sense and respond every one or two milliseconds. That's 500 to 1000 times per second, or about 100X faster than the model takes to think. Same for grasping: the moment a finger touches an object, forces spike, and if you don't modulate within a few milliseconds you either crush it or drop it. You get an inversion of authority: the system component that understands the task can't react in time, and the component that can react in time doesn't understand the task.
- Limitation of simulation.
Sim-to-real works great for locomotion, because rigid-body contact with the ground is something physics engines model reasonably well. However, it works poorly for dexterous manipulation, because grasping involves soft contact, friction, deformable objects, and multi-finger force closure — all the places simulators are least accurate. That explains why you see so many robots walking, jumping, doing flips in the air, but you don’t see many robots using their hands to open a can of Coke. But it’s the hand where most of the economic value of a robot actually is.
- Compounding errors over long horizons.
Behavior cloning learns to imitate the demonstration distribution. The moment the robot’s actions and environment deviate slightly from that distribution, its observations become unfamiliar, its next action gets worse, and the deviations accelerate. Humanoid tasks are typically long; for instance, "unload the dishwasher" is dozens of sequential sub-behaviors, each of which has to end in a state the next one recognizes. Success rate compounds multiplicatively: eight steps at 95% each gives you 66% overall. That might not be sufficient for customers to pay for.
- Failure recovery might be missing from the data.
Human demonstrators are competent. They rarely drop things, and when they do, the take usually gets discarded. So the training set is full of successful execution and nearly empty of failure recovery, which means the robot has no idea what to do once something goes wrong, which is precisely when you need it to know. Deliberately collecting failure-and-recovery data is expensive and unnatural for teleoperators. Another issue arises when exploration of failure modes means breaking hardware or destroying the environment. Those instances often require a human to stand the robot back up and rearrange the scene. It’s very different from reinforcement learning deployed for software AI, and, you can imagine, very expensive.
- Knowledge transfer is imperfect.
Training a model on one robot doesn't automatically make it work on another. The robot's body is the interface to the world, and every body is different. A humanoid trained on one set of joints, motors, and sensors will have a different experience than a humanoid with a different set. The model has to learn how to map its understanding of the world to the specific kinematics and dynamics of its own body.
The field's main data hedge is pooling across robot types, but a humanoid's kinematics, workspace, and dynamics differ enough that knowledge transfer can only be partial. Even two humanoid models from different vendors have different joint limits, mass distributions, and hand designs. Some knowledge can transfer more easily (semantics, task structure), but the low-level motor mapping largely doesn't. Every hardware revision risks partially invalidating the data collected for the model training with the older generation hardware.
This is exactly the problem the robotics field is trying to solve. Rather than accepting that intelligence is trapped inside a specific body, some companies are trying to decouple the intelligence layer from the hardware entirely. For example, Skild AI is building a single foundation model on data pooled across many robot morphologies (quadrupeds, humanoids, arms, mobile manipulators, and human video) to create "omni-bodied" Skild Brain, an intelligence layer that can power different robotics applications. Pooling data across many bodies becomes a scalable solution for the robotics data problem. If that approach works, and Skild AI has published early demonstrations of this working across body types, the diversity of hardware stops being a tax on the data and becomes the thing that feeds the flywheel.
Why is this a strategic finance problem?
Every problem mentioned above is eventually requiring a decision about where to spend a finite amount of capital. And the humanoid industry has a structural bias toward spending it in the wrong place.
For instance, let’s talk about gravity again. Simulation is cheap for locomotion and nearly useless for manipulation. So the cheapest data produces walking, jumping, and backflips — the things that make a compelling demo and raise the next round. But the true economic value of a humanoid sits in the hand, in reliable manipulation of objects with valuable use cases, which is exactly the data that has to be manufactured slowly and expensively, commonly through teleoperation and real robot time. A company optimizing for what it can show on stage and a company optimizing for what a customer will pay for are pointed in different directions. Left alone, engineering and marketing pull toward the demo. The job of strategic finance is to resist the temptation and to influence capital allocation towards the boring but more commercially oriented data collection and training, because that is what eventually will push the company from being an entertainment company to a commercially viable business.
Willingness to pay for a physical task is a step function tied to reliability metrics: if a humanoid can accomplish the task only 66% of the time, customers are highly unlikely to pay for it. It has to perform with an acceptable level of reliability to be deployed in the real world and paid for. This means there is a wide valley between the impressive demo and the first payable task, and a company has to cross that valley on investor capital with no revenue underneath. What commercializable use cases to target first, what hardware setup that will require, how much capital it takes to achieve the reliability threshold, what’s the path to expanding the scope of capabilities in a capital-efficient manner should be amongst the most important questions on the C-suite agenda, and these are finance questions before they become engineering ones. Market sequencing is a capital allocation problem, the one that strategic finance must be tasked to evaluate before the engineering roadmap is locked in.
The unit economics also doesn’t behave like it does in pure software products, and pretending that it does is how these companies die. In software, the marginal cost trends toward zero and scale turns into margin. Here it doesn’t always behave like that: failure-and-recovery data has to be deliberately produced and never stops being expensive, new use cases might require new data collection efforts or new hardware support, every hardware revision partially invalidates the data behind it, and the sensor stack that generates the training data is itself a cost and a constraint. On the hardware side, a two-to-four-hour battery is a hard ceiling on utilization, and utilization is a key revenue driver for a robot. The number the buyer actually computes is total cost of ownership, including purchase, power, maintenance, warranty, replacement cycles, not the sticker price quoted for a demo robot. A $6,000 humanoid that falls off that stage is not a $6,000 product. The sticker price is a number that has to be reconciled against a bill of materials, a cost to serve, and maintenance cycles before it means anything. And that, again, is where strategic finance must be heavily involved to create a full economics profile and value proposition for the customer in order to guide pricing and engineering decisions.
Decoupling intelligence from hardware isn't only an engineering choice; it's a choice that impacts capital structure, unit economics profile, and value pool allocation. A company that builds the brain independent of the body can amortize combined data-collection cost across many robot platforms and customers instead of funding new model training with every hardware revision. It also can choose to push the costs of hardware, maintenance, and operations to their customers, supplying and monetizing the intelligence layer only. That is a more attractive capital story than vertical integration, of course assuming that the generalization across different hardware architectures actually holds. If it does work, there are still many strategic finance questions to address, including data economics, value share framework, pricing structure, integration / adaptation costs recovery, customer support framework, etc.
So what does strategic finance actually need to do? It needs to decide which task to commercialize first, mapping the reliability thresholds against customer willingness to pay. It needs to allocate capital across the data strategy — real robot time vs. teleoperation vs. simulation vs. human video, as each one has a different cost structure and risk profile. It needs to choose the business model — outright sale, robotics-as-a-service, outcome-based pricing — which is really a decision about who carries the utilization and reliability risk as the technology matures. It needs to establish the pricing guidelines that correspond with the value generation. It needs to build the investor and board narrative that models the unit economics honestly, instead of importing software assumptions that don't hold. And it needs to enforce the discipline of tying capital deployment to commercial-ready milestones rather than demo milestones.
None of these are questions engineering can answer, and none of them can wait until the technology is finished, because the technical choices and the financial ones are the same choices. The hand design sets the bill of materials and the data cost. The data strategy sets the capital profile. The first market sets the runway. A humanoid company that treats finance as the team that books the numbers after the fact will make its most expensive decisions blind.
The robots will make breakfast eventually. But at the cost of many companies failing to account for economic realities of hardware-software products or failing to run engineering choices through the lens of capital deployment optimization. The companies still standing to sell humanoids will be the ones that were as rigorous about the economics as they were about the engineering. And economic rigor is what strategic finance brings to the table.
August 2026