The physics of winning in Physical AI

I've been carrying around a mental model for Physical AI since the early days of Autopilot, back when we were still figuring out what the thing even was. The lesson that stayed with me wasn't about AI or sensors or neural networks. It was the gap between a working prototype and a production system spanning millions of vehicles.

That gap is not incremental. It's a chasm.

Physical AI is at that exact inflection point. In 2025, over 14,500 humanoid robots shipped globally, the first real transition from academic R&D to commercial deployment, and more than $14 billion poured into robotics startups. Yet beneath the record valuations and viral demos sits a structural reality most investors and operators miss.

The winners won't be determined by who has the best AI.

They'll be determined by who masters three interlocking problems: form factor economics, model architecture depth, and manufacturing scale. These are the load-bearing walls of Physical AI. Get any one of them wrong, and the whole structure eventually collapses.

The form factor debate misses the point

The hottest argument in robotics right now is whether humanoid or non-humanoid platforms will win. I've sat in enough investor meetings to know this debate can consume entire conversations. Both sides are missing the actual signal emerging from the market.

The humanoid thesis is backed by staggering capital. Figure AI raised over $1 billion in its September 2025 Series C at a $39 billion valuation, a 15x increase from $2.6 billion eighteen months earlier. Apptronik secured nearly $1 billion from Mercedes-Benz, John Deere, and Google. Tesla is converting production lines at Fremont to manufacture Optimus humanoids. The argument is the iPhone argument: our world is built for the human form (stairs, doors, tools, workstations), so one general-purpose humanoid replaces dozens of specialized robots, just as the smartphone absorbed dedicated cameras, GPS units, and music players.

What humanoid advocates don't talk about enough is the energy math.

Bipedal humanoids operate for two to four hours on a charge. Agility Robotics' Digit runs for 90 minutes at Amazon warehouses. Boston Dynamics' Atlas, for all its acrobatics, manages maybe 30 to 60 minutes of rigorous operation. The physics is unforgiving: bipedal walking is an inherently unstable inverted pendulum problem, so the robot is constantly about to fall over and must run high-frequency sensor feedback and balance correction just to stand still. A 30-kilogram bipedal robot burns 200 watts simply walking on flat ground.

Wheels eliminate that overhead entirely. Elephant Robotics' Mercury X1, a human-like upper body on a wheeled base, runs for eight hours: a full factory shift. The energy efficiency advantage of wheels over legs isn't marginal. It's 3-5x, and it translates directly into ROI a customer can calculate today.

This is why Sanctuary AI, after building one of the world's most dexterous humanoid hands, pivoted its Phoenix Gen 8 to a wheeled base in January 2025. Their explanation was blunt: customer feedback revealed that "bipedal legs are too frail to support a strong torso needed to carry out useful work precisely and safely." Russ Tedrake, who ran Large Behavior Models at Toyota Research Institute, has noted that practical legged locomotion remains "one of the fundamental unsolved problems in robotics."

So who's right? Both, on different timelines. The pattern in technology history is remarkably consistent: initial diversification, convergence toward general-purpose platforms, then re-diversification into optimized variants. Aviation went from biplanes to monoplanes to jets, but also spawned helicopters, drones, and cargo aircraft. Smartphones consolidated dozens of devices but didn't eliminate professional cameras.

My read: wheeled platforms capture the near-term value in structured environments like warehouses and factories, where their battery advantage prices into ROI immediately. But humanoids win the long game, because economies of scale compound in ways that are hard to see until they've already happened. Unitree's G1 humanoid starts at approximately $16,000 at current volumes. Bank of America projects humanoid bills of materials dropping to $13,000-$17,000 by 2030. At one million units per year, Tesla's stated target, unit costs could reach $10,000-$20,000: cheaper than the wheeled alternatives that currently undercut them. The same production volumes that make smartphones cost $200 instead of $2,000 will eventually make humanoids cost less than the specialized robots they're supposed to compete with.

Your robot needs a spinal cord

The form factor debate gets the attention, but the more consequential technical question is model architecture. The answer is coming from somewhere unexpected: neuroscience.

When you want to pick up a coffee mug, you're not consciously calculating every muscle fiber and its tension. You have an intention, "I want that mug," and your brain handles that high-level planning. A different system entirely handles execution: your spinal cord, managing the inverse kinematics of moving your arm without knocking anything over. These are fundamentally different computational problems. One is slow, deliberate, and tolerant of latency. The other is not: if it's too slow, you miss the mug or knock it over.

Every major Physical AI lab is converging on this exact insight, independently.

Google DeepMind released two separate models in March 2025: Gemini Robotics-ER (Embodied Reasoning) for high-level planning, and Gemini Robotics for physical execution. The reasoning model analyzes physical spaces, predicts trajectories, and generates step-by-step instructions. The action model converts those plans into smooth, reactive motor commands. This is literally a brain telling a spinal cord what to do.

NVIDIA's GR00T N1 makes the split explicit, citing Daniel Kahneman's "Thinking, Fast and Slow" framework. Its System 2 module, a vision-language model called Eagle-2, runs at a deliberate 10 Hz to interpret scenes and language. Its output feeds System 1: a diffusion transformer generating continuous motor commands at 120 Hz. That 12x frequency disparity isn't arbitrary. A planning model can tolerate 200 milliseconds of latency while deciding which tool to grab. A motor control model needs sub-20-millisecond response times, or the bipedal system collapses under gravity. Physical Intelligence's π0.6 takes a different approach, grounding learning in real-world experience rather than a fixed dual-system split: its RECAP reinforcement learning method reportedly doubled operational throughput on complex manufacturing manipulation tasks and cut human intervention rates in half in real-world logistics deployments.

Here's the strategic implication few people are discussing. After the LLM explosion, market analysts widely predicted that Physical AI foundation models would rapidly commoditize. That prediction misreads even the digital precedent: LLMs didn't commoditize, they hyper-specialized. Claude dominates long-context reasoning and coding; Gemini excels at multimodal data ingestion. Physical AI models will specialize even more severely, because the physical world imposes hardware constraints that text processing never encounters.

The high-level "brain" models may indeed trend toward commodity availability as tech giants release open-weights reasoning engines. The "spinal cord" models cannot. A motor control policy trained on Tesla Optimus's specific actuator weights, limb lengths, sensor latencies, and center of gravity cannot simply be downloaded into a Boston Dynamics Atlas or a Unitree G1. The execution layer remains bound to the proprietary geometry and physics of the hardware it controls.

This is why deep hardware-software integration will remain a defensible moat. The spinal cord won't commoditize.

Manufacturing is the variable everyone underprices

The variable investors most consistently underestimate in Physical AI is manufacturing capability. Not theoretical AI prowess. Not impressive demos. The brutal, low-margin economics of scaled production and global supply chain logistics.

Boston Dynamics is the cautionary tale. Regarded for decades as the pinnacle of robotic hardware engineering, by 2025 they had shipped over 1,500 Spot units and a smaller number of Stretch warehouse robots, generating approximately $130 million in revenue while costing parent company Hyundai over $150 million annually to operate. Leading mechanics without scalable manufacturing competence produces expensive laboratory demonstrations, not market dominance.

The automotive parallel is even starker. Rivian entered the EV market with over $24 billion in total investment, desirable designs, and exceptional software. After five years of production, it delivered 42,247 vehicles in 2025, losing approximately $85,000 per vehicle, and eventually abandoned its 2027 profitability target. What Rivian lacked was the institutional knowledge of scaled manufacturing: the tens of thousands of microscopic supply chain optimizations, tooling adjustments, and scrap reduction methodologies that separate a profitable assembly line from cash incineration.

There's a story from the early days of EV manufacturing, probably apocryphal but true in what it captures, of engineers driving to every electronics store in Northern California buying USB cables to keep the production line running because the supply chain wasn't ready. That solve-it-now mentality applied to manufacturing is what separates companies that survive from companies that run out of money. Nobody writes about it in the tech press. It's the difference between building a business and building a science project.

The contrast with Chinese manufacturers is stark. Unitree shipped 5,500 humanoid robots in 2025, more than every American humanoid company combined. Revenue grew 335% year over year to $235 million, and the company is profitable. AgiBot, founded three years earlier, shipped 5,168 units and captured 39% of global market share. Unitree's G1 starts at roughly $16,000; Agility's Digit costs $250,000. China controls approximately 70% of the global supply chain for humanoid robot components.

Tesla is probably the most credible Western manufacturing play in this space. The transferable advantages are real: custom actuators that share precision manufacturing with automotive gearboxes, integrated battery and thermal management, identical cameras and AI inference chips across vehicles and robots. But even Tesla produced only hundreds of Optimus units in 2025 against an original target of 5,000.

Pure software companies face the structural headwind from the other direction. Skild AI raised $1.4 billion from SoftBank in January 2026 at a $14 billion valuation to build an "omni-bodied brain" that can operate any robot for any task, and grew revenue from zero to $30 million in months. But a 466x valuation-to-revenue ratio assumes near-perfect execution of a business model that has never been done before. The Android analogy that justifies it cuts both ways: Android won massive market share but suffered fragmentation, lowest-common-denominator hardware optimization, and an inability to capture the integrated value Apple extracts through hardware-software co-design. Physical AI magnifies that problem with real-world physics.

Waymo versus Tesla illustrates the principle from another angle. Waymo operates a technically superior service right now: 450,000 paid rides per week across 10 cities, a 90% reduction in serious injury crashes over 200 million autonomous miles. But each vehicle costs an estimated $75,000-$200,000, and the service only works in pre-mapped, geofenced areas. Here's something I think about constantly: no human has ever driven with centimeter-accurate HD maps. A paper map is accurate to maybe 10 meters, and that's always been good enough. Waymo's infrastructure has to be built city by city; it can't scale to Nairobi or rural India or anywhere the economics don't support the investment. Tesla's camera-only approach costs $40,000-$50,000 per vehicle, draws on 6.9 billion miles of supervised driving data from over 5 million vehicles, and scales globally without per-city mapping.

The lesson for Physical AI: the approach that can scale production and data simultaneously will eventually dominate, even if it starts behind on raw performance.

The common denominator is data

Here's what I keep coming back to. Regardless of which form factor dominates, which architecture prevails, or which manufacturer achieves scale, every single player in Physical AI needs training data infrastructure. And the data requirements are qualitatively different from, and vastly harder than, anything the AI industry has faced before.

The largest open robotics dataset, Google DeepMind's Open X-Embodiment, contains just over one million real robot trajectories across 22 robot types. Frontier language models train on trillions of tokens. The gap is staggering.

The harder problem is what kind of data. Robotics data follows power-law scaling with a critical twist: environment and object diversity matters far more than raw demonstration quantity. A robotic foundation model doesn't generalize by observing an arm pick up an identical mug ten thousand times in a sterile lab. It generalizes by attempting the task across thousands of distinct kitchens, counter heights, mug geometries, friction coefficients, and lighting conditions. What robots need is data from thousands of distinct physical environments (homes, factories, warehouses, hospitals), not more repetitions in a single lab.

The architecture specialization I described earlier multiplies the complexity. A perception layer needs diverse visual and spatial data. A planning layer needs task-level demonstrations with semantic labels. A motor control layer needs high-frequency proprioceptive data, the robot's sense of its own joint positions and movements, at 50-1000 Hz from specific actuator configurations. Each layer requires its own pipeline, its own labeling standards, its own quality assurance. Synchronizing LiDAR point clouds, RGB images, depth maps, force/torque readings, and proprioceptive data across time is orders of magnitude harder than labeling 2D images.

And here's the flywheel: manufacturing scale creates data advantages that software-only companies cannot replicate. Tesla's 5 million vehicles generate an estimated 50 billion miles of driving data per year, feeding both FSD and Optimus development through shared perception architectures. More deployed robots generate more data, which trains better models, which make robots more useful, which drives more deployments. Right now, Chinese manufacturers shipping thousands of units are best positioned to start that cycle.

So the bet isn't on which robot wins. It's on the data infrastructure every robot requires. Whether you're building humanoids or wheeled platforms, training unified models or hierarchical specialists, manufacturing in Shenzhen or Fremont, you need annotation pipelines that handle 4D multi-modal sensor data with full traceability at scale.

The next 24 months separate demos from businesses

Physical AI in early 2026 presents a genuine paradox: record capital deployment alongside fundamental unsolved problems. Goldman Sachs projects a $38 billion humanoid market by 2035; Morgan Stanley's scenario analysis reaches $5 trillion by 2050. Yet Rodney Brooks, one of the most accomplished practitioners in the field, estimates "more than ten years" before the first profitable humanoid deployment. Both can be true. The market will be enormous. Getting there will be brutal.

The resolution is that Physical AI is not one problem but three, and nobody has solved all of them. Tesla has manufacturing and data but is behind on robot capability. Figure AI has capital and promising AI but no manufacturing track record. Chinese manufacturers have volume and cost but trail on foundation model sophistication. Physical Intelligence and Skild AI lead on model architecture but lack hardware integration.

One dynamic changes the calculus for whoever closes those gaps first: when a human worker learns a new skill, that knowledge stays in one brain. When a robot masters a novel manipulation task, that policy propagates instantly across millions of identically configured units via over-the-air updates. The hardware appreciates through software improvement. And the thing that makes the software improve is data.

The next 24 months will separate the companies that can bridge the gap between impressive demos and profitable production. The winners won't be the ones with the best AI or the flashiest hardware. They'll be the ones that understand what every manufacturing veteran knows instinctively: in the physical world, execution at scale is the only moat that matters.

And execution at scale starts with getting the data right.