There's a version of the robotics story everyone knows: a humanoid backflipping at a launch event, priced as though it already runs your household. It gets funded, and for good reason; It just isn't where we're spending our time.

Robotics and physical AI startups raised either $18.8 billion or $55.8 billion in the first half of 2026, depending on who's counting. Crunchbase says the former, excluding autonomous vehicles and defense; Dealroom's broader definition says the latter. Both already exceed any full prior year, and the three-fold gap between them tells you the category has outgrown anyone's ability to define its edges.

Our bet at Cade is that general-purpose humanoids are near certain, but further out than the funding implies, and for reasons no model update solves. The training data doesn't exist at the scale real deployment demands. Dexterous manipulation outside a controlled demo remains largely unsolved. And even if both were fixed tomorrow, humans still have to be willing to share a hallway, a factory floor, or a kitchen with something shaped like them.

That last constraint doesn't respond to engineering. Resistance hardens wherever a machine's decisions carry moral or physical weight. Why is it that Waymo reports 94% fewer serious-injury crashes than human drivers across 220 million autonomous miles, yet autonomous vehicles remain contested anyway? Human error is granted context and intent; Machine error is not. The automatic elevator is another illustrative example of the cultural barrier to technological adoption - the automatic functionality was commercially viable as early as 1900, but unable to achieve adoption until the 1945 New York operator strike, plus a reassurance campaign of soothing recorded voices and emergency buttons that mattered more psychologically than mechanically. If unattended elevators in public spaces took fifty years for widespread acceptance, how long will it take for humans to accept fully autonomous humanoids at-large in their homes?

So we're long on the interim to humanoids, and on one discipline throughout: take a bounded problem, solve it completely, own something that compounds every time it runs, and be honest about how much of the work a person is still doing.

| THE RIGHT SHAPE FOR THE JOB

Most commercial tasks, and most near-term residential ones, don't need a general-purpose body so much as one sized to the single thing it does. The error in most humanoid bets is conflating two kinds of generality that don't have to travel together. General intelligence is a software property that gets cheaper to copy with every unit,  while general embodiment is a hardware property that gets more expensive with every joint and certification it carries. A trained policy transfers across bodies for nearly free, but a hand doesn't. BofA Global Research puts actuators at 51% of a humanoid's projected 2030 bill of materials and dexterous hands at another 19%, concluding that humanoid cost reduction is mostly an actuator manufacturing problem rather than an AI one. Seven-tenths of the machine sits in the parts that don't get cheaper when the model gets smarter, which is why the smartest capital is buying transferability at the brain layer and leaving the body specialized: Skild AI raised $1.4 billion in January, led by SoftBank and Nvidia, for a cross-embodiment platform, not for a robot.

| THE HONEST BRIDGE: TELEOPERATION

The second shape is teleoperation, and quadrupeds are already earning their keep in manufacturing with a person quietly holding the controls. Waymo didn't declare itself driverless on day one; it ran safety drivers for years, and every mile became training data for the system that replaced them.

The logic is about data rather than humility. Autonomy fails at the long tail, the jammed pallet and the hallway that resembles nothing in training, and that tail is the one dataset you cannot scrape, buy, or simulate. Demonstration data is cheap and redundant. Intervention data is scarce and sits exactly on the distribution of failures blocking deployment. A teleoperated fleet isn't a company waiting to become autonomous; it's one being paid by customers to collect the only data that gets it there. Starlife, a Cade portfolio portfolio company, runs on this structure: remote-operated robots, deployed in a variety of commercial events where labor costs loom large.

So the metric worth underwriting isn't "percent autonomous," which anyone can flatter, but the operator-to-robot ratio and whether it compounds. Rising ratio with falling interventions is a real curve. A flat one is labor arbitrage with a robot-shaped interface, and should be priced accordingly.

| THE SAME LOGIC, IN SOFTWARE

A general-purpose platform for all of robotics fails for the same reason a general-purpose body does: unbounded scope, nothing compounding, priced on ambition. What we look for below the robot is what we look for above it, one unglamorous function solved completely, with an asset that thickens every time it runs.

Whether that function sells into one industry or many is a separate question, one of strategy as much as technology. A billing layer built only for warehouse robotics-as-a-service, or a certification practice that only ever touches surgical robots, can be a better business than a horizontal version of either, because the depth is the moat and the buyer is easier to reach. What matters is that the problem is bounded and something accumulates. Five problems remain ripe for innovation:

  • Data. Mecka AI, which we backed at seed, has scaled significantly in under two years on real-world deployment data that compounds and gets harder for a customer to leave behind.
  • Integration and deployment. Deloitte found only 25% of organizations have moved even 40% of AI pilots into production. Rollouts die through small, unclosed gaps compounding into an integration bill nobody budgeted for. Whoever closes that last mile earns the account, and keeps it.
  • Financing and billing. Invoicing, utilization tracking, residual value. An industry selling machines by the hour still lacks the back office to do it.
  • Safety and certification. Standards were written for deterministic machines. Nobody has cracked certifying a policy whose behavior emerges from a training run, and whoever does it for a given regulated setting will set the terms for everyone shipping into it.
  • Fleet orchestration. Hardware and models commoditize; coordination compounds. YC's Fall 2026 RFS asked for both an operating system for physical work and real-world data collection, two independent asks sitting under the machine rather than in it.

‍

This is also why the market stays fragmented rather than winner-take-all. The buyers are fragmented too, and a warehouse operator, a manufacturer, and a hospital system have little in common beyond needing the same problems solved in their own context. Some of that gets served by one company selling broadly. Much of it gets served by companies that went deep on a single vertical and became impossible to displace there.

| WHAT DRAWS US IN

What we’re looking for: A reason to stay that compounds, through contract, operations, or accumulated data, rather than one that resets with every model release. A problem that survives a smarter model, because if the value disappears when a larger company gives the capability away, the scope was never the moat. Claims that match the deployment, hybrid mode described as hybrid mode. Founders who understand a supply chain, a balance sheet, or a regulator, not just a codebase. And revenue we can underwrite honestly, where services-heavy is fine this early given a credible path to recurring.

We could be wrong on timing. What we're confident about is narrower: the form factor, the operator, the data, and the plumbing all have to exist before anything flashier earns its keep at scale.

| IF YOU’RE BUILDING THIS, GET IN TOUCH

Robots built for one job rather than one body. Teleoperated fleets with a real human in the loop and a ratio that's improving. Or any of the five layers underneath, built for one industry or for many.

We invest at pre-seed and seed, and we're glad to talk well before things are polished. Hybrid autonomy, services revenue, a pilot that hasn't converted yet: none of that puts us off, and we'd much rather hear it from you than work it out later. 

Email us at inquiries@cadeventure.com and tell us what's working and what still needs a human.

| SOURCES

‍