Cost per hour is how most robotics teams buy data, and cost per hour is the wrong metric. The four layers of the robotics data stack differ by orders of magnitude in what an hour costs. Yet the budgets that fail rarely fail on price. They fail through four predictable leaks that no hourly rate reveals, and the research of 2025 and 2026 documents the mechanism behind each one.
The stack is a pyramid. At the base sits internet-scale human video, effectively free, but carrying no robot actions and no proprioception (the robot's internal sense of joint positions and forces). At the top sits real robot data, mostly produced by teleoperation (a human remotely driving a robot to demonstrate tasks), where every frame carries the exact action and force signal a policy needs and every hour is slow and expensive. Between them sit two layers that try to buy strong signal cheaply: simulation and world-model synthetic data (video generated by models that predict how a scene evolves in response to actions), and egocentric (first-person) human demonstration.
The money moving through this pyramid is getting serious. Bessemer Venture Partners estimates aggregate robotic data costs will exceed $3 billion within the next two years, spanning every collection modality from teleoperation to tactile and force data. That is a named venture-capital estimate, not audited data, but the direction is hard to dispute.
We have already argued that data, not architecture, now decides who ships. What nobody has written down is how to spend. Consider this the audit: what each layer costs, what it alone can teach, and the four places the money leaks.
The pyramid prices abundance against signal
The models in question, Vision-Language-Action (VLA) policies (models that map camera images and language instructions directly to robot actions) and world models, train on data sources whose prices differ by orders of magnitude, and the gradient tracks the scarcity of embodied signal, not the usefulness of an hour. Bessemer's scale summary makes the asymmetry concrete: roughly 300,000 hours of robot manipulation data exist globally, against approximately 1 billion hours of internet video and 300 trillion tokens of text. That is roughly 3,000x more internet video than robot data, and the ratio is the entire reason the pyramid exists. Every serious lab imports capability from the abundant layers while spending as little as possible at the scarce top.
Internet video, the base layer, is the cheapest capability multiplier in the stack, and what it buys is physical priors: how objects move, how humans apply force, what plausibly follows what. Its real cost is curation compute: NVIDIA's Cosmos pipeline curated roughly 100 million clips from a 20-million-hour raw video collection, and Cosmos itself trained on 9,000 trillion tokens drawn from that footage (arXiv 2501.03575). "Nearly free" describes the footage, not the layer. Even the free layer is an industrial pipeline in disguise.
The canonical result is Meta's V-JEPA 2: pretrained on more than 1 million hours of internet video, its action-conditioned variant was post-trained on under 62 hours of unlabeled DROID (Distributed Robot Interaction Dataset) robot video and ran zero-shot on Franka arms in two labs at 65–80% pick-and-place success (arXiv 2506.09985). Under 62 hours. The cheapest layer plus a sliver of the most expensive one produced a working policy.
Ego4D anchored the layer earlier: 3,670 hours of first-person video from 931 wearers across 74 locations (arXiv 2110.07058) became the corpus that R3M-style encoders built on to establish that egocentric video transfers to manipulation representations. The base layer's signal is real but weak: it sets priors, not policies, and every result here required a top-layer shim of real robot data.
Simulation and world models convert GPU-hours and engineering into scale, with almost no human wages in the loop. NVIDIA says its GR00T-Dreams pipeline generated the synthetic data behind GR00T N1.5 "in just 36 hours" versus "nearly three months" of manual collection; NVIDIA states the underlying DreamGen pipeline expands effective robot training data roughly 10x and cuts manual teleoperation needs "by a factor of 20."
Those figures come from NVIDIA's GTC (GPU Technology Conference) and Computex 2025 keynotes and company blogs: internally consistent and directionally credible, but not independently audited. The compute is not trivial either. 1X reports its 14-billion-parameter world model needs 11 seconds of multi-GPU inference to generate 5 seconds of video plus 1 second of extracted actions (company-published figures).
Egocentric human demonstration is the current sweet spot on price, because it needs no robot at all. Multiple peer-review-track papers put collection at roughly 3x faster than teleoperation: DexViTac reports more than 248 demonstrations per hour from its handheld and glove system against 84 for teleoperation in the same comparison (arXiv 2603.17851), and the UMI (Universal Manipulation Interface) handheld gripper was more than 3x faster than spacemouse (a 3D joystick-style input device) teleoperation on cup arrangement while completing dynamic tossing tasks teleoperation could not do at all (Chi et al., RSS 2024; arXiv 2402.10329).
The flagship economics: 1X trained its world model on 900 hours of egocentric human video and only 70 hours of robot data, company-published figures from January 2026. The layer now has a scaling law too: EgoScale, from NVIDIA's GEAR lab (Generalist Embodied Agent Research), trained a VLA on 20,854 hours of action-labeled egocentric video and reports a "log-linear scaling law between human data scale and validation loss" that "strongly correlates with downstream real-robot performance," improving average success 54% over a no-pretraining baseline on a 22-degree-of-freedom hand (Zheng et al., arXiv 2602.16710, February 2026).
Real robot data sits at the top and costs the most per hour of any layer. Hardware alone: a base ALOHA (A Low-cost Open-source Hardware system for bimanual teleoperation) rig runs $18–20K (arXiv 2304.13705), Mobile ALOHA roughly $32K (arXiv 2401.02117), and dual Franka arms $80K or more.
On top of hardware sit operator wages and a hard throughput ceiling: a skilled operator collects roughly 20–40 demonstrations per hour (DexViTac's 84-per-hour teleoperation baseline marks the task-dependent high end), and a meaningful fraction gets discarded for quality. Jim Fan, NVIDIA's Director of AI and Distinguished Scientist, put the top-layer economics memorably at Sequoia's AI Ascent: "The real robot data is the human fuel. It's worse than fossil fuel. You're burning human fuel."
The price gradient is real. It is also the least useful fact about the pyramid.
Cost per hour is the wrong denominator
What an hour teaches varies far more than what an hour costs, and the best current evidence says the thing worth paying for is diversity, not volume. Lin et al., "Data Scaling Laws in Imitation Learning for Robotic Manipulation" (ICLR 2025; arXiv 2410.18647), found that generalization follows a roughly power-law relationship with the number of environments and objects, while "the diversity of environments and objects is far more important than the absolute number of demonstrations." Once demonstrations per environment cross a threshold, additional demonstrations of the same scene contribute almost nothing.
That result reprices every layer. An hour of teleoperation in a new environment with new objects and an hour repeating a solved scene cost the same on the invoice. They are not worth the same to the policy. (The apparent counterexamples confirm the rule: AgiBot's GO-1 and Physical Intelligence's π0 both report clean power-law gains in trajectory count, but that is scaling of a diverse, curated corpus, not of repeated easy cases.)
Four leaks follow from ignoring this. Most teams will recognize at least two.
Leak one: paying for repetition the model ignores
The first leak is redundant easy-case teleoperation that starves the long tail (the vast set of rare situations that individually appear almost never and collectively dominate deployment). Per Lin et al., demonstrations added to an already-covered scene saturate quickly, so the marginal dollar spent there buys approximately nothing, while the uncovered environments that will break the policy in deployment stay uncovered.
The datasets that moved the field spent on spread, not depth. DROID collected 76,000 trajectories and 350 hours, but across 564 scenes, 86 tasks, and 52 buildings, with 50 collectors at 13 institutions (arXiv 2403.12945). Physical Intelligence's π0.5 trained on roughly 400 hours of mobile-manipulation data inside a much larger co-training mix (97.6% of its training data came from sources other than the target robot) and cleaned kitchens and bedrooms in homes never seen in training, with performance improving substantially as training scaled from 3 to 104 homes (arXiv 2504.16054). Homes, not hours, were the axis that mattered.
The operational rule is simple to state and rarely followed: when marginal demonstrations in a scene stop improving held-out success, stop collecting that scene and move the rig. If a new task appears to need more than a few hundred well-targeted demonstrations per environment, the problem is your coverage, not your volume: add environments, not demonstrations.
Leak two: simulation compute without a real anchor
Simulation spend leaks when it substitutes for real-world validation instead of multiplying it. The steelman first, because simulation's wins are real: Tobin et al. introduced domain randomization (varying simulator parameters widely during training) for zero-shot transfer at IROS in 2017 (arXiv 1703.06907), OpenAI's Dactyl solved a Rubik's cube one-handed after training almost entirely in simulation (arXiv 1910.07113), and DeepMind's Genie 3 (August 2025, still a research preview) generates interactive worlds at 720p and 24fps with consistency over "a few minutes". These systems narrow the gap. They do not close it.
OpenAI itself reported that "domain randomization alone is not enough": friction, elasticity, and dynamics resisted modeling and required automatic domain randomization on top. Randomization covers random visual variation but misses the structured variation of real environments, meaning layout, object inventory, and deformability. We have covered why simulation isn't enough at length; the budget-level summary is that simulation converts compute into an approximation of data, and a second, quieter risk compounds it: model collapse, the degradation that comes from training a model on its own synthetic output.
Notice how NVIDIA itself structures its loudest synthetic-data claim. GR00T-Dreams trajectories are labeled with pseudo-actions from an inverse-dynamics model (a model that infers the actions that produced an observed motion), and their value is validated against real deployment: a force multiplier on real data, not a replacement for it. That anchor doubles as the guard against model collapse, because synthetic output validated against real deployment is not a model feeding on itself unchecked.
The audit rule follows. If the sim-to-real gap holds or widens as you add simulation compute, that is a reality gap, not a data shortage; more simulation compute deepens the hole, and real validation data is what climbs out of it. Simulation is a multiplier looking for something real to multiply.
Leak three: datasets depreciate, loops compound
A one-time dataset is a depreciating asset: it loses value the moment your embodiment, sensor stack, or deployment environment changes. Much of the market sells exactly that, and buying it feels productive because the invoice maps to a deliverable.
Cross-embodiment transfer works, which is what makes pooled data tempting. Open X-Embodiment assembled more than 1 million trajectories across 22 embodiments, 527 skills, and 60 datasets from 21 institutions, and RT-1-X improved average success 50% across five labs versus models built specifically for each robot (arXiv 2310.08864).
But naive pooling also fails in documented ways: 1X found that for in-distribution tasks already well covered by robot data, adding egocentric human data diluted the mix and produced neutral-to-negative changes. Pooled naively, it is negative transfer with good marketing. The same ablations cut the other way: omitting the egocentric data caused the model to overfit to simple grasping and fail at nuanced tasks like scrubbing. Which data helps is an empirical, per-task question, and answering it requires the one thing a static dataset purchase never includes: the loop.
The systems that win are closed loops, pipelines where deployment failures decide what gets collected, labeled, and retrained next: 1X's flywheel, Physical Intelligence's co-training recipe, and the data engines Waymo and Tesla describe publicly (public descriptions, to be clear, not audited pipeline documentation). Collection is infrastructure whether you treat it that way or not, and the layers compound when collection, 4D reconstruction (rebuilding a scene in three dimensions plus time from fused sensor streams), auto-labeling (machine-generating candidate labels for humans to check), human verification, and retraining share one pipeline. Vertically integrated infrastructure is the capability that compounds. A purchased dataset is the capability that decays.
Leak four: unvaried, unverified data scores 95% and collapses below 30%
Cheap data fails in two ways, and neither shows up as a line-item overrun. The first is data you never varied. LIBERO-Plus systematically perturbed leading VLA models across seven dimensions at evaluation and found "models exhibit extreme sensitivity to perturbation factors, including camera viewpoints and robot initial states, with performance dropping from 95% to below 30% under modest perturbations," adding that "models tend to ignore language instructions completely" (arXiv 2510.13626; CVPR 2026).
OpenVLA fell from 76.5% success to 1.1% under a single camera-viewpoint change. A benchmark score can be an artifact of everything you never varied and never verified.
The second is data you never verified. Verification gets cut from budgets because its absence is invisible until deployment: no headline benchmark number reveals whether the episodes behind it were checked. The counterexample is AgiBot World: 1,001,552 trajectories and 2,976 hours across 217 tasks, collected by more than 100 robots with human-in-the-loop verification, deliberately retaining roughly 1% failure-recovery data. Quality there is not an afterthought. It is the protocol.
Regulation is now pricing this in. EU AI Act Article 10 requires that training, validation, and test datasets for high-risk AI be "relevant, sufficiently representative, and to the best extent possible, free of errors and complete in view of the intended purpose," with documented provenance (a traceable record of where each sample came from and how it was handled), bias examination, and identification of relevant data gaps. Non-compliance with the high-risk data-governance duties can draw fines of up to 3% of global annual turnover or EUR 15 million, whichever is higher (EU AI Act Article 99). Provenance is now a performance lever and a legal one.
For anyone selling robots into the EU, that stops being a nice-to-have and becomes an auditable obligation, which structurally favors pipelines that can produce the evidence on demand. A data budget without a verification budget buys a more expensive demo; deployment goes to the teams that can prove the robot works, rather than show that it can.
The strongest objection: fleets will make collection free
The sharpest counterargument comes from 1X CEO Bernt Børnich, in a January 2026 Bloomberg Technology interview: "Your intelligence doesn't scale with the amount of data you can collect with humans anymore. It actually scales with the number of robots you've deployed." If deployed fleets generate data as a byproduct of work, why budget for collection at all?
He is right about the direction. Waymo's study with reinsurer Swiss Re (December 2024) analyzed 25.3 million fully autonomous miles and found an 88% reduction in property-damage claims and a 92% reduction in bodily-injury claims versus human drivers: 9 property-damage and 2 bodily-injury claims where 78 and 26 were expected, benchmarked against more than 500,000 claims over 200 billion miles of human exposure. The study is company-commissioned, but the claims baseline belongs to the reinsurer. That is what the top of the pyramid looks like at fleet scale: verified real data, in a closed loop, producing outcomes an insurer will underwrite.
But the fleet is the output of the loop, not a substitute for it. A fleet without reconstruction, labeling, verification, and retraining produces exhaust, not training data, and Waymo's result sits downstream of years of exactly that infrastructure.
And the transition he describes has a catch. A robot can be teleoperated at most 24 hours per day, and far less in practice, so the moment fleets lift that ceiling is precisely the moment when loop infrastructure, not raw volume, becomes the binding constraint. Fleets change who collects. They do not change what the data must pass through.
Budget by what each layer alone can do
If the audit is right, the allocation follows. Spend the least at the base, where open corpora and pretrained encoders like V-JEPA 2 already exist. Spend deliberately in the middle layers to buy scale and priors cheaply, always anchored to real validation. Reserve top-layer spend for the two things only real robot data provides: a small embodiment "shim" (1X's 70-hour figure is the benchmark), and diverse, verified coverage of the long tail. Then put real money into the layer nobody demos: the closed loop that collects, reconstructs, labels, verifies, and re-feeds. Everyone demos the model; nobody demos the pipeline that produced it.
The honest caveats. Fully loaded dollar-per-hour teleoperation figures circulating online come largely from data vendors' own marketing and could not be independently verified; treat them as directional, while hardware costs and the 20–40 demonstrations-per-hour range are better corroborated across independent papers. Auto-labeling speedup multipliers are mostly vendor claims as well; the most rigorous analog is PATO, a policy-assisted teleoperation system that showed large throughput gains in a controlled user study (arXiv 2212.04708). And EgoScale and DexViTac are 2026 preprints: early evidence, not settled science.
Here is what follows if the thesis holds. Budgets get restated in environments and verified episodes rather than hours. Synthetic data gets priced as a multiplier on real anchors, never as a substitute. Provenance moves from paperwork to product requirement as EU enforcement arrives. And the more than $3 billion in data spend Bessemer projects flows to whoever operates the loop, not whoever sells the most hours.
The pyramid's economics reward the operator of the loop, and the teams that internalize those economics stop buying hours and start buying proof. The teams that ship will be the ones that can show, for any frame in the training set, where it came from, who verified it, and which failure sent the rig back out to collect it.
Related: how a teleoperation session moves through a 4D pipeline and what VLA models need from their training data.
