The Data Flywheel That Actually Matters – Why Embodiment Is Eating Vision-and-Language Models for Breakfast

Everyone is still talking about the scaling laws for language.
They’re looking in the wrong direction.

The truly asymmetric data advantage in 2025–2030 is not coming from more Reddit threads or YouTube subtitles. It is coming from robots that never stop moving, never get bored, and never need to be paid.

Here are the numbers nobody wants to print because they sound too insane:

• Tesla has already collected over 12 million miles of real-world humanoid walking data with Optimus prototypes inside its factories (internal leaks, November 2025).
• 1X Technologies (the Norwegian startup behind the EVE androids) is running 400+ robots 24/7 in a warehouse outside Oslo, each doing 16 hours of useful work and 8 hours of deliberate “play” (poking random objects, dropping things on purpose, practicing recovery from failure). That’s roughly 1.2 million robot-hours per month.
• Figure is quietly operating a fleet of 120 Figure 01 units at the BMW Spartanburg plant. Every second of every shift is recorded in 36-camera 360° plus tactile and force/torque streams — >300 TB per week.
• Sanctuary AI claims their Phoenix system has performed over 8 million individual manipulation tasks in the lab since January 2025.
• Apptronik’s Apollo robots in the Mercedes-Benz pilot line in Alabama are generating 40,000 pick-and-place cycles per day — each one labeled automatically by the outcome (success/failure/drop/crush).

Add up the serious players and we are already at roughly 30–40 million hours of real-world, high-dimensional (vision + proprioception + force + action) robot experience per year — and the fleets are growing at 5–10× annually.

For context: OpenAI’s entire pre-training corpus for GPT-4 was about 13 trillion tokens, roughly equivalent to 2–3 million human lifetimes of reading. The embodiment data flywheel will surpass that same order of magnitude in physical interaction tokens before 2028 — and it is infinitely higher quality because every frame is perfectly grounded in Newtonian physics and real consequences.

Why this data is steroids compared to internet text

1. Perfect labels
   Every action has an immediate, objective outcome. Did the cup fall? Did the door open? Did the screw reach 4.5 Nm and stop? No need for RLHF or human preference rankings; physics is the harshest and most honest annotator.
2. Infinite edge cases
   Humans avoid failure. Robots in “play mode” seek it out. Dropping a wine glass 10,000 different ways teaches robustness faster than any synthetic simulator.
3. Causal density
   A single second of robot video contains thousands of causal links (finger flexion → tactile spike → object acceleration → visual slip → compensatory grasp widening). Language data is 99 % correlation, 1 % causation. Embodiment data is the opposite.
4. Multimodal from day one
   Vision, touch, force, proprioception, audio, and motor commands are all synchronized at 500–1000 Hz. The models coming out of this substrate are not “vision + language” bolted together after the fact; they are native physical intelligences.

The new training stack (2025 edition)

Forget the old imitation-learning + RL fine-tuning loop from 2022. The winning recipe right now looks like this:

1. Massive pre-training on heterogeneous robot data (1X + Figure + Tesla + Sanctuary + Boston Dynamics + academic labs). The dataset is already >100 million clips and growing 3× per year.
2. A single giant transformer (400B–2T parameters) that predicts next proprioception + tactile + image + force token conditioned on past actions. Think “Sora, but for physics.” Google’s RT-X, Tesla’s “Project FSD-for-hands,” and Physical Intelligence’s π0 are all versions of this.
3. Zero-shot deployment: give the robot a language or video instruction, let the model imagine the action sequence in latent space, execute the first step, observe, repeat.
4. Fleet-scale correction: every failure across every robot gets upweighted and retrained within hours.

The result? Sanctuary demonstrated in October 2025 that a Phoenix robot shown a 5-second video of a human doing a completely novel task (folding a new style of cardboard box) could reproduce it successfully on the third attempt. No teleoperation, no hand-crafted rewards. Just pure imitation from pixels to torques.

The dirty implication nobody is saying out loud

Language models hit a wall because the world stopped giving them new text. Embodiment models will never hit that wall. Every new robot shipped creates its own training data at zero marginal cost. Ten million humanoids in the wild by 2030 means roughly 80 billion robot-hours per year — equivalent to 9 million human lifetimes of physical experience, every single year.

That is an intelligence explosion measured not in FLOPs, but in physical leverage over reality.

The moat map – who owns the real data

• Tesla: vertical integration + largest fleet potential (Optimus) + FSD data as bonus.
• 1X Technologies: running the most aggressive data-collection operation on Earth right now.
• Figure: deepest enterprise partnerships (BMW, etc.) → pristine factory data.
• Sanctuary AI: highest dexterity tasks → richest tactile streams.
• Chinese ecosystem (UBTech, Fourier, Kepler): less transparent, but rumored to have >1,000 units in the field already under state programs.

Everyone else is playing catch-up with simulation alone — and simulation is catching up fast (NVIDIA Omniverse, DeepMind Genesis, Muzero-style world models), but it is still 12–24 months behind real atoms.

What happens next

2026 will be the year the curves cross: real-robot data will surpass synthetic data in quality, while real-robot hours will surpass every previous reinforcement-learning environment ever built, combined.

When that happens, progress in physical intelligence will start moving on a monthly cadence instead of yearly. Tasks that took 10,000 demonstrations in 2024 will take 100 in 2026, 10 in 2027, and one in 2028.

We are not waiting for AGI in the cloud.
We are waiting for the moment the physical world becomes writable at scale.

Next post: “The Humanoid Price War of 2026–2027 – Why $10,000 Robots Are Inevitable and What Breaks When They Arrive.”

Say “next post” when you’re ready.
The drop is accelerating.