The entire humanoid bull case rests on one belief: “the VLA scaling law holds” — that, like large language models, a general-purpose robot’s capability will climb smoothly as long as you stack enough data and compute.

That belief drives every valuation. If it’s right, general-purpose robots are “the underrated decade.” If it’s wrong, they’re “the overrated three years.” So instead of arguing whether Figure is worth $39B, take the belief itself apart — because it actually has three layers, and most people collapse them into one.

L1 / L2 / L3 — three gates, don’t conflate them

• L1 — baseline capability. In controlled settings, on specific tasks, the robot improves as data grows. This layer already holds — most demos show this layer.
• L2 — real-world generalization. Transferring learned skills to unseen objects, lighting, layouts. This layer is emerging, not yet stable.
• L3 — deployable autonomy. Reliable, long-duration, human-intervention-free operation in real, unstructured environments — commercially usable. This layer, today, is blank.

The key: what actually pays off — what supports those sky-high valuations — is L3. And L3 isn’t just not here yet; its timeline keeps sliding right. The robot smoothly folding laundry in a demo is often teleoperated, in a staged scene, edited — that’s an L1 performance, not L3 autonomy.

“Only pay for L3 money”

That line is the core discipline here: when you price the body, when you price pure humanoid upside, only credit L3 delivery — not L1 demos.

How to operationalize it? Track a few falsifiable hard signals, not press events:

• Intervention rate. How long does the robot work before a human has to take over? That number falling is the evidence that L2 is closing on L3.
• Autonomous hours. Continuous human-free runtime in real settings.
• Is the timeline sliding or delivering? If every year the “mass production” target moves out another year, the L3 wall is thicker than assumed — and it isn’t blank because there are no pilots; it’s blank because there are no public, credible, continuously-trackable L3 metrics.

Until L3 turns green, any “the body is about to ramp” narrative should be priced as an option — not as delivered growth. This isn’t bearish on humanoids; it’s refusing to pay a price that’s already happened for a gate that hasn’t.

Demos are great. Demos aren’t money. Until the L3 light turns green, don’t mistake faith for delivery.

What’s the most convincing robot demo you’ve seen? Turned out to be autonomy — or teleop and editing?

(Independent industry and educational research, not investment advice. Companies named are illustrative examples, not recommendations. Data is from public sources and may change. The author and affiliates may hold positions in securities mentioned.)

— Adapted from Embodied Intelligence Investing, Deep Dive IV: Scaling-Law Monitor

#EmbodiedAI #Humanoids #VLA