Six billion dollars of venture capital poured into companies promising to build "World Models" in the first quarter of 2026 alone. There's no sudden breakthrough in mechanical reasoning behind it. The term is doing a job physics can't yet do itself: physics is a remarkably difficult thing to scale. The industry is rife with mislabeled models. Companies slap the world model label on what are essentially Vision-Language-Action models, or VLAs. A VLA is a reactive loop that maps visual perception and language directly to motor commands. Machines like Physical Intelligence's pi-0.5 or Figure AI's Helix-02 (which splits fast reflexes from slower reasoning across separate systems) are fundamentally seeing-and-doing machines. They react to pixels in real time.

Fei-Fei Li, the industry's godmother, said in a recent interview that her team got fed up with how the term was being used. Her team also published a taxonomy post to draw the lines. Her functional taxonomy defines three tiers: the Renderer, which generates beautiful pixels; the Planner, which calculates a reactive next move; and the Simulator, the actual World Model that respects the underlying physics and dynamics of reality. Most current humanoids are merely Planners being sold as Simulators, because the latter commands a billion-dollar premium.

A few outliers are treating robotics as a physics-pretraining challenge rather than a simple fine-tuning problem. Tesla runs both its cars and Optimus through the same neural world simulator to predict future states, while 1X Technologies has established a dedicated World Model Lab to build its 1XWM model. In January 2026, 1X published its architecture: the model generates a video of what the task should look like first, then extracts the robot's movement sequence from that footage. 1X calls its $20,000 NEO a home robot. 1XWM is the research layer underneath, teaching NEO what the physical world is supposed to look like before it moves. Companies like 1X and Tesla argue that general-purpose autonomy requires an internal model of how the physical world actually behaves. When Ashok Elluswamy, head of Tesla AI, was asked about the payoff of building a real world model, he said you have to spend a lot of money to do it. That's the whole answer.

For the investors signing the checks, these technical nuances are secondary to the credibility stamp the label provides. Investors don't care which architecture it's running as long as the branding suggests a leap forward in embodied AI (systems that operate in and interact with the physical world). The terminology is doing work the technology hasn't earned yet. Physics doesn't take venture debt as a down payment, and soon a six-billion-dollar bill will come due.

Figure AI has claimed to produce one Figure 03 per hour at their BotQ facility and has set a price target of $20,000 per unit. A significant gap exists between Figure AI’s consumer marketing and the actual cost of a Western factory pilot unit. Bank of America reported the cost closer to $90,000 to $100,000 for a test machine used in a trial phase. This pricing is likely driven by the necessary inclusion of high-torque actuators, which are powerful motors that move the robot’s joints, and industrial safety ratings, which are the expensive certifications required for machines to work alongside humans without cages.

The $25-per-hour operating cost cited by one industry source for the BMW Spartanburg deployment ignores the physical reality of hardware decline. Although it is presented as a notable milestone, the machine is not yet a sustainable financial asset. The math doesn't check out on the difference between optimistic consumer pricing and the actual cost of industrial-grade machinery.

The depreciation (loss in value of an asset over time) exposes economic instability at the BMW Spartanburg plant. During the initial ten-month trial, the machines recorded 1,250 operating hours, which annualizes to approximately 1,364 hours. When a standard three-year amortization (process of spreading the machine’s high purchase price over its expected working life) is applied, the numbers become clear. A $90,000 machine used for 1,364 hours a year results in a depreciation cost of $22 to $24 for every hour it is powered on. If the robot costs $25 an hour to run but loses $24 an hour in value just by existing, there is only $1 remaining for electricity, replacement parts, and the expensive software keeping the unit upright.

To me, a machine paying for its own decline and not its output isn't an asset, it’s a very expensive clock running down. If these humanoids are barely paying for their own existence, how much is the industry losing on every hour it can't account for?

Two pieces this week, one thread. The Signal looks at how "world model" became a credibility stamp rather than a technical specification. Below the Fold runs the depreciation math on Figure 03's real operating cost. The gap between what this industry claims and what the numbers show is where the risk is building.