Here is the data: Over the past 18 months, total disclosed funding for 'physical AI' and 'world models' reached $13.36 billion. That includes robotics, 4D generative models, and simulation platforms. At the same time, early-stage foundation model rounds are slowing โ the 'stack parameters and pray' model is closing its window.
This is not a prediction. It is a capital flow statement. And if you are trading crypto infrastructure, you need to read it.
Context: The Three-Layer Casino
The AI investment landscape now splits into three distinct layers: - Foundation Models: Closed rounds concentrated in a few oligarchs (OpenAI, Anthropic). No new entry. - Infrastructure: Chips, cloud, data centers โ steady demand, but capex-heavy. - Applications: AIGC tools โ revenue exists, but zero moats. No winner yet.
The new money is avoiding the crowded middle. It is going straight to the physical layer: robots, world models, embodied intelligence.
Why does a crypto trader care? Because physical AI changes the compute profile entirely. LLMs are matrix-multiplication heavy. World models demand real-time 3D rendering, physics simulation, and sensor fusion โ a workload that current GPU supply is not designed for. The 2023โ2024 GPU shortage was driven by training. The next wave will be driven by inference of 4D space-time models. That is a different bottleneck.
Core: The Decentralized Compute Angle
I spent three months in early 2025 stress-testing an AI-agent trading platform. The agent failed on a regulatory news event โ 10% drawdown in 90 minutes. Its reasoning engine lacked any understanding of non-market signals. That is a toy problem compared to what world models face: causal reasoning over physical events.
Let's be precise. A world model must simulate Newtonian physics, object permanence, and multi-agent interactions in real time. The computational cost per query is orders of magnitude above a GPT-4 inference. This creates a structural demand for compute that centralised cloud providers cannot meet at scale without massive latency trade-offs.
That is where decentralised compute networks (Render Network, io.net, Akash) enter the picture. Their value proposition โ distributed GPU availability with low latency for batch jobs โ aligns with the training phase of world models. But the inference phase requires sub-100ms response times. Physics simulation is not a batch job when a robot arm is moving.
The key insight: physical AI will bifurcate the compute market into two segments: high-latency training (where decentralised solutions can compete) and ultra-low-latency inference (where edge hardware wins). Crypto infrastructure plays only the first segment. Anyone pitching 'decentralised inference for humanoid robots' is selling PowerPoint, not product.
Contrarian: The Consensus Trap
The 2023 EigenLayer restaking fiasco taught me that when everyone agrees on a narrative, the risk is the narrative itself. Today, 'physical AI' is the maximum consensus. Every venture fund has a dedicated partner for 'embodied intelligence.' The problem: the technology is pre-POC. There is no benchmark for world model performance. No reproducible evaluation suite. No clear path from simulation to reality.
I have seen this movie before: the 2022 Terra collapse. Everyone agreed that 'novel yield mechanisms' were the future. The data showed inflows โ just like today. The flaw was ignoring the failure mode. Physical AI's failure mode is a decade-long Sim-to-Real gap. If the first generation of world models cannot transfer to physical hardware, the $13.36 billion will be written off as R&D spend. Not revenue.
A battle-tested trader knows that the best time to sell a narrative is when it becomes consensus. The fact that there is 'no pure public play' (as the Serenity report notes) means all the enthusiasm is priced into private rounds. Retail investors cannot access that alpha directly. They buy token proxies โ compute tokens, sensor oracle tokens โ that are driven by speculation, not by actual demand from robots.
Takeaway: Position the Picks and Shovels
The asymmetric trade is not on physical AI tokens. It is on the infrastructure that serves both physical AI and conventional AI. Decentralised compute networks, simulation platforms (Unity, Omniverse integration), and data oracle chains that handle 3D verification. These assets have a floor from existing demand (gaming, simulation, AI training) and upside if physical AI lands.
If the technology works, the infrastructure layer prints. If it fails, you are still holding assets with proven utility. That is the trade structure I am building into Q3 2026.
If you cannot explain the slashing conditions of a restaking protocol, you are the exit liquidity. Apply the same principle to physical AI: if you cannot explain the Sim-to-Real gap, you are the funding round.