The ledger remembers what the hype forgot: Chinese venture capital just fired a warning shot across the bows of the pure-play LLM thesis. Serenity, a prominent VC, reports that ¥8.7B has rotated from foundational model startups to physical AI and world models since Q1 2024. The market's initial reaction is euphoria—a new narrative, a new frontier. But as a forensic analyst who tracked the Terra/Luna collapse line by line, I see the same pattern: capital fleeing a dying narrative, rushing into a new one before the technical reality catches up.
Context: Why Now? This isn’t organic evolution; it’s a desperate pivot driven by three tectonic forces. First, the 'scaling law' of LLMs is showing diminishing returns in China. With H100 export bans choking compute, Chinese labs can't match GPT-4 class models. Second, the hype cycle for pure AIGC applications (AI video, chat) has peaked—too many me-too products, zero moats. Third, the state's new industrial policy pushes 'hard tech' over software. The result: VCs are now pretending that physical robots are the next big thing in AI, conveniently ignoring that world models require 100x the data, 10x the compute, and a god-level hardware supply chain.
Core: The Technical Chasm Alpha is silent until the chart screams. Let’s read the chart of physical AI. The core difference between a language model and a world model is that the former learns tokens; the latter must learn physics—gravity, friction, causality. Current state-of-the-art in physical AI (e.g., NVIDIA's Isaac Gym) still relies on synthetic data because real-world interaction data is prohibitively expensive. A single hour of human teleoperation for a robot can cost thousands of dollars. Meanwhile, the inference latency required for real-time robotics is in single-digit milliseconds—a feat that even the most optimized LLM deployment struggles to achieve.
Chinese startups entering this space are betting on two unproven assumptions: that they can produce cheap hardware (China's manufacturing edge) and that they can compensate for software deficits with data from domestic factories. But here's the kicker: the foundational simulation engines—like NVIDIA's Omniverse or Google's MuJoCo—are all American. Chinese firms are essentially building on sand, then pretending it’s bedrock. They’re wrapping proprietary data around foreign open-source cores, a model that works until the licenses change or the export controls tighten.
Furthermore, the capital rotation is reminiscent of the 2021 NFT mania I dissected. Back then, I traced anomalous wallet clusters to metadata manipulation. Today, I see VCs funding 'world model' startups that can barely produce a demo without heavy manual intervention. The few that have shown progress (like Figure AI’s humanoid) are burning $2M+ per month on R&D with zero revenue. Investors are buying lottery tickets, not businesses.
Contrarian: The Unreported Blindspot Everyone is celebrating the 'diversification' of AI funding. No one is asking: what happens when these physical AI companies fail to deliver? The structural risk is existential—not just for the startups, but for the entire narrative of 'AI industrialization.' In crypto, we saw L2s fragment liquidity; in AI, physical AI is fragmenting capital into dozens of unsolvable technical problems. The biggest blindspot is safety. A hallucination in a text model gives you a nonsense answer. A hallucination in a physical AI robot can break bones. There is no regulatory framework for that in China, and the US is just starting to draft guidelines.
Another contrarian angle: this pivot is admission that China has lost the decadelong race for AGI. By focusing on physical AI, they are ceding the high ground of fundamental reasoning to OpenAI and Anthropic. They are betting on a niche that may not exist as a standalone market. If US labs solve physical AI as a side effect of AGI (which they are, via multimodal models), Chinese hardware-centric approaches will be obsolete before they ship.
Finally, let's talk about the liquidity illusion. Serenity’s report shows ¥13.36B in physical AI deals since 2023. That sounds massive until you realize that just one US AI company (OpenAI) raised more than that in a single round. The 87.9B in LLM funding over the same period dwarfs physical AI by 6.6x. The 'acceleration' is from a tiny base. This is not a flood; it's a trickle that feels like a wave because the alternative (LLM funding) has dried up.
Takeaway: The Next Clock to Watch We build on sand, then pretend it’s bedrock. The question every crypto native should ask is not 'will physical AI work?' but 'which physical AI startups will implode first, and will that contagion spill into adjacent tokenized compute and hardware markets?' The future is a bug report waiting to happen. In the next 12 months, look for the first high-profile physical AI bankruptcy or a major safety incident involving a Chinese robot. When that happens, the capital that rushed in will rush out even faster. And those still holding the bag will learn what I learned in 2022: speed kills, but in crypto, stillness is death. Stay liquid.