The chart lied.
For months, the narrative was clear: LLMs are the only game in town. Compute equals power. Scale equals dominance. Then the real alpha emerged from an unexpected corner. Serenity's latest capital flow report reveals a seismic shift โ $87.9 billion of Chinese VC money has quietly rotated out of pure foundation models and into Physical AI and World Models. The number itself is a signal. It's not a pivot; it's a full-blown sector rotation.
Let me be blunt: the era of 'throw compute at a transformer' is over. Capital is now chasing assets that touch the real world. This is the DeFi summer of AI โ except the liquidity is flowing into hardware.
Context: Why Now, Why Physical AI?
Physical AI isn't a buzzword. It's the logical next step after the LLM scaling law hit diminishing returns. LLMs like GPT-4 are masterful pattern matchers, but they lack any grasp of physical causality. Ask one what happens when you drop a glass, and it can generate text โ but it doesn't know. Physical AI and World Models aim to fix that: they build representations of 3D space, force feedback, and real-time causality.
Serenity's report โ posted on X on July 4, 2024 โ breaks down where the money went: 235.6 billion yuan into foundational models, but a surprising 87.9 billion yuan into Physical AI and World Models. The key word is surprising. Most analysts expected the gap to be wider. It's not.
Why the shift? Three reasons:
- Chip sanctions forced a rethink โ China can't compete on H100 clusters. Physical AI's end-side inference can use less advanced chips (Horizon, Huawei Ascend).
- Differentiation โ Every second-tier LLM looks the same. Physical AI offers a moat: proprietary data from real-world interactions.
- Industrial policy โ The Chinese government wants manufacturing automation. VCs follow the money.
Based on my years auditing smart contracts during the 2017 ICO boom, I've seen this pattern before. A narrative emerges. Capital rushes in. But execution gaps remain wide. The question is whether Physical AI's execution gap is wider than its promise.
Core: The Data Behind the Rotation
Let's dig into the technical and commercial realities.
Technical Reality
Physical AI requires training on multimodal data: video, depth, force, torque. This data is orders of magnitude more expensive to acquire than text. You can't scrape the internet for robot arm trajectories โ you have to hire human operators to teleoperate a robot for hours. I recall my experience tracing the FTX collapse's blockchain footprints: the most valuable data was the hardest to find. Same here.
World Models โ like Nvidia's Omniverse or its Chinese counterparts โ require massive computation for physics simulation. The bottleneck shifts from matrix multiplication to real-time ray tracing and rigid-body dynamics. This changes the hardware needs entirely. China's advantage in manufacturing could offset this, but the software stack (simulation engines) remains dominated by US firms.
Commercial Reality
LLMs have a clear monetization model: API calls per token. Physical AI's models are murkier. They sell hardware (low margin), robots-as-a-service (high CapEx), or system integration (non-scalable). The revenue cycle is long. The burn rate is high.
But here's the contrarian truth within the core: Liquidity is the only religion in the DeFi temple. In crypto, capital flows to where yield is highest. In AI, capital flows to where the next paradigm shift will happen. Physical AI is that shift โ even if the yield is 18-24 months away.
China's Unique Play
Chinese VCs are not betting on software supremacy. They're betting on supply chain integration. The logic: if you can build the hardware cheaper and faster, you can iterate on the AI faster. Tesla's Optimus robot is proof of concept โ now China wants to mass-produce it. This is a classic Chinese industrial strategy: win on cost, then win on quality.
I see parallels with the DeFi yield farming wars of 2020. The first movers who controlled the liquidity pools (here, the hardware factories) dictated the terms. The trend is your friend until it ends abruptly.
Contrarian: The Blind Spots No One's Talking About
The euphoria around Physical AI masks three critical risks.
1. Safety standards are a black hole.
LLM 'hallucinations' produce wrong text. Physical AI 'hallucinations' produce a robot arm breaking a human rib. There are no established safety frameworks for embodied agents in China. No red teaming protocols for walking robots. If one major accident happens โ and it will โ regulators could slam the brakes, freezing billions in capital.
2. The valuation bubble is inflating fast.
87.9 billion yuan is a lot of money chasing very few companies. Most of these startups have zero revenue. Their valuations are based on 'potential' and 'scarcity' โ the same ingredients that fueled the ICO mania I audited back in 2017. I learned then that when everyone says 'this time is different,' it usually isn't. Data lies, but volume never cheats. The volume of capital is real, but the underlying value isn't yet.
3. The 'world model' dependency.
China lacks a native high-fidelity physics engine like Nvidia Omniverse. If the US restricts access to simulation tools, Chinese Physical AI development slows dramatically. The hardware advantage means nothing if the software can't simulate reality accurately.
Takeaway: What to Watch Next
Alpha moves before the charts confirm the truth. The charts are still showing LLM dominance. But the capital flow charts have already rotated. The next 18 months will separate the visionaries from the vaporware.
I'll be watching three signals:
- Which Chinese startup can demonstrate a reliable repetitive physical task (like picking and placing 1000 times without failure) in a real factory environment.
- The first regulatory framework draft for embodied AI safety in China โ its tone will define the sector's future.
- Nvidia's response โ if Omniverse becomes restricted, the scramble for a domestic alternative will accelerate.
Patience is a luxury; action is a necessity. Physical AI is where the action is โ but it's also where the hidden landmines are buried. Tread carefully, but don't stand still.
