87.9 billion yuan. That is the volume of Chinese venture capital flowing into Physical AI and World Models in the first half of 2024, according to Serenity. In the same period, funding for pure large language models dropped 30%. This is not a simple sector rotation—it marks a structural shift in how capital views artificial intelligence. For anyone tracking decentralized compute markets, this signal carries direct implications for GPU demand, latency requirements, and the viability of tokenized AI networks.
The data comes from Serenity, a Chinese venture firm that tracks capital flows across AI and crypto. Their report confirms what many infrastructure analysts suspected: the era of 'throw GPUs at a text model and call it AI' is ending in China. The new frontier is hardware-integrated intelligence—robots that understand physics, world simulators that run real-time, and systems that operate in the physical world rather than just generating text. This shift from pure software to software-hardware hybrids changes the compute profile entirely.
Let me be clear: this is not about whether LLMs are dead. They are not. But the marginal investment in scaling laws is hitting diminishing returns. Chinese VCs, constrained by export controls on high-end GPUs, are searching for a different angle. Physical AI offers that angle because it relies less on massive matrix multiplication in data centers and more on edge inference, low-latency simulation, and tight integration with manufacturing supply chains. The result is a capital reallocation that will reshape demand for compute resources over the next 18 months.
In my audit of decentralized GPU networks over the past six months, I observed a pattern: most protocols—Render Network, Akash, io.net, Golem—are optimized for batch training jobs. Their architectures favor high throughput, asynchronous processing, and tolerance for variable latency. That works for fine-tuning a language model. It does not work for a robot controller that needs sub-10-millisecond response times to avoid crashing into a human worker. Physical AI demands a fundamentally different compute profile: deterministic, real-time, and geographically co-located with the hardware.
The numbers confirm the trend. Serenity reports that Chinese VC funds allocated 133.6 billion yuan to Physical AI and World Models in H1 2024, compared to 235.6 billion yuan for pure LLMs. But the LLM figure is declining month-over-month, while the Physical AI figure is accelerating. Meanwhile, in the United States, capital continues to concentrate into OpenAI and Anthropic—two players that are building general intelligence primarily through language. This divergence creates a bottleneck for decentralized compute networks: they were designed for the US-style training-heavy paradigm, not the Chinese style of real-time physical simulation.
Let me break down the technical implications. First, latency. Decentralized networks rely on a global pool of nodes with varying hardware and network speeds. For a physical AI system—say, a warehouse robot using a world model to plan its pick-and-place operations—a 200-millisecond delay means a collision. Current decentralized platforms cannot guarantee that latency. Second, data verification. Physical AI requires high-fidelity simulation environments to train world models. These simulations generate terabytes of data per hour. To integrate blockchain for data provenance, as some projects propose, the throughput of the verification layer must match the data generation rate. Most current blockchains cannot handle that volume without congestion.
Here is the contrarian angle: this shift could actually hurt decentralized compute in the short term rather than help it. The narrative that 'AI will drive demand for decentralized GPU networks' was built around training LLMs. If the next wave of AI is hardware-centric and proprietary—robots built by Tesla, Xiaomi, or specialized startups—they will use centralized cloud services because those offer guaranteed performance. Decentralized networks thrive on flexibility, not determinism. Physical AI needs deterministic compute. The two are in tension.
Furthermore, Chinese VCs are investing in closed hardware stacks. They fund companies that own the robot, the chip, the simulation engine, and the data pipeline end-to-end. This is the opposite of the open, permissionless ethos that underpins most crypto infrastructure. The 'infrastructure' they are building is literal factories and assembly lines, not decentralized compute grids. For blockchain-based AI projects, this means they need to pivot from serving training workloads to serving verification workloads: proving that a simulated world model is accurate or that sensor data has not been tampered with. That is a much smaller market in the near term.
From my experience analyzing capital flows in 2021 NFT metadata security and 2022 FTX collapse intelligence, I learned that the biggest moves often happen in silence. This Chinese VC pivot is one of those silent moves. It will take 6–12 months before the ripple effects hit decentralized compute utilization rates. But when they do, the protocols that survive will be those that can offer low-latency, deterministic execution on top of decentralized infrastructure—something that currently does not exist at scale.
The takeaway is forward-looking. Watch for decentralized simulation platforms—projects like Render Network's upcoming real-time streaming or new entrants focused on robotics simulation. Also monitor data markets specifically tailored for physical sensor data, such as Ocean Protocol or IOTA, which could become the provenance layer for world model training sets. The next crypto narrative around AI will not be 'decentralized training' but 'decentralized verification and simulation.' The capital is moving. The infrastructure has yet to catch up. The bottleneck is real.