The chain says scarcity, but the flow says abundance. China’s top AI models processed 98 trillion tokens in May 2026—nearly double the 53 trillion handled by their U.S. counterparts. Apollo Global Management and The Kobeissi Letter released these numbers, and the crypto-native read them as a liquidity event. But I’ve spent enough years watching AMM pools and order books to know: high volume can mask structural fragility. Let me trace the ghost in the liquidity protocol.
Context: The Macro Liquidity Map of AI Inference
The data is clear: the top-50 most-used models now include 20 Chinese entities, up from just five a year ago. U.S. models dropped from 33 to 28. Month-over-month token growth: China at +113%, U.S. at +43%. On the surface, this is a tectonic shift—China has overtaken the U.S. in raw inference volume. But as a digital asset fund manager who survived the 2022 derivatives crash, I know that volume without margin is a debt bomb.
Alibaba banned employees from using Claude Code, citing ‘backdoor risks,’ and redirected everyone to Qoder. Meanwhile, Anthropic accuses Alibaba of mass distillation—essentially stealing model outputs to train competing architectures. These are not just geopolitical gestures; they are liquidity valves being shut off. The U.S. is trying to restrict GPU exports, China is consolidating its app ecosystem (removing 14,000+ unregulated AI products), and both sides are building parallel compute grids. Code is law, but narrative is leverage—and the narrative here is that Chinese AI has won the usage war. But has it won the quality war?
Core: Token Volume as a False Signal
Let’s treat tokens like on-chain transaction volumes. In DeFi Summer 2020, Uniswap’s daily volume exploded—but most of it was bots and wash trading. The real signal was in fee generation and unique active users. Similarly, China’s 98 trillion token count may be inflated by price wars. DeepSeek slashed API costs to near zero in 2025, and Qwen followed. Cheap inference attracts massive testing and free-tier consumption, not necessarily high-value commercial workloads. From my experience designing dynamic hedging strategies for impermanent loss, I know that when you drop the price, you attract the noise, not the signal.
Based on my audit of liquidity protocols, I’ve learned that a deeper book doesn’t mean a healthier market. The same applies here: each token from a Chinese model may carry lower economic value per unit than a GPT-5 or Claude 4 token. U.S. models handle more complex tasks—code review, long-document analysis, multi-step reasoning—that command higher margins. The average revenue per token for U.S. models could be 3–5x higher. So the 85% volume lead might translate into a revenue deficit.
Furthermore, the infrastructure cost of 98 trillion tokens is enormous. Assuming an average inference cost of ~2 FLOP per token, that’s over 150 petaFLOPs of sustained compute. China has built this capacity, but largely on restricted GPUs like H20 and domestic alternatives like Ascend 910B. The U.S. is blocking further exports—Anthropic has actively lobbied Washington to tighten controls. If the next generation of GPU restrictions hits, China’s token volume could stall or even contract. Volatility is the price of admission, and the chip supply chain is the most volatile variable in this equation.
Contrarian: The Decoupling Thesis Is Premature
Many analysts argue that the U.S. lead in AI is over. They point to the token volume gap and the shrinking list of U.S. models in the top 50. But I see a decoupling that might not materialize because the two ecosystems are still intertwined. American models are built on American chips; Chinese models are increasingly built on Chinese chips, but the software stack—PyTorch, CUDA alternatives, transformer libraries—remains globally shared. You can’t fully decouple without forking the entire infrastructure layer. The architecture of digital scarcity isn’t just about GPU die size; it’s about where the developer talent goes.
Consider this: China’s top 10 models (DeepSeek, Qwen, GLM, etc.) have closed the gap on benchmarks like MMLU and HumanEval, but they still lag on reasoning tasks like MATH and on instruction-following consistency. The U.S. top 10 (GPT-5, Claude 4, Gemini 2.5) maintain a statistically significant edge. The token volume gap is a reflection of China’s massive domestic internet user base—less a technology triumph and more a population dividend. When 1.4 billion people are online and hungry for low-cost AI, the token meters spin fast. But as any fund manager knows, retail-driven volume peaks first in a downturn.
Takeaway: Positioning for the Cycle
I’m not betting against Chinese AI. I’m betting that the market doesn’t price in the cost of these tokens. When the liquidity tap turns—via chip restrictions, a funding winter, or a shift to high-margin enterprise use—China’s volume will compress faster than America’s. The key signal to watch is not token count but API revenue per token and net dollar flows to Chinese AI companies. Until China’s top models achieve sustainable unit economics, the volume leadership is a mirage. Tracing the ghost in the liquidity protocol means understanding that in both crypto and AI, the real game is capital efficiency, not gross throughput. Volatility is the price of admission, and the cycle is about to teach us which models have the staying power to survive the next correction.