Hook The data suggests a quiet inflection point: China's AI models now process 98 trillion tokens per month. America's models handle 53 trillion. That is an 85% lead, and the gap is widening at 113% month-over-month growth versus 43%. These numbers come from Apollo Global Management, not a Chinese propaganda outlet. For anyone who has watched DeFi protocols fake their TVL with token incentives, this smells familiar. But the signal here is real — and it carries a warning for how we value centralized infrastructure.
Context In crypto, we obsess over on-chain activity as a proxy for network health. Token volume, active addresses, transaction counts — these are the raw metrics that precede revenue and valuation. AI models now operate under a similar paradigm. Every token processed requires compute, storage, and inference latency. The top 50 most-used AI models now include 20 Chinese entries, up from just 5 a year ago. American models dropped from 33 to 28. This is not a slow drift; it is a structural rebalancing of global AI compute demand.
The numbers were published by The Kobeissi Letter, citing Apollo's analysis. The same week, Anthropic accused Alibaba of running a massive model distillation campaign — effectively stealing Claude's reasoning patterns. Alibaba responded by banning all employees from using Claude Code, citing "backdoor risks" and forcing them onto its own Qoder platform. Simultaneously, Chinese regulators removed over 14,000 unregistered AI products from the market. These events are not coincidental. They are the surface evidence of a deeper war over control of the AI value chain.
Core: Quantitative Reality Check Let me break down what 98 trillion tokens per month actually means for infrastructure. Assuming an average inference cost of 1.5 FLOPs per token (a conservative estimate for models like DeepSeek-V4 or Qwen4), China's monthly inference compute demand reaches approximately 147 petaFLOPs. That requires a sustained cluster of thousands of H100-class GPUs. America's 53 trillion tokens demand roughly half that — but American models are often larger, with higher per-token compute requirements, so the gap in raw GPU hours may be narrower.
More importantly, the growth rate differential is impossible to ignore. China's 113% month-over-month token growth suggests a market still in hyper-adoption phase. America's 43% growth is strong, but it signals a maturing user base. If current trends hold, by Q4 2026 China's token volume could exceed 300 trillion per month — more than five times America's projected volume. This parallels what we saw in DeFi during 2020-2021: a smaller number of protocols (models) capturing disproportionately large shares of activity, fueled by aggressive pricing and network effects.
But here is the catch: token volume does not equal revenue. In crypto, we learned that TVL can be rented. In AI, token volume can be subsidized. DeepSeek and Qwen have been offering inference at prices 60-80% below OpenAI's GPT-4o. This is classic market-share warfare. The question is whether these Chinese firms can monetize without bleeding cash. If their unit economics are negative at scale, then the 98 trillion tokens become a liability, not an asset.
Contrarian: The Blind Spot in the Token Narrative Conventional wisdom says that more usage equals more value. Logic is binary; intent is often ambiguous. The AI industry is repeating a pattern we saw with Ethereum's gas fees during the NFT mania: high activity, low utility per transaction. Chinese models may be processing massive volumes, but what fraction is high-quality reasoning versus cheap text generation or spam? The 14,000 products removed by Chinese regulators suggest that a significant portion of the demand came from low-value, possibly non-compliant applications. When those are eliminated, the token volume could drop sharply.
Furthermore, the Alibaba-Claude incident exposes a deeper trust asymmetry. Anthropic claims systematic distillation. Alibaba claims security backdoors. Both are plausible. But the real risk is that the two ecosystems are becoming mutually hostile. American companies may refuse to deploy on Chinese cloud infrastructure. Chinese companies will shift entirely to domestic AI stacks. The global developer community could fragment into two parallel markets, each with its own models, pricing, and compliance rules. This is the nightmare scenario for anyone building cross-border AI applications — and it mirrors exactly the blockchain trilemma of security, decentralization, and scalability, except here the trade-off is between capability, cost, and sovereignty.
Takeaway Token volume is a leading indicator, not a lagging one. The 98 trillion figure tells me that China's AI infrastructure is scaled and hungry. But it does not tell me if that scale is profitable or sustainable. For crypto, this is a canary: the same forces driving AI centralization — compute monopolies, regulatory fragmentation, and geopolitical distrust — are the forces pushing decentralized compute networks like Bittensor and Akash into relevance. If the AI world cannot trust itself, it might have to trust code. Watch the next vapor wave: not of tokens, but of tokenized compute.