China’s artificial intelligence industry has fully embraced the “token economy,” treating the fundamental unit of LLM computation as a utility akin to electricity. Driven by the viral popularity of digital assistants and agentic platforms like OpenClaw, daily token consumption in the country skyrocketed from 100 trillion at the end of 2025 to 140 trillion by March 2026. However, a new analysis by Reuters Breakingviews columnist Robyn Mak, published on April 16, 2026, suggests that this volume-driven strategy may have a fundamental ceiling.
The core argument is simple but profound: tokens are not fungible. While Chinese tech giants have successfully driven down inference costs to a fraction of US levels, the quality of the underlying models still dictates their utility for high-value enterprise applications. A million tokens processed by a mid-tier model cannot simply be substituted for a million tokens processed by a frontier system like Anthropic’s Mythos or OpenAI’s latest iterations.
The Economics of the Token Rush
The appeal of the token model for Chinese companies is clear. As Mak notes, charging for tokens rather than selling software subscriptions is a much simpler business model that aligns perfectly with China’s traditional industrial playbook: focus on efficiency, scale, and volume.
This strategy has been remarkably successful in driving adoption. According to Jefferies analysts cited in the Reuters piece, Chinese models are, on average, one-sixth the price per token of US offerings. Startups like MiniMax have aggressively pushed this narrative, describing their February release—which costs just $1 to run continuously for an hour at 100 tokens per second—as delivering on the promise of “intelligence too cheap to meter.”
This price war has been fueled by China’s structural advantages in inference, including cheap electricity for data centers and significant algorithmic breakthroughs (such as the widespread adoption of Mixture-of-Experts architectures) that compensate for the lack of cutting-edge US chips.
The shift is so pronounced that Alibaba recently separated its AI business from its cloud computing arm, rebranding it as the Token Hub Business Group. This unit recently launched Meoo, a free no-code app builder that integrates multiple models, further demonstrating the company’s commitment to driving token volume through accessible applications.
The Quality Ceiling and the Fungibility Problem
Despite these impressive metrics, Mak argues that the “too-cheap-to-meter” approach is only applicable to a certain extent. For enterprises looking to outsource complex, high-value work to digital agents, model quality will matter just as much, if not more, than cost efficiency.
While Chinese models from Alibaba, DeepSeek, Moonshot, and Zhipu have narrowed the performance gap in specific benchmarks like coding and math, they still trail in overall frontier capabilities—a broader measure of aggregate performance, versatility, and reliability.
Mak points to Anthropic’s latest Mythos system as a prime example. The company claims the model is so powerful that it will initially only be available to vetted firms, including JPMorgan, Amazon, and Microsoft. The implication is clear: the intelligence generated by a million tokens of Mythos is fundamentally more valuable than the intelligence generated by a million tokens of a cheaper, less capable model.
This lack of fungibility poses a significant challenge to the Chinese strategy. If the most lucrative enterprise applications require frontier-level capabilities, competing solely on price and volume will leave Chinese AI providers in lower-margin segments of the market.
The Compute Bottleneck and Investor Skepticism
The path to closing this capability gap is fraught with obstacles, primarily due to US export controls on advanced semiconductors. While China subsidizes chipmaking at 3.6x the US rate to build domestic alternatives, the reality remains that domestic chipmakers like Huawei are struggling to match the performance of Nvidia’s older H200 processors.
These older chips are already two generations behind Nvidia’s highly anticipated Vera Rubin architecture, which promises a 10x improvement in performance per watt over its Grace Blackwell predecessor. Because of US export rules, these next-generation systems will likely be unavailable to Chinese labs.
This compute bottleneck forces Chinese companies to make difficult trade-offs. Justin Lin, who formerly led Alibaba’s open-source models division, is quoted in the Reuters piece lamenting this reality: “A massive amount of OpenAI’s compute is dedicated to next-generation research, whereas we are stretched thin — just meeting delivery demands consumes most of our resources.”
This dynamic appears to be reflected in investor sentiment. Mak highlights a stark contrast in valuations: Alibaba’s market capitalization hovers around $315 billion, significantly below the private valuations of OpenAI ($852 billion) and Anthropic ($380 billion). Given that JPMorgan analysts project Alibaba’s e-commerce platform alone will generate $29 billion in earnings by 2027, applying a conservative 10x multiple implies that investors ascribe almost zero value to Alibaba’s cloud and AI businesses—despite their target of $100 billion in combined annual revenue within five years.
While China’s low-cost token strategy will undoubtedly accelerate the adoption of AI across its domestic economy, the Reuters analysis serves as a sobering reminder that volume cannot entirely substitute for capability. As the global AI race increasingly focuses on frontier intelligence, the ultimate success of China’s token obsession will depend on its ability to break through the quality ceiling.