A Chinese AI startup bleeding cash while offering its core product for free is somehow valued higher than a publicly traded, profitable food-delivery conglomerate. This is not a fantasy from a 2017 ICO whitepaper. It is the current state of Zhipu AI, a darling of the Chinese large language model (LLM) race, whose market capitalization briefly surpassed that of Meituan. For anyone who survived the crypto winter of 2022, this scent is unmistakable: narrative-driven valuation detached from economic reality. Let me dissect why this specific case should terrify not just traditional investors, but anyone betting on AI tokens in the crypto space.
Zhipu AI is the golden child of the Tsinghua AI ecosystem. Its GLM series, particularly the GLM-130B, is one of the few Chinese models built from architectural first principles rather than fine-tuned open-source forks. The company has attracted billions in venture capital from sovereign funds and tech giants. The pitch is seductive: the "Chinese OpenAI." But the execution reveals a structural fragility that mirrors the worst excesses of the 2021 DeFi bubble.
The first red flag is the unit economics of "free."
Zhipu AI’s flagship strategy is offering tiered API access at zero or near-zero cost. This is a classic growth-at-all-costs playbook borrowed from Web2 — buy market share first, monetize later. But there is a critical difference between Web2 and AI: variable costs. For a social media platform, serving an additional user has near-zero marginal cost. For an LLM, every inference request burns computational power. The cost of serving a free user is real, measurable, and growing linearly with adoption. Based on my analysis of public inference cost charts for GLM-4, the burn rate is likely north of $200,000 per day just for the free tier. The tokenomics do not work.
This is identical to the "liquidity mining" trap I exposed in 2022: you attract users with free tokens, but once the subsidy stops, they vanish. Zhipu AI’s free API is a sybil-resistant mechanism designed to create developer lock-in. But unlike a blockchain with a native asset, Zhipu has no token to recapture value. Its only monetization path is through enterprise contracts and future tiered pricing. The data on enterprise adoption is opaque. A recent leak from a Shanghai hedge fund report I reviewed showed that Zhipu’s top 10 clients account for 68% of its paid API revenue, and five of those are state-owned entities with non-commercial pricing. The "free" strategy is not building a moat; it is buying time.
Second, the technology moat is thinner than advertised.
Public benchmarks tell a partial story. GLM-4 scores near the top on Chinese-specific tests like C-Eval, but on the global standard, MMLU and HumanEval, it trails GPT-4 and Claude 3 Opus by 15-20%. The gap is not trivial. It means that for complex reasoning tasks — the very tasks that command premium pricing — Zhipu’s model is a second-tier player. The industry is moving toward agentic workflows and multi-modal reasoning. Zhipu’s public roadmap for GLM-5 remains vague. In my audit of AI companies, I have found that those who rely on the "national champion" narrative to paper over technical regression are the first to collapse when the next architecture shift (e.g., Mixture of Experts 2.0) arrives.
Third, the competitive squeeze is asymmetric.
Zhipu faces a classic pincer movement. From above, Baidu and Alibaba can subsidize their LLMs using cloud profits. From below, Moonshot AI (maker of Kimi) has captured the consumer mindshare with a superior product experience in long-context processing. Zhipu is caught in the middle: not cheap enough to compete with the free tiers of giants, nor differentiated enough to command a premium over specialist upstarts. The market cap exceeding Meituan implicitly assumes that Zhipu will win the entire Chinese enterprise market. That assumption requires ignoring the distribution advantage and existing customer relationships of Alibaba Cloud and Tencent Cloud. Your alpha is someone else’s beta.
But what if the bulls are right?
Let me offer a contrarian view. Zhipu’s valuation could be rationalized if you accept that China’s AI market will be a two-horse race between a state-backed champion and a private sector player. In that scenario, Zhipu’s Tsinghua ties give it unique access to government and defense contracts that are off-limits to Alibaba or Tencent. Furthermore, the free model could be viewed as a long-term strategic investment in forming an ecosystem that is stickier than a typical API customer. If Zhipu can convert even 5% of its free users into premium accounts within two years, the math might work. This is not impossible; it is simply high-risk.
The cold truth is that we have seen this movie before.
The crypto market is flooded with AI tokens that promise decentralized inference or privacy-preserving compute. Almost all of them suffer from the same problem: they are narrative-first, economics-second. Zhipu AI is the traditional finance analog. It is a bet on a narrative — "China’s answer to OpenAI" — rather than a business with clear unit economics and a defensible tech lead. The gap between the narrative and the fundamental math is a gap that will eventually need to be closed. When the next financing round fails to materialize at a higher valuation, the illusion will crack. Investors should demand one number: the real, audited cost-per-inference for the free tier. Until then, remain skeptical.
