Tracing the liquidity ghosts through the ICO fog. Fifty thousand slots. One billion tokens each. Gone in hours. The first round of Zhipu AI’s free token giveaway for its GLM-5.3 model crashed under the weight of demand. When the second round opened, it was a repeat of the frenzy. The scene is eerily familiar: a digital asset distributed for free, a scramble to claim, a promise of future value. But this time, the token is not a blockchain asset. It is a compute credit—an AI model’s output measured in tokens. And the platform is ZCode, a walled garden for developers. The parallels to the 2017 ICO mania are not accidental; they are structural. We are witnessing the birth of a new liquidity cycle, one that bridges the gap between artificial intelligence and decentralized finance.
Zhipu AI, a Beijing-based AI leader backed by Alibaba, Tencent, and Sequoia, launched GLM-5.3, the latest iteration of its GLM series. The model follows the Transformer architecture, but the company has not released benchmarks or parameter counts. Instead, it chose to market the model through a developer acquisition stunt: 1 billion free tokens per new user on the ZCode platform, limited to 50,000 copies. The tokens are non-transferable, expire after an unspecified period, and can only be used within ZCode. The stated goal is to attract developers to build agents and applications on Zhipu’s ecosystem. The first round was paused due to “overwhelming demand,” a phrase that echoes the “sold out in minutes” of ICO token sales. The second round resumed with the same quota, suggesting the company deliberately limited supply to create urgency.
Let me be clear: this is not a blockchain project. But the playbook is identical. Free distribution of a scarce digital resource to bootstrap network effects. The hope is that developers will build on ZCode, become dependent on GLM-5.3, and eventually convert to paying API customers. Zhipu is spending an estimated $1.4 to $3.5 million on this campaign—a rounding error for a company that has raised over $3 billion in cumulative funding. The cost per developer acquired is roughly $28 to $70, far cheaper than traditional enterprise sales. For a bull market in AI, this is pocket change.
The Core Insight: Compute Tokens as Macro-Liquidity Instruments
From a macro-liquidity perspective, Zhipu’s free tokens are a direct injection of purchasing power into the AI developer economy. Each token represents a unit of computational output—a form of digital labor. By distributing 50 quadrillion tokens (50,000 users × 1 billion), Zhipu is effectively creating a temporary subsidy for the cost of AI inference. This is analogous to central bank quantitative easing, but targeted at a specific asset class: AI-generated content. The developers who claim these tokens will use them to build agents, process data, and generate code—activities that would otherwise be priced out by commercial API rates. The result is a burst of innovation, but also a distortion of true market signals. The real cost of GLM-5.3 inference is hidden by the subsidy. When the tokens expire, the developers will face a price shock, much like how crypto markets react when liquidity is withdrawn.
My own experience modeling the 2017 ICO bubble’s liquidity illusion taught me to watch for recycled capital. Back then, 60% of initial token sale liquidity was recycled within hours, creating a false sense of organic demand. Here, the recycling is different: developers will consume tokens, generate outputs, and potentially build products that attract real users. But the underlying model cost is still opaque. Zhipu has not disclosed the inference cost per token. If the cost is high, the free tokens are a temporary band-aid; if low, the conversion to paid API could be sticky. The data from this campaign—especially the token consumption patterns—will be more valuable than the campaign itself. Zhipu is collecting a treasure trove of user behavior: which prompts dominate, how long sessions last, how often agents fail. This is the data flywheel that can improve GLM-5.3’s alignment and efficiency.
The Contrarian Angle: Free Tokens Are a Trap
Conventional wisdom says free tokens accelerate adoption. I argue the opposite: they create a dependency that is fragile and potentially toxic. The tokens are locked inside ZCode, a proprietary platform. Developers cannot take their fine-tuned models or agents elsewhere. They are building on a platform that may not survive the next iteration of AI competition. The “omnichain app” narrative in crypto taught us that users don’t care about the underlying infrastructure; they care about the application. Similarly, developers will follow the best model, not the best token distribution. If GPT-5 or Claude 4 surpasses GLM-5.3, the free tokens become a sunk cost. This is a classic vendor lock-in strategy, disguised as generosity.
Furthermore, the lack of transparency around GLM-5.3’s performance is a red flag. In a market where benchmark scores are currency, Zhipu’s silence on metrics suggests the model is not ready for prime time. The free tokens may be a desperate attempt to gather feedback before competitors steal the narrative. The first round’s “demand” could have been artificial—bots, resellers, or speculators hoarding tokens. The company did not release registration numbers, only the quota. This is the same opacity that plagued ICOs: hype without substance.
Takeaway: The Machine-to-Machine Economy Is Coming
Ignore the free token hype. The real story is the convergence of AI compute and blockchain payments. As AI agents proliferate, they will need to pay for each other’s services—micropayments for model inference, data access, and storage. Zhipu’s campaign is a beta test for a future where AI tokens are settled on-chain. The next step is a tokenized payment layer for AI, where compute credits are tradable, programmable, and interoperable. Watch for Zhipu or a competitor to launch a crypto token tied to model usage. That will be the signal of a true paradigm shift. Until then, the free tokens are just noise—liquidity ghosts in the fog of the AI boom.