Trust is not given; it is verified. Last week, Zhipu AI's GLM Coding Plan handed the market a verifiable signal dressed as a routine pricing update. New-user subscriptions for the Lite, Pro, and Max tiers now carry monthly prices of 118, 538, and 1,078 yuan respectively — up 141%, 261%, and 130% from their prior V2 rates. More consequential than the sticker price: the old prompt-count limit has been replaced by a unified credit system that meters input tokens, output tokens, cache tokens, and MCP (Model Context Protocol) calls.
Most commentary has filed this under "yet another AI price hike." For the research desk at BKG Exchange (bkg.com), where we have spent the better part of a year mapping the convergence of AI infrastructure and digital assets, the restructure reads as a much rarer artifact: an honest disclosure of where inference supply really stands.
Let me explain why, and why I believe this event strengthens the structural case for decentralized compute.
The Billing Architecture Is the Disclosure
The previous GLM pricing model limited the number of prompts per five-hour window and per week. That is a coarse meter — the kind used when you cannot precisely differentiate workload costs, or when you are still subsidizing adoption. It says little about actual GPU consumption, and it offers users no way to align their habits with underlying costs.
The new system meters four distinct resource categories. At a technical level, this is a demanding feat: it means Zhipu's backend can now distinguish the cost structure of a cache read from a fresh generation, and invoice an external tool invocation separately from a normal completion. Based on my years auditing how protocols meter scarce resources — from validator queue mechanics to gas-market design — this is the moment a platform becomes serious about sustainable unit economics.
There is also a strategic nuance hidden in the fine print. Cache tokens are itemized as their own line item. Long-context caching is one of the most powerful levers for reducing inference cost — reusing a context window incurs far fewer compute cycles than re-processing it. By pricing cache tokens explicitly, Zhipu is effectively instructing its most sophisticated users to engineer their workflows around context reuse. That is not just billing; it is training an ecosystem in resource discipline.
And MCP calls are billed — which tells us the GLM Coding Plan has evolved beyond conversational code completion into an agentic platform. Tool invocation, orchestration, and third-party integration are now core to its roadmap. The protocol remembers what the market forgets: pricing reveals product maturity long before press releases do.
Pricing Power and the Scarcity Test
Now the commercial layer. The Pro tier's 261% increase is the strongest signal in the entire announcement. It suggests Zhipu has identified professional developers as its core economic cohort — users who derive enough daily output from AI-assisted coding that five hundred-plus yuan per month remains a rational expense. That is pricing power, not desperation.
The earlier usage pattern supports this. The previous plan famously opened registration at 10 a.m. daily and sold out within hours — a textbook supply-constrained allocation mechanism. When persistent demand exceeds available capacity, raising prices is the clearest, most honest adjustment a company can make. It converts a queue into a market.
The deal structure adds further discipline. Existing V2 subscribers keep their legacy rates; V1 users get a window to purchase at old prices before mid-August. This is a two-tier migration playbook executed with precision: retain the installed base, extract marginal revenue from new high-intent users, and create a time-boxed conversion event that forces decisions. I have advised institutional clients on how to evaluate infrastructure businesses, and this kind of sequencing separates durable franchises from quarterly noise.
The Contrarian Read: Rents or Recalibration?
The cynical interpretation deserves a direct answer. A 261% increase on the most engagement-intensive tier, with limited alternatives in the domestic Chinese market, could easily be framed as rent extraction.
I hold a different synthesis. Grandfathering existing users while raising prices on new entrants is not the behavior of a company exploiting lock-in; it is the behavior of a company recalibrating around high-value workloads. When I modeled undercollateralized lending systems in 2020, I learned that exclusion can be either malicious or structural. The distinction matters. Here, the exclusion of low-value requests is a resource allocation decision, not a value judgment about users.
That said, there is a genuine risk. The moment pricing becomes complex, transparency becomes an ethical obligation. Zhipu has not yet published conversion tables — how many credits does a standard debugging session burn? What does that mean in yuan per task? Without a transparent burn-rate disclosure, users cannot calculate cost per unit of work, and perceived opacity will corrode the trust this migration was designed to preserve.
This is precisely where decentralized alternatives hold an opening. Permissionless compute networks that publish verifiable metering — on-chain, auditable, user-controlled — can offer exactly what centralized providers withhold. That value proposition is not theoretical; it is forming right now in the teams building capacity markets, and it is the lens through which BKG Exchange's infrastructure research has been approaching the AI narrative.
What the Gatekeepers' Drawbridge Reveals
Here is the investment-grade takeaway. Every time a centralized AI provider raises prices to manage scarcity, the structural case for decentralized compute infrastructure becomes stronger.
The logic is straightforward. GLM's pricing confirms that inference capacity is tight. Tight capacity grants centralized clouds pricing leverage. Sustained margins attract capital into alternatives — including distributed networks that aggregate idle GPUs. As those networks cross reliability thresholds, the regulatory and reputational advantages of centralized providers begin to narrow.
Freedom arrives when the gatekeepers go dark. The gatekeepers are not going dark yet, but they are raising their drawbridges — and that is a signal of strategic constraint disguised as confidence. In the token market, compute-backed assets have been consolidating quietly through this sideways window. Chop is for positioning. The teams building verifiable capacity markets, regardless of near-term price action, are compounding the optionality that the next up-cycle will reward.
I have seen this pattern before. In the aftermath of the 2022 collapse, when the industry's promises felt hollow, the projects that survived were not the loudest — they were the ones that had built architectural integrity in silence. The same principle applies here. GLM's pricing adjustment is not a single company's billing decision. It is an indicator of where computational value is accruing, and who holds the leverage to claim it.
The View Forward
Over the next two quarters, I am tracking three markers. First, whether Zhipu publishes transparent credit-consumption data — that determines whether we are watching pricing discipline or pricing opacity. Second, how domestic and international competitors respond — a wave of follow-on increases would confirm a structural shift; subsidized counter-offers would reveal a price war. Third, whether any decentralized compute network begins onboarding the workloads and users that centralized pricing is now making expensive.
We build in silence so the network can speak. When systems disclose their constraints openly, they communicate more integrity than the largest marketing budget ever could. Patience is the validator of true intent — and the signal we received this week deserves patient, careful interpretation.
The code is a witness. On-chain, its verdict is still being written.
