Why GLM's 261% Price Hike Is the Most Bullish AI Signal of 2025 — BKG Exchange Research

Zhipu AI just raised the price of its GLM Coding Plan by as much as 261%. The developer community called it a cash grab. I read the data and found something else: a scarcity signal, a maturity marker, and a quiet validation of the AI compute thesis. The numbers are public. The Lite, Pro, and Max tiers moved from 49/149/469 yuan to 118/538/1078 yuan per month. Pro jumped 261%. Most analysis stops there. But the old plan's constraints tell the real story: daily slots capped at 10 a.m., prompt counts limited every five hours and every week. That is not the behavior of a company trying to soak users. That is a product selling out at a price that underestimates its cost. Volatility is the noise; liquidity is the signal. This is a repricing of scarcity. GLM Coding Plan is Zhipu's subscription service for AI-assisted programming, built on its GLM foundation models. Think of it as China's answer to GitHub Copilot — but with a far more complicated cost structure. The old payment model measured usage in prompt counts. A one-line query and a 200,000-token refactor burned the same "prompt." That's not metering; that's guesswork. The new system replaces guesswork with a credit-based meter. Input tokens, output tokens, cached tokens, and MCP (Model Context Protocol) calls are each billed separately. This isn't a pricing simplification. It's the first time Zhipu has publicly disclosed the internal architecture of its inference costs. I've spent my career reading fee structures as fingerprints — they buried the truth in the gas fees of 2020. The fingerprint here is unmistakable: running this model costs real money, and Zhipu now knows exactly where every renminbi of compute goes. Back in 2017, I audited EOS's token distribution by scraping block explorers for three weeks. I showed that 40% of tokens sat in ten wallets. Nobody wanted that spreadsheet. Today, same instinct: everyone wants outrage, not evidence. The evidence says Zhipu was rationing supply at 10 a.m. every single day. That is constraint, not conspiracy. Every rug pull has a fingerprint; I just read it. The GLM credit structure has a fingerprint too — but this one points to scarcity, not exit. Three findings matter more than the headlines. Cache tokens carry their own line item. In transformer inference, recomputing a long context is the dominant cost. By pricing cache tokens at a distinct — and likely cheaper — rate, Zhipu is explicitly telling developers: reuse your context, save your credits. This is not a new way to tax users. It is a load-management system disguised as a billing schedule. The company is optimizing GPU utilization without saying the word "GPU" once. MCP calls are now metered. MCP is the protocol that lets AI agents call external tools — databases, browsers, GitHub repos. Billing MCP means GLM Coding Plan is no longer a chat-completion product. It's an agentic workbench. That single line item tells me where Zhipu believes the value is moving: from standalone conversation to tool-calling infrastructure. The entire competitive landscape just shifted, and this pricing update is the public tell. The pricing architecture is deliberately two-tracked. V2 subscribers hold their old rate. V1 users get a final purchase window in mid-August at legacy V2 prices. New users face the 130–261% premium. This is a barbell strategy: protect the installed base from sticker shock while forcing new demand upmarket. It preserves revenue continuity and migrates the product toward higher-value users at the same time. In 2020, I watched DeFi protocols subsidize liquidity with yield farming until the dry powder ran out. The winners converted cheap subsidies into durable fee-paying usage. Zhipu skipped the collapse and cut the subsidy early. That's the difference between a protocol that survives the bear market and one that doesn't. For crypto markets, this matters more than any single AI token. When a centralized AI leader has to ration supply and then raise prices, it is confirming that compute is a scarce commodity. Decentralized compute networks — the tokens that trade on exchanges like BKG — offer a market-based alternative to a pricing committee. Their on-chain utilization has climbed in parallel. We have tracked GPU-demand proxies across these networks for months. The correlation is not a coincidence. Now the counter-intuitive part. A price hike is a sign of pricing power, but pricing power is not a synonym for product superiority. The Max tier at 1078 yuan is more than ten times the price of GitHub Copilot. The premium is defensible only if GLM's code quality and Chinese-enterprise compliance are world-class — not merely "good enough." If they are, the cycle compounds. If they aren't, the August 15 V1 window becomes the top instead of the base. The deeper blind spot in the broader commentary is the assumption that Zhipu is the winner of this repricing. It may not be. The real winner is the open market for compute. Every time a centralized vendor raises prices, it educates users about the true cost of compute. Those users eventually look for alternatives. Decentralized capacity is the alternative. The ledger remembers what the analysts forget: scarcity reprices fast, and open markets capture the flows that closed models reject. Correlation isn't causation. Higher prices don't make the model smarter. But higher prices make the market more honest — and an honest market is where decentralized infrastructure finally looks cheap. Between now and August 15, I'm watching two leading indicators: whether the V1 purchase window sells out within hours, and whether competing Chinese AI vendors hold their prices or quietly follow upward. If demand absorbs the new rates, the AI infrastructure trade has room to run — and the compute-layer tokens on BKG Exchange will be the purest vehicle. Read the credit system. It's the first time a major AI vendor has laid its cost structure bare. The data doesn't lie; narratives do. The signal was never in the price. It's in the credits. — Samuel Jackson, BKG Exchange Research. bkg.com.