For the first time globally, there is an open-source model that surpasses the practical intelligence threshold and can be infinitely modified.

Some startups are already taking action. The most focused discussion is quite interesting: not about coding, but about moving toward long-term agentic RL algorithms outside of coding. Healthcare, legal, finance, manufacturing... countless vertical fields with long-term tasks, which in the past remained in papers due to insufficient base model capabilities and inability to freely modify. GLM-5.2 has lowered the barrier to an open-source playable level.
If the beginning of the year OpenClaw accelerated the first wave of agentic consensus (solving "can it move"), then GLM-5.2 is the second acceleration (solving "can it continuously operate in complex environments"). The difference is—OpenClaw is a laboratory toy, while GLM-5.2 is an industrial-grade base model.
MIT summarized it in three sentences: freely modify, freely sell, and the only obligation is to retain the copyright statement. This is the premise for the "mass post-training movement" to succeed. Zero intellectual property risks, commercial use allowed, embedded in business, no need to disclose modifications.
Especially with recent events—Anthropic’s flagship model was required to be taken offline globally due to US export controls. "Domestic models + domestic computing power" has shifted from an alternative to a necessity. On the first day of GLM-5.2’s release, all mainstream domestic computing platforms were adapted, and the stock price roughly doubled in a week.
The biggest benefit isn’t from GPUs themselves (open source + adaptation by 8 domestic platforms → inference chip deflation). The real inflation pressure is layered in four areas (these four leading stocks are very clear, with AI mining some niche stocks that haven't surged significantly, with higher risks):
🥇 HBM—most critical globally
Multi-hop reasoning in long-term agentic tasks causes exponential bandwidth consumption. The three major manufacturers’ capacities are sold out, with a 50%-60% gap.
Wanrun Technology (002654): collaborated with Yangtze Memory to complete 12-layer HBM3E packaging samples adapted for Nvidia H200, securing the first batch of 300,000 units. The actual controller is the Hubei State-owned Assets Supervision and Administration Commission, same as Yangtze Memory. UBS and Morgan Stanley increased holdings in Q1.
🥈 Optical chips/InP—CPO inflation amplifier
InP substrate shortage exceeds 70%.
Yunnan Geology Industry (002428): subsidiary XinYao Semiconductor is a leader in InP substrates, already certified by Nvidia. The only domestic 6-inch InP substrate mass production, with a yield of over 70%. Huawei Hubble invested. Dual scarcity of germanium and indium phosphide.
🥉 ABF substrates—certainty of shovels
Ajinomoto ABF film price increased by 30%.
MacroChips Electronics (603002): GBF multilayer films have been sent for samples to TSMC and Changdian Technology.
🥄 CCL M9—material layer inflation
M9 price is ten times that of ordinary FR4.
South Asia New Materials (688519): M6-M8 has been supplied in bulk, M9 is in the process of being introduced. Small volume, high flexibility.
GLM-5.2 has pulled the trigger. The bullets are still flying.