OpenClaw has allowed model companies to taste the sweetness of Token economy for the first time.

On June 17th, GLM-5.2 was open-sourced. This time, it's different.
MIT License: modify freely, sell freely, only obligation is to retain the copyright statement.
Intellectual property risks are eliminated, companies can confidently embed models into their commercial products, and modifications do not need to be shared. Compared to GPL's "use my code and you must open source," MIT removes the barrier.
More importantly— all models now enter the era of long reasoning.
What exactly has GLM-5.2 changed?
In the OpenClaw era, agents work in "short bursts"— each task window is limited, planning-execution-termination, KV Cache size is controllable, hardware pressure mainly on computing power.
GLM-5.2's long-term reasoning is a "marathon"— 1 million tokens of lossless context, the model can hold all code, all decision history, all constraints in a single task. In actual tests, it processed 880k tokens in one go, nearly filling the window.
What does this change mean?
In the past, AI was "question and answer," each token consumption ended after one turn.
After GLM-5.2, agents begin to run real long-term tasks: decomposing goals → multi-round planning → repeated verification → tool tuning → coding and running code → re-planning based on feedback. A single task triggers hundreds of reasoning cycles.
Each cycle loads the full context into memory for recalculation.
Continuous computation, continuous communication, continuous read/write.
These three "continuals" completely change the hardware pricing logic.
What are the benefits of long-term agent reasoning?
🥇 HBM
KV Cache grows linearly with dialogue turns and context length, quickly exhausting GPU HBM capacity. Once KV Cache leaves the GPU local, bandwidth drops from TB/s to hundreds of GB/s— the problem shifts from "computing power" to "memory bandwidth."
Major manufacturers' capacities are sold out, with a 50%-60% gap, and the market size is projected to reach $54.6 billion in 2026.
🥈 Optical chips/InP
Long-term reasoning runs on clusters, each cycle requires inter-card synchronization. The longer the task and the more cycles, the more terrifying the communication.
Optical modules will reach a market of $26 billion in 2026, with a 60% annual growth. InP substrate shortages exceed 70%, indium prices up 90% year-over-year.
🥉 CPU
Long-term tasks require continuous task decomposition, tool invocation, process management, KV Cache scheduling. These tasks are hard for GPUs to handle, relying on CPUs.
The CPU/GPU ratio is approaching 1:1 from 1:8, and Intel's CEO publicly said "Many company CEOs are calling to rush CPU supplies."
❄ Liquid cooling
Short reasoning is pulse load, long reasoning is continuous full load. The same card, the actual power consumption for long tasks is 3-5 times that of short reasoning.
Rack power consumption jumps from 36kW to 200kW, air cooling can't handle it, liquid cooling shifts from "optional" to "mandatory."
🔌 Switches
Inference cluster bandwidth requirements jump from 100G to 400G, tens of thousands of cards need scheduling. InfiniBand and high-speed Ethernet benefit across the board.
📦 ABF substrate
Clusters expand from thousands to tens of thousands of cards, each chip needs packaging. Mitsui monopolizes over 90% of ABF film, with a 42% shortage projected for 2028.
Flour prices rise, bread will only get more expensive.
🧪 CCL M9
Mainboards and backplanes of inference clusters all require high-speed substrates. M9's unit price is 10 times that of ordinary FR4, and the AI CCL market will reach $18.7 billion in 2027, growing faster than optical modules.
OpenClaw ignited the fire, GLM-5.2 provides the fuel.
The former allowed model companies to earn their first tokens, the latter is bringing this market from labs into industry.
Simple operations still rely on storage, on light, enjoying the AI bubble.
$MU $SKHYNIX $LITE