Must C

研报英文原文证据摘录

Must C

However, constraints cover the full stack of energization —

transformers, switchgear, turbines, fuel cells, substation work, and

skilled labor. That is why the market is increasingly bifurcating between

those who can secure physical capacity and those who must buy access

from them. This strengthens the hand of hyperscalers and infrastructure

owners relative to model labs that do not control cloud, chips, sites, or

power procurement.

Improving energy efficiency is a key focus, through both better

performance per watt and power usage effectiveness (PUE). Better

performance per watt takes place at the chip and software level through

improvements in liquid cooling designs, re-architecting software and

newer silicon generation to increase the output per watt. Put differently,

PUE improvements occur at the facility level before electricity even gets to

the servers.

As AI costs rise, we note a shift under way in select use cases toward

open-weights (partially open) models, reflecting their lower inference

costs relative to frontier closed-source models.3 (Toms Hardware, May

23) However, lower inference costs would drive higher overall compute

demand, increasing data center utilization — consistent with Jevons

Paradox, where falling costs lead to greater usage. (Bloomberg, June 10)

Model use in some areas would partly migrate from frontier models such as

GPT-5.5 and Opus 4.8 to those such as GLM 5.2 or V4 Pro or their

equivalents in the future. The scaling law from Hoffman (2022), which had

shown that optimal performance under a fixed compute budget is

achieved by scaling parameters and training tokens jointly and

approximately linearly4 (and which has helped justify vast investments in

the past), has since been superseded by over-training smaller architecture