China Technology

研报英文原文证据摘录

China Technology

Macquarie Equity Research China Technology

Compute shortage drives another leg of AI capex

• The most consistent message across our meetings with IDC and cloud service companies

was that GPU supply, not demand, is the binding constraint. This supply tightness is

translating directly into price inflation. We note GPU rental prices have tripled over the past

two quarters. Meanwhile, the capex cycle of hyperscalers and some AI labs, which have

recently replenished with new capital, has not peaked. Companies are accelerating IDC and

GPU rental commitments for their growing model training and inference scales.

Þ Neoclouds gaining traction as compute buyer pool diversifies. Demand for

independent compute providers appears to be expanding beyond frontier labs into

enterprises, AI-native startups and application developers that need flexible access

to scarce GPU resources. This creates a favourable backdrop for neocloud providers,

particularly those that can offer fast provisioning. We also see incremental demand

from training larger models and inference-heavy use cases, where agentic workflows can

generate persistent token consumption even after the initial model training phase.

Þ 'Token factories' one of the buzzwords this year. A striking number of companies

are now branding themselves around token production, token infrastructure or token

delivery. We interpret this as an indirect signal that compute remains structurally tight

in China. Near term, the competitive edge, like neoclouds', still lies in securing access to

high-performance chips, stable clusters and reliable deployment environments. Longer

term, however, the source of advantage is likely to shift from simply owning compute to

using compute more efficiently via optimising token throughput.