Risk eased as open-model pressure moderated, while China’s cloud build-out keeps capex and competitive risks elevated.

The public risk reading eased from 3.250 to 3.125 as GLM 5.3’s measured capability shortfall to the leading proprietary comparator widened to 15.92% from 11.62%, reducing immediate open-model rent-compression pressure despite its much lower list price. The wider investment picture is otherwise broadly unchanged: spending and infrastructure commitments remain large, Chinese AI-cloud investment is accelerating, and monetisation still trails the pace of capital deployment.

• Capex remains unusually high relative to cash returns and measured cloud-revenue conversion, leaving returns dependent on sustained utilisation and pricing.
• Alibaba reported 75% quarterly capex growth alongside 45% AI-cloud and compute-services revenue growth, illustrating that Chinese commercial progress still requires heavy infrastructure spending.
• AI-data-centre financing is broadening: ByteDance secured a reported US$29.6 billion loan, while Barclays warned that Asian lenders are becoming more selective on data-centre projects.
• Digital Realty’s large development pipeline and IREN’s costly conversion of legacy power capacity toward AI cloud underscore how much capital is being committed ahead of realised end-demand revenues and asset lives remaining uncertainly short.

• NVIDIA reported data-centre revenue of US$89.0 billion, up 117% year over year, with a US$108 billion next-quarter revenue outlook; accelerator demand remains exceptionally strong.
• Hut 8 disclosed long-term, take-or-pay AI-campus leases with an investment-grade tenant, providing an example of contracted rather than purely speculative capacity demand.
• Alibaba Cloud’s external revenue grew 45% and segment margin expanded, while Baidu’s AI Cloud Infra revenue grew 50%; this is evidence of commercial uptake, not just model releases.
• Chinese progress is not unambiguously negative for the cycle: lower-cost capable models can broaden adoption, increase inference volumes and raise demand for cloud and power even if they pressure frontier-model pricing.

• Whether hyperscalers translate AI capacity into faster cloud revenue, improved free cash flow and clearer enterprise productivity gains.
• Whether data-centre lending selectivity, power constraints or permitting delays produce cancellations, delayed energisation or higher required returns.
• Whether Chinese domestic chips and Qwen/other model deployment narrow performance gaps while lowering inference costs.
• Whether demand strength broadens beyond the current AI infrastructure leaders, reducing concentration and valuation fragility.