AI demand remains firm, but financing, power and China’s lower-cost competition keep returns risk elevated.

The underlying risk picture is broadly unchanged. Strong AI cloud contracting and enterprise infrastructure demand continue to support accelerator, server and networking demand, while large spending commitments still outrun demonstrated sector-wide cash returns. Financing and infrastructure execution became more salient: Oracle’s New Mexico project delay highlighted power and contract risk, and Texas froze new data-centre permits pending grid audits. China added no verified immediate shock, but Alibaba’s full-stack expansion and Huawei’s domestic-chip progress reinforce the longer-run possibility that cheaper capable supply broadens adoption while pressuring model and cloud margins.

• Hyperscaler spending remains far ahead of directly evidenced AI monetisation; Goldman estimates roughly $300 billion of annual AI revenue would be needed to break even on the largest US hyperscalers’ investment.
• Credit investors are demanding greater selectivity and concessions for AI-linked borrowers, even without a broad investment-grade stress event.
• Power availability, permitting and community opposition are moving from planning constraints to project-level execution and financing risks.
• The dashboard’s high capex/cash-return, monetisation, concentration and macro readings remain jointly adverse despite strong semiconductor demand.

• Oracle reported more than $30 billion of additional AI-cloud contracts, a $664 billion backlog and cash burn below expectations, indicating that some capacity is being underwritten by customer commitments and prepayments.
• HPE raised forecasts on AI server and networking demand; management described demand as exceeding supply, with enterprise adoption contributing to orders.
• NVIDIA data-centre revenue growth remains exceptionally strong in the authoritative reading, arguing against a near-term broad accelerator-demand break.
• Cheaper Chinese and other open-weight models may raise AI adoption, inference volumes and demand for compute rather than simply displace spending.

• Whether the Oracle/Blue Owl power delay changes project financing terms, tenant commitments or construction schedules beyond the individual campus.
• Evidence that AI-linked debt concessions widen materially relative to comparable investment-grade issuers.
• Texas grid-audit outcomes, permitting policy and whether other jurisdictions impose comparable filters on data-centre projects.
• Conversion of Oracle and other cloud backlogs into revenue and cash flow, rather than further contracted capacity alone.