AI cycle risk is unchanged: financing and grid bottlenecks persist, while China adds capacity but no verified shock.

The underlying risk picture is broadly unchanged and the public risk reading remains 3.125. Financing pressure is held up by greater reliance on outside capital and renewed concern about delayed data-centre commitments, while Texas grid-connection restrictions underscore physical delivery risk. Offsetting this, surveyed enterprises report measurable AI value and Microsoft continues to report substantial paid-Copilot adoption. China’s new domestic chip and model initiatives add future competitive supply, but have not yet established a new near-term shock to global AI revenues, pricing or infrastructure demand.

• AI infrastructure funding is moving beyond corporate balance sheets into private credit and project structures, increasing opacity and sensitivity to delivery delays.
• Oracle-related lease and force-majeure concerns show that permits, power and construction timing can defer contracted compute revenue while financing obligations continue.
• Texas has paused new data-centre grid connections for an audit, highlighting that power availability can constrain deployment even when chip demand is strong.
• China’s expanding domestic chip, cloud and open-model ecosystem could lower inference costs and intensify pressure on proprietary-model economics over time.

• Microsoft reported Azure revenue above $100 billion and more than 30 million paid Microsoft 365 Copilot seats, supporting real commercial demand.
• KPMG’s survey found nearly six in ten large-company leaders reporting measurable AI value, with productivity the most frequently cited benefit.
• S&P Global notes that most large hyperscalers retain comparatively strong balance sheets and capacity for further leverage; Oracle is the notable weaker-credit exception.
• Chinese competition can also expand global adoption and token demand, increasing usage of cloud, networking, power and compute rather than simply displacing US profits.

• Whether delayed or contested data-centre projects produce cancellations, revised capacity commitments or weaker project-finance terms.
• Evidence that enterprise productivity gains translate into durable software, cloud and inference revenue rather than pilots and surveys.
• Chinese model pricing, domestic-chip availability and cloud monetisation, especially whether they alter global inference pricing or compute demand.
• Grid interconnection decisions, power-price allocation and memory/networking supply constraints.