Risk stays elevated as spending and financing outpace cash conversion; China adds competition but no new cycle break.

The underlying risk picture is broadly unchanged. AWS cited accelerating AI-linked revenue and a large order backlog, supporting demand, while recent reporting still shows substantial AI-related borrowing, power-delivery frictions and financing sensitivity at data-centre projects. China advanced AI Plus implementation and domestic Ascend software tools, which could lower Chinese deployment costs and eventually intensify competition, but no measured migration, pricing change or global revenue impact was reported that would establish a material China-driven change in the cycle.

• Fact: AWS demand evidence does not remove the wider mismatch between very rapid infrastructure spending and slower recognised cloud-revenue growth and aggregate free-cash-flow conversion.
• Fact: AI borrowers continue to tap capital markets despite higher rates; Project Jupiter remains a live example of how power and permitting delays can reach lease and financing structures.
• Interpretation: If build schedules slip while capex and interest costs continue, project-level cash conversion and lender protections become more important than contracted headline demand.
• Fact: China is pairing AI Plus adoption policy with domestic Ascend ecosystem development and technology-export controls, reinforcing a separate Chinese AI stack and potential competitive fragmentation.

• Fact: AWS reported second-quarter revenue growth of 36.7% and said its AI revenue run rate exceeded $25 billion and was growing at a triple-digit annual rate, evidence that some infrastructure demand is converting into revenue.
• Fact: September global AI-debt issuance reportedly fell from August and well below June's peak, a partial indication that financing volume is not accelerating monotonically.
• Fact: Oracle, Blue Owl and STACK said their Project Jupiter financial commitments remained intact despite the force-majeure notice.
• Interpretation: Cheaper Chinese models and domestic hardware compatibility can broaden AI adoption and inference demand, potentially raising total compute consumption even if they pressure some model-layer rents.

• Evidence that cloud backlog converts into recognised revenue and free cash flow rather than additional contracted capacity.
• Any change to Project Jupiter power timing, rent obligations, lender protections, syndication or tenant commitments.
• Whether data-centre power workarounds and permitting delays spread beyond individual campuses.
• Measured production adoption of DeepSeek-compatible Ascend tools, including developer usage and non-Huawei workloads.