AI build-out remains cash-intensive and power-constrained, while China’s stack advances without a verified broad commercial break-out.

The underlying risk picture is broadly unchanged. Amazon reported stronger AWS revenue and operating income but AI-led equipment spending pushed trailing free cash flow negative; Meta continued very high capital spending with minimal quarterly free cash flow. Project Jupiter’s power-related delay continues to test data-centre contract and financing structures, while higher AI-cloud pricing supports near-term capacity demand. China added evidence of domestic-stack progress through DeepSeek and Huawei, but higher domestic chip prices and no verified broad commercial adoption leave its near-term impact on global AI economics mixed.

• Amazon said AI investment drove a $66.1 billion year-on-year increase in trailing equipment purchases, turning trailing free cash flow to a $7.6 billion outflow.
• Meta reported $31.1 billion of quarterly capital expenditure and only $784 million of free cash flow, while maintaining $130-145 billion 2026 capex guidance.
• Oracle’s Project Jupiter force-majeure notice was tied to power delays; Reuters reported an approximately one-year delay, illustrating that grid risk can defer project returns and complicate financing.
• DeepSeek and Huawei’s open-source Ascend tooling reduces a software-portability obstacle for China’s domestic hardware ecosystem; if adoption scales, it could shift demand and bargaining power away from incumbent US stacks.

• AWS second-quarter revenue grew 37% year on year and operating income rose 64%, showing material cloud demand and operating leverage alongside the spending increase.
• Nebius raised selected GPU rental prices by 17-21% from October 1 and reported four contracts averaging more than $1 billion, consistent with scarce compute capacity rather than a broad demand retrenchment.
• Oracle, Blue Owl and STACK said they remained aligned and that Project Jupiter’s multiyear financial commitments were unchanged, limiting evidence so far of a wider funding withdrawal.
• China’s domestic accelerator effort still faces cost constraints: Reuters reported 20-50% indicated-price increases for Huawei’s Ascend 950DT amid high-bandwidth-memory pressure.

• Whether Project Jupiter’s power delay changes tenant payments, lender protections, syndication or construction milestones beyond the reported notice.
• Whether AWS and other cloud backlog converts into recognized revenue and free cash flow fast enough to narrow the investment-to-returns gap.
• Whether higher GPU-cloud prices persist through new capacity additions or instead trigger demand deferral and price competition.
• Whether Ascend-compatible tooling delivers measured production utilisation outside announced partnerships, and whether Chinese low-cost models gain global paid adoption.