Risk is unchanged: financing strain persists, while China’s chip-cost bottleneck tempers competitive pressure.

The underlying risk picture is broadly unchanged. Recent reporting points to larger and more complex debt funding for AI infrastructure, including greater exposure to smaller data-centre and neocloud counterparties, while Amazon’s long-term Qualcomm commitment and Oracle’s cloud growth support continuing infrastructure demand. In China, official capacity and efficiency initiatives reinforce domestic AI expansion, but reported increases in Huawei and Cambricon chip prices due to HBM scarcity limit near-term price-performance gains.

• S&P-related reporting says hyperscaler funding structures are becoming more complex and that credit quality is gradually weakening as capex rises and returns may take years to realise.
• The investment cycle remains exposed to a returns gap: aggregate hyperscaler capex growth is 72.29% while aggregate free-cash-flow growth is -16.11%, and AWS plus Google Cloud growth trails the capex pace used in the monetisation proxy.
• Amazon’s potential purchase of up to $60 billion of Qualcomm AI data-centre products includes equity warrants, illustrating that major infrastructure commitments can embed financing-like incentives and counterparty interdependence.
• China’s domestic accelerator push faces an immediate HBM constraint: reported 20% to 50% price increases for Huawei’s Ascend 950DT and 20% to 30% increases for Cambricon’s next-generation chip raise domestic deployment costs.

• NVIDIA Data Center revenue growth of 117% remains strong demand evidence, so there is no verified broad accelerator-demand break.
• Oracle reported 62% cloud revenue growth and 121% IaaS growth, alongside more than $30 billion of additional AI-cloud contracts, supporting the proposition that some capacity is being commercialised.
• Amazon’s Qualcomm partnership broadens the supply base for inference chips and signals durable customer demand beyond NVIDIA-centric procurement.
• China’s national plans target more than a fivefold increase in intelligent-computing capacity by 2030 and explicitly promote domestic-chip adaptation, while reported token and application activity points to expanding local adoption.

• Whether AI-data-centre borrowers without secured power, permits or investment-grade counterparties face meaningfully wider funding costs or project delays.
• Whether cloud and enterprise AI revenue accelerates enough to narrow the capex-to-monetisation gap.
• Whether custom inference-chip agreements translate into delivered volume, lower unit costs and reduced dependence on NVIDIA.
• Whether China can secure HBM and scale domestic accelerators without price inflation eroding its claimed cost advantage.