China’s HBM bottleneck tempers domestic chip competition, while global capacity and debt commitments keep cycle risk elevated.

The public risk reading is unchanged at 3.125 and the underlying risk picture is broadly unchanged: investment remains far ahead of demonstrated cloud-revenue conversion, free cash flow remains under pressure, and data-centre construction is still expanding rapidly. The notable new development is that Chinese accelerator suppliers have raised prices amid an HBM shortage, which may slow China’s domestic chip substitution and reduce its near-term price advantage; this is offset by ByteDance’s $29.6 billion loan and continuing regional capacity commitments, which sustain AI infrastructure demand and capital intensity.

• Capex and cash generation remain misaligned: aggregate hyperscaler capex is up 72.29% while aggregate free cash flow is down 16.11%.
• Cloud-revenue growth of 24.30% remains well below the 64.73% capex growth used in the monetisation proxy, leaving returns dependent on future utilisation and pricing.
• AI financing is becoming more credit-sensitive: recent reporting shows weaker demand metrics and wider concessions for large hyperscaler bond issues as supply grows.
• Large new capacity commitments in Australia and Malaysia validate demand but add to the fixed-cost, power and delivery burden before end-customer returns are fully proven.

• NVIDIA Data Center revenue growth of 117% remains strong evidence that accelerator demand has not broken.
• OpenAI’s multi-year Firmus capacity agreement and NVIDIA’s planned Australian buildout provide concrete demand commitments rather than speculative announcements.
• China’s HBM shortage and higher domestic accelerator pricing constrain the speed and cost advantage of domestic alternatives, partially countering near-term competitive pressure on US suppliers.
• Chinese commercialisation evidence is mixed rather than uniformly disruptive: Zhipu reported sharply higher revenue, but remained loss-making and faces intense price competition.

• Whether large AI-related bond and loan issuance produces sustained spread widening, weaker order books, or tougher covenants.
• Whether contracted capacity converts into paying utilisation and cloud revenue rather than reserved but underused infrastructure.
• HBM availability, pricing and allocation, especially whether shortages broaden beyond China or delay accelerator deployments.
• Evidence that lower-cost Chinese and open-weight models gain enterprise adoption outside China and alter proprietary-model pricing or compute demand.