Ping An CTO Ray Wang: From Tokens to Value Creation

Ping An reported that average daily AI token consumption exceeded 300 billion, up tenfold from 30 billion at year-end, supported by computing capacity expansion to 1,500 PF from 800 PF. Full-stack optimization increased token output per unit of computing power by 368% and cut cost per token by 76%, while its AI Express Service surpassed 100 million interactions with a 92% end-to-end resolution rate and processing times more than 50% lower than traditional models. AI also generated about 80% of the Group's software code output and helped Ping An P&C achieve RMB7.11 billion in first-half claims savings through fraud detection.

Analysis

The investable question is not model scale but whether automation converts into a lower expense ratio without offsetting increases in cloud, data-governance and model-validation spend. For 2318 HK, even a modest sustained reduction in acquisition, servicing and claims-handling costs could support ROE and justify a valuation premium to China Life (2628 HK) and CPIC (2601 HK), whose business mixes are more exposed to agency productivity and investment-income sensitivity. The next proof point is the FY26 operating expense ratio and new-business-value margin, not usage metrics.

The more differentiated potential benefit is claims severity control: fraud detection and triage can improve combined ratios while reducing customer friction, a combination that competitors will find difficult to replicate without comparable proprietary claims and healthcare data. Conversely, broader automated service may weaken product differentiation if it becomes an industry standard; Alibaba (9988 HK) and Tencent (0700 HK) are likely to capture part of the value through inference infrastructure and enterprise-model tooling rather than allowing insurers to retain all productivity gains.

Near term, this is unlikely to rerate the stock materially absent audited cost savings or raised earnings guidance; investors have learned to discount enterprise-AI adoption claims. Over 6-18 months, successful healthcare workflow deployment could increase customer retention and cross-sell while lowering loss ratios, but regulatory scrutiny of AI-mediated clinical recommendations is the asymmetric risk: a high-profile error, tightened data rules, or mandated human review could eliminate much of the labor-cost benefit. The contrarian view is that rapidly rising inference demand may become a cost problem before it becomes an earnings catalyst if unit-cost declines cease or domestic compute availability tightens.

Key Decisions for Investors

• Maintain a watch-list bias to long Ping An (2318 HK) versus short China Life (2628 HK) over the next 1-3 months, but enter only if Ping An reports a sequential decline in operating expense ratio or upgrades FY26 profit guidance. Target 8-12% relative upside; exit if expense ratio fails to improve by FY26 results or if the relative spread widens 7% from entry.
• Do not underwrite a standalone AI rerating before audited disclosure separates recurring savings from AI infrastructure, personnel and compliance costs. Set an earnings alert for claims-ratio, expense-ratio, NBV-margin and technology-expense disclosures; absent these, treat the news as sentiment-positive rather than a catalyst.
• For 6-18 month exposure, prefer a small long 2318 HK / short 2601 HK pair if Ping An demonstrates lower claims severity while preserving premium growth. The pair isolates insurer beta and rates risk; thesis is falsified by CPIC matching the efficiency improvement or Ping An's health-related costs rising faster than servicing savings.
• Monitor Alibaba (9988 HK) and Tencent (0700 HK) enterprise-AI commentary for inference-price compression and financial-services workloads. If hyperscaler pricing falls materially, Ping An's cost advantage may prove non-proprietary, reducing the case for a 2318 HK multiple premium.