The Pensions Regulator has warned that apparently diversified portfolios can carry common exposure to the artificial-intelligence investment cycle. A fund may hold equities, private credit, infrastructure and property while several positions still depend on the same data-centre demand, semiconductor supply chain or financing conditions.[1][2]
Illustrative portfolio review questions
| Area | Example assessment |
|---|---|
| Listed equities | Exposure to AI developers, chips, cloud and power suppliers |
| Private markets | Valuation and refinancing sensitivity of AI-linked assets |
| Credit | Borrower dependence on AI capital spending or customer concentration |
| Infrastructure | Demand assumptions for data centres, power and networks |
| Adviser oversight | Whether models capture common factors across funds |
The issue is not that every AI-related investment is unsuitable. It is that a portfolio’s labels may understate the extent to which different holdings respond to the same shock.[1]
Portfolio-wide stress tests
A useful scenario could combine lower AI-company valuations, weaker data-centre demand, higher financing costs and reduced exit liquidity. Testing each fund independently may miss the effect if several managers own different securities exposed to that same chain.[1]
Trustees are also expected to question advisers about assumptions, data and time horizons. A risk model built on a period of rapidly rising investment may not capture a sudden reversal, particularly in private assets where valuations update slowly.[2]
This analysis does not imply that the regulator has imposed a specific AI allocation limit. The published expectation is stronger governance: identify hidden concentration, understand the drivers and test whether the scheme can tolerate a severe but plausible change.[1]
Sources & further reading
- TPR sets expectations on hidden concentration risks The Pensions Regulator · 22 Oct 2026 · Accessed 30 Oct 2026
- UK regulator urges pension trustees to stress-test AI concentration risk Reuters · 22 Oct 2026 · Accessed 22 Oct 2026