Allocators overhaul risk models to avoid hidden AI concentration

Key Points

As the expansion of AI drives unprecedented capital spending, investors are overhauling risk models, setting aside simple label-based classification in favour of causation to analyse portfolios that might appear well-balanced across private credit, real assets and listed equities.

Tony Appiah
Nuveen

"Investors need to look through labels and focus on underlying economic risk factors," said Tony Appiah, managing director and client portfolio manager at Nuveen, speaking at a recent AsianInvestor event.

"To use a simple example: suppose you own equity in a hyperscaler, its corporate bonds, a data centre where the hyperscaler is the anchor tenant, and an infrastructure project fully powering that data centre. On paper, those are four separate asset classes. In reality, you are exposed to the exact same AI investment ecosystem," he said.

Appiah stressed that this is not a reason to avoid the sector entirely, noting that "AI represents a multi-year structural cycle…It's about ensuring you aren't taking unintended, concentrated bets across your portfolio."

UNCOVERING HIDDEN CORRELATIONS

To identify these hidden concentrations, allocators are increasingly moving beyond traditional correlation analysis and examining the underlying economic drivers shared across different investments. That means de-smoothing return data and shifting risk tests from return correlation to common fundamental drivers.

"Risk models are moving from classification to causation," said Ninghui Liu, head of investment strategy and research, Apac, at State Street Investment Management. Rather than grouping assets strictly by type, geography or sector, more useful models group them by cash-flow drivers, such as power demand, global capital spending intensity, long-duration discount rates and US dollar funding conditions.

Ninghui Liu
State Street IM

"In AI power generation and data-centre real estate, the common drivers are straightforward to identify, namely the credit quality of a small group of hyperscale tenants, the cost and availability of electricity, and the path of long-term interest rates that underpins both project financing and equity valuations," Liu said. "When those three variables move adversely at the same time, the private credit position, the real asset and the listed equity are effectively one trade. Scenario analysis captures that; a historical correlation matrix generally does not."

DISCIPLINED FRAMEWORKS AND METRICS

"The most useful approach is a total-portfolio one, in which the theme is given an explicit risk budget at the whole-portfolio level rather than being accumulated incidentally inside several asset-class mandates," Liu said. "Alongside that budget, allocators should track a small number of metrics consistently. Effective number of independent bets is more informative than a simple count of holdings, because it captures how much genuine diversification the portfolio retains."

He added that allocators should also monitor look-through revenue exposure across public and private holdings to measure how much cash flow depends on one narrative, alongside marginal contribution to risk. "Rebalancing rules should then be pre-committed, because the discipline is hardest to apply when the theme is performing well," he noted.

Other asset managers report similar shifts in portfolio construction frameworks, with a growing emphasis on measuring risk through underlying factors rather than traditional asset-class buckets.

Matthias Scheiber
Allspring Global Investments

Matthias Scheiber, senior portfolio manager and the head of the multi-asset team at Allspring Global Investments, said that many allocators are using factor-based portfolio construction, scenario analysis and contribution-to-risk frameworks to ensure exposures remain balanced.

"Metrics such as risk-budget allocations, factor exposure, concentration limits and stress-test outcomes are becoming more important than simple asset-class weights. The objective is to maintain exposure to long-term themes while avoiding excessive dependence on a single return driver," he said.

At the same time, risk models are evolving towards factor and regime-based frameworks, including AI and machine learning, to map exposures across growth, inflation, rates, liquidity and behavioural drivers under different market environments.

For allocators, the evolution reflects a broader recognition that diversification can no longer be assessed solely by counting asset classes or geographic exposures.