Frans Verhaar: Subject the entire portfolio to an AI stress test
This column was originally written in Dutch. This is an English translation.
By Frans Verhaar, Managing Director, Head of Continental Europe, bfinance
When assessing AI risks within investment portfolios, the focus tends to be on equities. However, to identify all blind spots, it is just as important to scrutinise bonds as well.
In bond portfolios, exposure to large technology companies may increase without you consciously opting for more AI. After all, new bond issues alter the composition of the benchmark, particularly in the case of high-grade, long-dated corporate bonds. A portfolio-wide stress test can reveal the extent to which these investments are sensitive to the same setbacks as equities and private financing, and where differences actually offer protection.
Concentration behind the index average
Our analysis of Bloomberg data shows that, at the end of September 2026, Alphabet, Amazon, Meta, Microsoft and Oracle together accounted for 4.85 per cent of the market value of the Bloomberg US Corporate Index. At the end of 2024, this figure stood at 2.67 per cent, representing almost a doubling. Nevertheless, these companies do not dominate the broad index. Banks collectively carry much greater weight.
The picture changes as soon as we zoom in on high-quality corporate bonds. Alphabet, Amazon, Meta and Microsoft together represent around a third of the AA-or-higher segment within that same US index. Investors who select on the basis of high credit quality may therefore have considerably greater exposure to this group than the broad index figure suggests. Their weighting may also be greater in long-term credit mandates. A high average rating does not tell the whole story about diversification.
Three mandates may be looking in the same direction
Those who report exclusively by asset class may miss this interconnection. A group may feature in both the equity portfolio and the bond mandate. Furthermore, private financing of data centres may depend on demand for computing capacity or on contracts with the same technology companies.
Each manager may remain within their limits, whilst the overall portfolio becomes more vulnerable to a disappointing AI investment cycle. Therefore, in addition to the size of each position, shared dependencies are also relevant. Different mandates may ultimately rely on income from the same investment wave.
Look beyond the label to the repayment terms
Simply adding up these exposures does not in itself provide a measure of risk. Disappointing profit growth can affect a share, whilst interest and principal are still being paid. Bondholders and shareholders have different rights, so guarantees, collateral and the position within the capital structure matter.
One key distinction lies in the source of repayment. Does it come from the broad cash flows of a profitable group, or from a single data centre with a limited number of customers? Who bears the risk if a tenant pulls out? And how much protection does the collateral offer if demand for data centres falls short of expectations? These are questions to which a sector label such as ‘technology’ or ‘infrastructure’ provides no answer.
Ask all managers to consider the same setback
A practical starting point is a joint overview of the relevant companies, projects and guarantees, cutting across all mandates. Then ask the managers involved to assess the same scenario: the investments go ahead, but the income comes in later than anticipated. Use the same assumptions and the same time horizon. This ensures the results are comparable. It is a stress test, not a prediction.
What happens then to valuations, credit spreads and the need for refinancing? Where do cash buffers absorb the setback and where is new capital required? And which investments continue to offer protection under the same assumptions? The same setback can affect equities, corporate bonds and data centre financing in very different ways.
What the combined picture adds
The overview of exposures and the comparable scenario outcomes together reveal where risks are accumulating. This makes it clearer whether the combined exposure has been deliberately chosen and fits within the agreed risk appetite. This interconnection is easily overlooked when each mandate is assessed in isolation.
Perhaps the returns and protection are convincing. In that case, there is no need to sell anything, and a different benchmark is not automatically better. However, if the concentration turns out to be greater than intended, investigate specifically whether mandate limits or the allocation need to be adjusted.
The value of a portfolio-wide AI stress test lies in that overall picture: where do different investments rely on the same expectations, and how much protection remains if those expectations fall short?