Norges Bank’s AI risk outlook
The $2 trillion sovereign wealth fund finds no safe havens against systemic risks such as an AI-led market crash.
The idea that the AI industry constitutes a risk to investors is well known. The story goes that precarious, opaque and circular transactions exist among involving LLM modelers like OpenAI, its Big Tech backers, data center and energy buildouts, off-balance sheet lending from private capital, and private capital’s own borrowing from banks. And the AI labs sorely lack the revenues to justify all of this, so poof go the securities markets.
(There’s also the end-of-the-world theme, but in that case, an equities wipeout won’t matter.)
How do institutional investors address such a market risk? Is the risk itself so hyped that it’s “been priced into the market”? How do we think about protecting portfolios?
This is not just a question for individuals or rich individuals. Financial centers earn that title by recycling the investments at scale of the largest institutions, including sovereign wealth funds, pension funds, and insurance companies.
These vast real-money investors must allocate to the biggest, most liquid asset classes, as well as to alternatives. They have benefited greatly over the past three years from the terrific performance posted by AI-related stocks (the Mag Seven or Eight US names), and these have become the biggest drivers of performance in US equities, and to some extent, global equities.
Mag-Seven explain more than half the S&P 500’s gains over the last yree years, and today they represent about one-third of its market cap. That is incredible concentration that shows the extent to which equity performance relies on a handful of Big Tech hyperscalers.
But when the music stops, could we be facing a market like October 2008? Or a bubble collapse like 2001, or Japan in 1998? Value investors and perma-bears have been saying yes (of course) for a while now, but their concerns are no longer fringe.
Norges Bank Investment Management, the $2 trillion manager of Norway’s oil receipts, released its annual review of investment risk, for the second year running. AI-related tail risk is at the heart of its stress-testing framework.
NBIM treats AI as a systemic risk, not an isolated tech theme or a company-specific story. It can drive broad equity crashes, interact with geopolitics and fiscal stress, and overwhelm traditional diversification measures.
AI is one of four system risks NBIM has identified, along with geopolitical fragmentation, a debt crisis, and a climate crisis. Individual calls on stocks or bonds or alternatives don’t warrant this level of attention. We live in a world dominated by flows, not fundamentals.
It’s notable, however, that NBIM has changed its tune on the AI story. In 2024, it focused on the risk that over-optimistic AI valuations deflate as earnings disappoint, with losses concentrated in AI-heavy segments of the equity market and a classic “flight to safety” into government bonds that partly offsets the damage.
By the end of 2025, though, NBIM has observed that AI-related market concentration and capex have risen further, so a correction would now mean a broad equity market crash, with larger losses in stocks overall, and fixed income providing less relief.
Even this analysis is moderate in its fears, as it doesn’t account for the second- and third-level impacts on private capital and banks.
Nonetheless, NBIIM’s approach to stress testing is increasingly systemic. It outlines many possible outcomes and narrows these to a few scenarios that combine high severity and a probability of occurring (not a prediction, but not a negligible chance). NBIM applies past episodes (the Asian financial crisis of 1997, the dot.com bust, GFC, Covid, the 2022 interest-rate shock) to the current holdings of assets and reports simulated NAV impacts. In most cases, such as the dot.com crash or GFC, equity losses dominated, with fixed income providing modest losses, or modest upsides.
But these four new risk scenarios don’t behave the same way. The outcomes don’t match previous crises. The drawdowns are deeper and more widespread.
The good news, if we look for some, is that investors are waking up to the fact that there aren’t safe havens. This raises a question that NBIM doesn’t appear to have addressed: in previous crises, the US dollar and Treasury market was the safe haven, even in crises of America’s making. Will that be the case next time? Probably, because that’s the only place with liquidity at this scale. But still, worth a ponder.
But the not-so-good news is that in a systemic crisis, diversification becomes diluted. Diversification has always been the “free lunch” for investors: balancing portfolios among asset classes has always been a way to reduce volatility and other measurements of risk.
NBIM notes that a 70/30 equity‑bond mix can deliver long stretches of robust real returns but also multi‑year periods of deeply negative performance, underscoring that diversification across public equities and fixed income is no guarantee against large drawdowns.
What’s an asset owner to do?
Institutions must keep leverage (including securities lending) and illiquid exposure within tight bounds; hold onto a lot of cash or short-term government bonds; and avoid too much concentration in obvious danger points (mega-cap AI names, but also single counterparties or specific sovereigns).
But even this is tricky, because NBIM can’t predict which crisis it should try to mitigate. They are all existential but in different ways. (It also realizes that one systemic crisis could bleed into a second, in which case...yeah, bad.)
A debt crisis would harm equities and long-duration bonds, but may be limited to just certain markets. A fragmented-world scenario (economies retrench into blocs) would hurt stocks in both developed and emerging markets. The AI scenario hurts stocks but could be mitigated by a rally in bonds, but this requires tricky timing around duration. And extreme weather events could hit all risk assets, including alternatives, via inflationary food shocks. Meanwhile, keep too much in cash, and watch inflation eat your portfolio.
Put another way, NBIM sees that building exposures to one asset class sets it up for different risks. Should it lean in into equities and real assets? Yes if it fears a debt crisis, but then it’s vulnerable to an AI crash or geopolitical fragmentation and climate shocks.
So should it emphasize duration and sovereign bonds, and give up some upside? Not if a debt crisis strikes.
Go for real estate and infrastructure? NBIM does suggest real estate and infrastructure may retain some of their defensive characteristics. Yet bad things can happen in both interest-rate and growth shocks, and these assets can’t be tapped quickly if NBIM had to cover drawdowns in the liquid parts of its portfolio, meaning it would accelerate a downward spiral of asset prices. Alternatives are helpful so long as investors avoid concentrating too much in a handful of private-capital managers and have a sensible spread across geographies.
At least big investors have a realistic idea about what can go wrong, and know that all asset classes are vulnerable. The usual tools of volatility and tracking error are still important to risk management, but expected shortfalls may prove a more realistic way to understand what might happen.
In other words, risk management for large, real-money investors is now about survival. That’s it.