NextFin News - PGIM, the $1.4 trillion asset management arm of Prudential Financial, is urging limits on artificial-intelligence exposure inside collateralized loan obligations, arguing that unchecked concentration in AI-linked borrowers risks turning the CLO market's signature diversification into a single, correlated bet. The stance, flagged on September 29, 2026, marks a turning point for one of the largest managers in the roughly $1.4 trillion CLO market - a firm with about $77 billion of CLO assets under management, according to industry data from early 2025 - which now treats AI not as a growth theme but as a credit-underwriting problem that needs hard caps.
The push lands while software and AI-adjacent loans still make up the single-largest subsector in U.S. CLO collateral, and while the credit stress that began with the January-February software rout is still working through leveraged-loan portfolios. The question PGIM is forcing the industry to answer: when an entire theme can be disrupted by one technological shift, does owning 200 different issuers actually protect investors at all?
The Exposure PGIM Wants Capped
CLOs are structured to pool idiosyncratic risk. A typical deal holds senior secured loans to 200 or more borrowers, sliced into tranches that absorb losses in order, with overcollateralization tests designed to catch deterioration early. The pitch to investors has always been diversification - the idea that no single borrower can break the structure. That logic held through the pandemic, through the 2022 rate shock, and through the regional-bank turmoil of 2023, and it is why CLOs grew into a $1.4 trillion asset class within the broader $13.3 trillion structured-credit market.
AI disruption does not arrive as a single borrower. It arrives as a sector-wide repricing of business models, and by several measures the CLO market has already accumulated a meaningful position in it. Software alone accounts for roughly 12% of CLO holdings and about 15% of collateral in outstanding syndicated U.S. CLO deals, according to a February estimate from Morgan Stanley, making it the largest subsector by concentration. Moody's Ratings, in a February sector comment, put software at about 10% of assets in the U.S. CLOs it rates and 6% in European deals, while warning that legal business services and data-and-analytics providers sit among the most vulnerable sub-segments.
PGIM's own research, published in March, put the number differently but reached the same destination. Using data from Intex and Markit as of February 2026, PGIM estimated that approximately 11% of U.S. CLO portfolios and 7% of European portfolios are exposed to near-term AI disruption - with "meaningful levels of dispersion across managers." In other words, the average masks winners and losers: some managers built concentrated AI exposure quietly, inside a structure sold on the promise that concentration was impossible.
The stress is already visible in the tranche stack. Average levels on BB market-value overcollateralization tests fell about 1.2 percentage points to 104%, leaving only a four-point cushion before full repayment of BB tranches becomes impaired. The share of BB tranches with cushions below 3% has doubled this year to roughly 23%, and about 13% of BB tranches now show negative cushions - a level that implies potential principal impairment if collateral values deteriorate further. CLO equity valuations have been driven near post-pandemic troughs by the decline in software loan prices.
The timing matters. The software rout of January and February 2026 was triggered by the release of new AI tools that raised fears of widespread displacement across technology and professional services. Loan prices fell, CLO managers began exploring ways to reduce exposure, and the selling fed more selling. By March, spreads on CLOs had widened as investors weighed the possibility of a broader private-credit downturn in an industry estimated at $1.8 trillion.
Why Diversification Breaks When the Risk Is a Single Idea
The mechanism behind PGIM's concern is straightforward, and it is why a simple industry cap may not be enough. CLO concentration tests are written around industry classifications - how much of the pool sits in software, or healthcare, or business services. But AI disruption cuts across those categories. A data-and-analytics firm, an enterprise-software vendor, and a technology-enabled services company can all see their revenue models compress from the same shock, even though they sit in different industry buckets. A portfolio can pass every sector cap and still be entirely exposed to one idea.
PGIM's March analysis illustrated the point with two managers that had similar headline software weights but very different outcomes. One, underweight software relative to the market, still suffered outsized price moves because of idiosyncratic exposure to specific underperforming software issuers. The other, with more concentrated software exposure, was more sensitive to sector-wide price declines. The lesson: sector weight alone does not measure AI risk. Issuer-level exposure, underwriting quality, and portfolio construction do.
There is also a maturity dimension. PGIM noted that approximately half of the software exposure owned by several of the managers in its analysis will mature in the near term, creating a refinancing gauntlet at the very moment lenders are becoming more selective. A borrower that could roll debt easily 18 months ago may now face higher coupons, tighter covenants, or no market at all - and in a floating-rate structure, that stress transmits directly into the collateral pool's cash flows. Near-term maturities are the transmission channel through which a valuation shock becomes a cash-flow shock.
This is the core of the structural argument. The AI repricing is not a cyclical drawdown that mean-reverts when sentiment improves. It is a change in how entire categories of software and services revenue are valued - subscription seats replaced by AI agents, perpetual-license economics eroded, professional-services margins compressed. Public software multiples have already fallen more than one standard deviation from their eight-year average, according to PitchBook analysis, and private-market software valuations have followed. A concentration limit is an attempt to prevent a structural shock from being amplified by a structure designed for idiosyncratic ones.
The historical analog is instructive. In the early 2020s, commercial real estate - specifically office - built up as a correlated theme inside regional-bank balance sheets under the label of "real estate loans," diversified across borrowers and geographies. When remote work proved durable, the diversification proved illusory: the risk was not the borrower, it was the theme. CLOs now face a version of the same problem, except the theme is technology and the repricing is happening in months rather than years.
The Cyclical Layer: Selling Pressure and the Private-Credit Spillover
Not everything about the episode is structural. The market reaction has a distinctly cyclical texture. CLO spreads widened in March as fears of a broader private-credit downturn spooked investors, and several CLO managers began exploring ways to reduce software exposure.
Software is a sector where there is more selling coming from CLO managers than there is buying right now.
That was Jim Egan, co-head of securitized products research at Morgan Stanley, describing the dynamic. The imbalance is self-reinforcing in the short run. When multiple managers try to exit the same names, loan prices fall, overcollateralization tests tighten, and the structure itself forces further deleveraging. This is the cyclical leg of the problem: positioning, liquidity, and mark-to-market mechanics that can overshoot the underlying fundamental damage. It is also the leg most likely to reverse if software credits prove more resilient than feared and refinancing markets reopen.
But the cyclical overshoot does not invalidate the structural concern - it compounds it. Managers selling into a falling market crystallize losses for mezzanine and equity tranche holders, and the resulting caution raises the cost of capital for every AI-exposed borrower, including those with viable paths forward. The risk is that the industry's response is both too late to avoid the first wave of damage and too blunt to distinguish resilient borrowers from doomed ones.
There is also a distributional asymmetry worth noting. The pain from this dynamic is not shared evenly down the capital structure. Senior AAA and AA tranches are insulated by overcollateralization cushions and payment waterfalls; the first losses land on equity, then mezzanine. That is why PGIM's research found the price pressure concentrated in equity and mezzanine tranches while senior tranches remained stable. A concentration limit, then, is as much a defense of the senior tranche as it is a view on AI.
The Second-Order Consequence: Where the AI Risk Goes Next
The first-order effect of PGIM's push is obvious: less AI exposure inside CLO collateral. The second-order effects are more consequential, and they are not fully priced.
First, the CLO market could fragment along manager lines. Managers that kept AI exposure light gain a due-diligence narrative and likely cheaper funding; managers that built concentrated books face higher scrutiny and wider spreads on their new issues. Dispersion across managers, already elevated, becomes a permanent feature rather than a temporary dislocation. PGIM's own ETF footprint illustrates how much is at stake: its flagship PGIM AAA CLO ETF held roughly $13.45 billion in assets as of late September 2026, and the firm's aggregate-duration CLO vehicle, launched in June, adds another layer of publicly priced exposure that will mark the theme daily.
Second, and more important, AI-linked lending does not disappear - it migrates. Borrowers shut out of the CLO market can turn to direct-lending funds and other private-credit vehicles, where disclosure is thinner, concentration tests are less standardized, and the investor base is narrower. The risk does not leave the financial system; it moves to a part of the system with fewer guardrails and less price transparency. A CLO investor who buys an AAA tranche today can see the collateral list. A limited partner in a private-credit fund often cannot.
Third, the cost of capital for AI-exposed companies rises precisely when many of them need capital most - to fund the very AI transition that is disrupting their customers. That is a real-economy consequence of a portfolio-construction debate, and it is one that a concentration limit, however sensible for CLO holders, does nothing to soften.
And fourth, the definition of "AI exposure" itself is a moving target. Today's limit is written around software and business services. But the AI buildout is already migrating into data centers, electric utilities, and semiconductor supply chains - capital-intensive borrowers that CLOs are only beginning to underwrite at scale. A limit calibrated to 2026's exposed sectors may miss 2027's. That is the argument for risk-factor-based limits over sector labels, and it is where the industry is likely to end up even if it resists a blunt AI cap today.
The Strongest Case Against Hard Caps
The counter-argument is not weak, and it deserves weight. CLO structures already contain safeguards: industry concentration caps that typically run in the 10% to 15% range, overcollateralization and interest-coverage tests that trigger before senior tranches are touched, and active managers who can trade out of exposed credits. Moody's has argued that CLO exposure to genuinely at-risk issuers is lower than software's headline concentration implies, because the damage depends on issuer type - a hyperscaler's supplier is not the same credit as a legacy enterprise-software vendor.
A blunt AI cap, on this view, risks forcing indiscriminate selling at distressed prices, locking in losses for the mezzanine and equity tranches that can least absorb them, and pushing managers to chase yield in equally correlated but less obviously labeled corners of the loan market - data centers, utilities, semiconductor suppliers - where the next concentration problem is already building. The irony is sharp: a rule meant to reduce concentration could increase it, simply by relabeling where the risk sits.
The answer to that argument is that the existing safeguards are backward-looking. They test concentration by the industry labels of 2020, not by the correlated risk factors of 2026. A loan book can pass every industry cap while being entirely exposed to one idea. PGIM's point is not that safeguards should be discarded; it is that the taxonomy they measure has become obsolete. The forced-selling concern is legitimate, but it is a symptom of having reacted late, not a reason to avoid reacting at all.
What Comes Next
In the short term, expect more volatility in mezzanine and equity tranches as managers reposition, and for CLO spreads to stay wide while the refinancing gauntlet for software borrowers plays out. Over the medium term, AI-linked lending is likely to drift toward private-credit channels, and CLO issuance may slow as underwriters recalibrate what counts as acceptable concentration. Over the long term, the most durable outcome is a rewrite of the rulebook: risk-factor-based concentration limits - measuring correlated exposure, not just industry labels - are likely to become standard in new CLO documentation.
The base case is gradual adoption: large managers introduce AI or technology-theme caps over the next 12 to 18 months, and AI exposure in U.S. CLO collateral drifts down from roughly 12% toward the high single digits. The upside case is that AI disruption proves overblown, software borrowers refinance successfully, and BB market-value overcollateralization cushions recover above 107% - in which case the episode reads as a cyclical panic, not a structural fix. The downside case is a wave of software downgrades that triggers overcollateralization breaches and forced deleveraging, spreading stress from BB tranches into BBB.
One signal would falsify the structural thesis cleanly: if software's share of U.S. CLO collateral falls below 8% while the average BB market-value overcollateralization cushion recovers above 107% for two consecutive quarters, the problem was cyclical positioning, not a regime shift in credit risk. Until then, PGIM's push is best read as an admission about the CLO market itself - that its oldest promise, that diversification neutralizes risk, breaks the moment the risk is a single idea that everyone owns.
Explore more exclusive insights at nextfin.ai.
