Where does the software lending actually sit?
In application software, by a wide margin. Apollo's chief economist, Torsten Slok, published a cut of outstanding software loans by subsector on 22 August 2026, using PitchBook data as of 28 July 2026. Application software carries 146 billion dollars of loans outstanding. Cybersecurity and infrastructure-and-data together carry roughly a third of that.
Read from the chart, cybersecurity borrowers owe about 29 billion dollars and infrastructure and data borrowers about 21 billion. Application software is therefore roughly three quarters of the software loan book that Apollo counts, and the other two subsectors split the remaining quarter between them.
The rating mix is the second half of the picture. Apollo's text says most of the 146 billion is rated B- or lower. The chart bears that out: roughly 67 billion is rated B-, a further 16 billion or so is CCC+ or lower, and about 5 billion carries no rating, against about 59 billion rated B or higher. A borrower in application software is, on the median dollar, one notch above the level at which a single downgrade moves a loan into the tier that collateralized loan obligations are limited in holding.
The two smaller subsectors are better rated as well as smaller. On the chart, a little under half of the cybersecurity book and about half of the infrastructure-and-data book is rated B or higher, and the CCC-or-lower slice in each is a few billion dollars. Concentration and quality point the same way: the most exposed dollars are also the weakest rated.
| Subsector | Total outstanding | B or higher | B- | CCC+ or lower | Not rated |
|---|---|---|---|---|---|
| Application software | $146bn | ~$59bn | ~$67bn | ~$16bn | ~$5bn |
| Cybersecurity | ~$29bn | ~$11bn | ~$13bn | ~$5bn | nil |
| Infrastructure and data | ~$21bn | ~$11bn | ~$7bn | ~$4bn | nil |
Why is the exposure concentrated in that layer?
Because that is where the leveraged buyouts were. Application software has been the sponsor sector of choice for a decade: recurring revenue, high gross margins, low capital intensity, and a customer base that renews. Those are the qualities a lender underwrites, so the loans followed the deals, and the deals were concentrated in companies that sell seats.
Apollo's argument is that the pricing model is now the exposure. Application software companies charge by how many people use the product to run tasks. Those are the tasks that AI can most plausibly automate, so the seat count is the line in the revenue model that an AI deployment attacks first. Infrastructure, data and security providers are paid more as AI adds computing workloads and more systems to defend, so the same shift that threatens one subsector's revenue is a tailwind for the other two.
The lending market has already begun to sort on this. PitchBook LCD data as at 30 June 2026 has software falling to 8.6 percent of broadly syndicated loan issuance year to date, from 17.6 percent in 2025. New money is arriving more slowly; the stock of loans in the chart is what was written in the years before the sorting began.
That is the point for a borrower. The 146 billion is not a forecast. It is a book of loans already made, at leverage set on the old revenue model, held by lenders who now read every application software credit through the seat-count question whether or not the borrower has raised it.
“Sector concentration in a lender's book is a borrower's problem before it is the lender's. When most of a lender's software exposure is one subsector at one rating, every refinancing in that subsector is priced against the whole cohort, not against the company in front of them.”
What does it mean for an owner refinancing in that cohort?
That the lender's first question is about the product, not the numbers. A lender with a large B- application software book is managing a concentration, and the tool for managing a concentration is to add less of it and to charge more for what is added. An owner arriving to refinance a seat-priced business should expect a margin that reflects the cohort, a tighter leverage test, and less appetite for payment-in-kind flexibility than the same lender showed in 2023.
It changes what evidence carries weight. Net revenue retention by customer, the share of revenue that is consumption or outcome priced rather than seat priced, and the seat count trend inside the largest accounts are the figures a credit committee now wants ahead of the EBITDA bridge. A borrower who can show that usage per customer is rising while seat counts are flat has an answer to the question the lender is asking. A borrower who cannot is priced with the bar on the left of the chart.
For a business on the other side of the line, the reading reverses. An infrastructure, data or security borrower is in a subsector with a fraction of the exposure and a better rating mix, and a lender under-weight in that subsector will compete for the credit. That is the moment to test the market rather than roll with the incumbent, and to negotiate covenant headroom and reinvestment capacity rather than only price.
For a buyer, the chart is a diligence prompt. The debt package on an application software target was sized on the seat model, and the lender who wrote it will re-underwrite on the new one at the first amendment. A purchase price that assumes the existing facility rolls at existing terms is assuming something the lender has already stopped assuming.
What would change the reading?
A shift in the rating mix inside the application software bar. If the B-or-higher slice grows while the B- and CCC slices shrink, the cohort is being cleaned up through repayment, upgrade or sale rather than through default, and the pricing premium on the subsector should narrow. If the CCC slice grows, the reverse.
The second is the total. The 146 billion dollar figure falling while new issuance is subdued would mean the book is amortizing and refinancing away from the sector, which is what the issuance share suggests is under way. The figure rising would mean lenders have decided the seat-count risk is priced and are adding again.
The third is the revenue model of the borrowers themselves. Every application software company that moves a material share of revenue from seats to usage or outcomes moves itself, in a lender's eyes, toward the right-hand side of the chart. That is a change a borrower can make ahead of the next refinancing, and it is the one most within the owner's control.
Apollo's note is one chart and four sentences. The sentence that matters is the last: the same AI shift that hits application software revenue pays infrastructure, data and security providers more. Lenders have most of their money on the first side of that sentence, and every borrower in the sector is now being read against it.
As of August 2026
Sources: Apollo Global Management, The Daily Spark by Torsten Slok, Software Lending Is Concentrated Where AI Hits Hardest, published 22 August 2026, read from the page and the chart image, chart sourced there to PitchBook and Apollo Chief Economist with data as of 28 July 2026, for 146 billion dollars of application software loans outstanding described as mostly rated B- or lower; for the statement that application software is the part of software most exposed to AI disruption and the part where lenders have the most money at risk; for the characterization of application software companies as charging by how many people use the software to run tasks AI can most plausibly automate; and for the statement that infrastructure, data and security providers are paid more as AI adds computing workloads and more systems to defend. The subsector totals for cybersecurity and infrastructure and data, and the rating split within each of the three bars, are our readings of the note's stacked bar chart, rounded to the nearest billion and approximate. PitchBook LCD, as at 30 June 2026, for software falling to 8.6 percent of broadly syndicated loan issuance year to date from 17.6 percent in 2025, as already cited in a companion article on this site. The description of the B- rating as one notch above the tier that collateralized loan obligations are limited in holding refers to the customary CCC basket in those vehicles and is general market knowledge, not a figure from either source. The shares, the roughly 88 billion dollar B-or-lower reading and the roughly three quarters concentration are our own arithmetic. Companion articles on this site cover the compression in public software multiples, whether private credit and the syndicated market are one market or two, and the CLO issuance window.