Golub Capital invests $5m in AI credit developer F2 and deploys it across underwriting

Golub Capital has put $5m into F2, an AI credit developer whose software will run across the manager's screening, diligence, underwriting and monitoring, and David Golub will chair a new invitation-only AI council that comes with the arrangement. The release names the four functions and the council, then omits the budget, headcount, timeline, credit count and displaced work.

The money is described as an investment rather than a software purchase, which implies a stake of some size in F2 itself, but the announcement gives no terms beyond the $5m, no ownership figure, and no indication of which part of the book gets looked at first.

Read the four functions as a workflow and the commitment sharpens. Screening, diligence and underwriting are origination functions applied to loans that have not been made yet; monitoring is the only one that looks backwards at credits already on the book — covenants, borrower reporting, the evidence behind a quarterly valuation — and it is the function a manager leans on when the book it already owns becomes the thing investors are asking about.

Where that software gets pointed is a question the market has spent the past few weeks answering in public.

9fin's second-quarter BDC watchlist flags 118 credits marked down five points or more, information technology accounting for just under 36% of the $4.19bn in affected fair value. That share is close to $1.5bn sitting in a single industry, and more than half of the 118 credits puts at least sixty on a maturity schedule that runs out by 2029.

A manager looking at broad credit deterioration would expect marks to spread across industries and maturities, while a manager looking at a cohort problem would expect them to bunch in one sector and one window of dates, and the watchlist shows the second pattern. The list is a floor rather than a total, since credits marked down by one to four points do not appear on it at all: $4.19bn is what clearing a five-point threshold caught, not what the book has lost.

The $4.19bn spread across 118 credits works out to roughly $35m a loan, a list of mid-sized positions rather than a handful of large ones, and the figures do not give information technology's share of the portfolios those credits sit in, so the concentration is documented within the markdowns first and only potentially within the underlying lending. Scope cuts the other way too: BDC marks are public because listed vehicles publish them every quarter, which makes the watchlist a clear map of the listed side of the market and silent on the rest.

The maturity profile is the more informative half of the disclosure, because more than half the flagged credits come due by 2029, a schedule consistent with paper originated in the low-rate years rather than anything written in the past twelve months. Broad deterioration argues for cutting exposure across a book, while a dated cohort argues for taking a defined set of loans back through diligence with a nearer maturity in view and deciding case by case which marks hold.

The second exercise is a project rather than a decision, and a repetitive, document-heavy one, which is the kind of work credit software is built to do. Whether Golub funded a developer and deployed its platform instead of staffing that exercise internally is inference rather than disclosure; what the announcement establishes is that the software goes live now, ahead of the third-quarter marks that will settle the cohort's condition either way.

A quarterly valuation refreshes a credit's numbers, while a re-underwrite goes further and revisits the reason the loan was made, asking whether the assumptions that justified it still survive contact with the borrower's current trading. A maturing cohort forces the harder exercise, because a maturity date converts a stale thesis into a refinancing decision, and the announcement does not say which of the two the platform performs.

There is a coincidence on the surface worth resisting: the industry carrying the largest share of the flagged marks is information technology, and the developer Golub backed is described as an AI credit developer, but the overlap is a matter of labels rather than evidence. 118 credits is not an unmanageable list for a team with a quarter to work through, which is why the case for tooling rests on repetition rather than volume; the same names run through the same tests every quarter, with a record of what moved between runs.

The gap between a mark and a credit event is the reason any of this gets done twice, and the week supplied a clean demonstration: Metrics Credit Partners, an Australian manager running around A$40bn, suspended redemptions in three ASX-listed funds after a dispute with KPMG over valuations and marked the vehicles down by as much as 12.16%. No borrower default is described behind the gate; what moved was the valuation rather than the loan, and investors could not transact while the disagreement ran.

The Bank of England's Financial Policy Committee used its September record to call private credit vulnerable as financing conditions tighten, pointing to elevated risk-taking and pressure on floating-rate borrowers' debt-servicing capacity, while its private-market stress test remains unpublished. That is the cash-flow version of the worry, landing on the cohort the watchlist identifies: credits maturing in 2029 will need refinancing into conditions the committee describes but has not sized.

Listed prices have been voting meanwhile: BDC shares fell again in a week when the wider market rose, and one week's share price cannot separate a repricing of funding costs across the whole book from a credit warning concentrated in one cohort. The third-quarter marks will separate them, and the credits held below 90 will do most of the separating, since a further step down in those names speaks to credit rather than to the cost of money.

The value of the tooling is easiest to state if the markdowns are a cohort problem with a 2029 boundary. The return on underwriting software lies less in writing better loans next year than in repricing the loans already made, and that work can be run between reporting dates so that a manager arrives at the third-quarter print holding its own view of which credits still support their marks.

The council is the governance half of the arrangement and the least specified part of it, because the investment comes with a new invitation-only AI council that David Golub will chair, and neither its membership nor its remit nor its meeting frequency is described. Attaching a named principal to a body whose conclusions bear on underwriting is a good deal of structure to hang on $5m, and it suggests the arrangement is meant to outlast a single deployment.

The next set of marks is the test: if the writedowns spread beyond information technology and beyond the 2029 maturities, the cohort reading fails and the broad-deterioration reading takes over, and neither outcome requires anyone to decide what the software is worth. Managers making comparable arrangements would matter as much, since one $5m investment is a single data point and a run of them would say the second look at the older vintages has moved from a choice into routine practice.