Goldman Sachs Says the AI Buildout Is Erasing the Lines Between Public and Private Capital Markets

What's New

Scott Goodwin, co-founder of Diameter Capital Partners, argued that the private credit market is systematically mispricing recovery rates and that the BDC redemption cycle now underway may represent retail investors acting as the smart money for the first time in a conversation at the Bloomberg Global Credit Forum. Goodwin disclosed that his firm has identified mark discrepancies as wide as 40 to 50 points on the same loan across public BDCs, with average mark differences expanding from roughly 1% pre-tightening to approximately 6% today. He argued that software loan recoveries will land at 10 to 30 cents rather than the 60 to 80 cents the industry assumes, based on the track record of BSL tech loan defaults over the past five to six years.

Why It Matters

The conventional recovery assumption in private credit is anchored to senior secured lending to asset-heavy businesses. Software companies do not fit that template. There are no hard assets to recover when the business fails. The leveraged loan market has already revealed this: tech loan recoveries have averaged 10 to 30 cents, not 60 to 80. If a third of the $1.5 trillion direct lending market is SAS-exposed and recovery rates prove to be half or less of what models assume, loss-given-default calculations across the asset class will need to be rewritten. Meanwhile, banks providing leverage to BDC vehicles are already capping SAS exposure at 20 to 25%, which will force de-risking and secondary sales of the highest-quality software loans.

Big Picture Drivers

• The prisoner's dilemma driving BDC redemptions is rational, not panicked: Goodwin frames the redemption cycle in game theory terms. If other investors redeem and you do not, you are left with the lower-quality residual portfolio. Public BDCs already trade at 85 to 95% of gross asset value, or roughly 70 to 80% of NAV. The public market is pricing a problem that private marks have not yet acknowledged.

• Mark discrepancies as an industry-wide credibility deficit: Goodwin's firm found one BDC marking a loan at par while another marked the same loan at 60. Diameter shorted the overmarked position and profited when the mark was eventually corrected. With average mark differences at 6% and rising, regulators and bank leverage providers are both increasing scrutiny.

• Recovery rate assumptions need recalibration for software-heavy portfolios: BSL tech loan recoveries over the past five to six years have landed at 10 to 30 cents. Software companies lack the hard assets that produce 60 to 80 cent recoveries in traditional industries. Direct lending managers who built their loss models on historic senior secured recovery rates will face a mismatch between assumptions and outcomes.

• Sponsors facing a binary choice that degrades lender recovery: A sponsor holding a software company with a 2031 maturity under AI threat can invest all cash flow in becoming an agentic company or extract dividends. The second option effectively converts the loan into an interest-only instrument, stripping the lender of the delevering trajectory that underpins recovery assumptions.

• Forced selling from BDC leverage providers will create a secondary opportunity: Banks are capping SAS exposure in the BDC vehicles they finance. Managers exceeding those caps will need to sell, and they will not sell the loans already trading at 50 because that does not improve their ratios. They will sell their highest-quality software loans at a discount to par. Goodwin is positioning to buy those.

By The Numbers

• 10 to 30 cents — Actual recovery rates on BSL tech loan defaults over the past five to six years, versus the 60 to 80 cents the private credit industry assumes

• 40 to 50 points — Widest mark discrepancy Goodwin's team identified on the same loan across different public BDCs

• ~6% — Current average mark difference for the same loan across BDCs, up from approximately 1% before the Fed tightening cycle

• 0.85 to 0.95 — Range at which public BDCs trade relative to gross asset value, implying 70 to 80% of NAV

• 5 to 10% — Default rate (hard defaults plus LME) expected by the Bloomberg audience for leveraged loan software borrowers by year-end 2027

• 20 to 25% — SAS concentration cap being imposed by banks providing leverage to BDC vehicles

Key Trends to Watch

• Regulatory focus on daily marking and mark convergence: Goodwin noted that Marc Rowan of Apollo and others are publicly advocating for daily marking in private credit. As regulators and leverage providers both push for consistency, the gap between reported NAVs and realizable values will narrow, potentially triggering a mark-down cycle across BDC portfolios.

• Software distress timeline tied to maturity walls in 2028 to 2030: Goodwin noted that LME typically happens two years before maturity. With software maturities clustering in 2028, 2029, and 2030, the acceleration in LME and hard defaults is beginning now and will build through 2027.

• AI data center financing approaching a saturation tipping point by mid-2028: Goodwin argued that hyperscaler debt issuance will eventually crowd out government and other investment grade issuance. The market is absorbing supply now, but the "bathtub is filling up" and differentiation on documentation and construction risk is declining as the herd chases the best excess spread opportunity that is not distressed.

• Regulatory and political risk to AI deployment underpriced: Goodwin flagged that AI companies are already adjusting their public messaging on jobs in response to low approval ratings. NIMBYism around data center construction in locations like Northern Virginia represents a real constraint. A regulatory intervention in AI rollout would directly impair the financing assumptions underlying the current buildout.

Memorable Quotes

• "A lot of people who've been in private credit for a long time say, 'Oh, senior secured loan, the recovery will be 60, 70, 80 cents.' If you look at BSLs and tech loans the past five or six years, what have the recoveries been? 10, 20, 30 cents." Goodwin on the gap between industry muscle memory and actual outcomes.

• "We found one guy who had it marked at 60 and another had the same one marked at par. So we shorted the one that was marked at par and then eventually they had to mark it at 60 and it went down." Goodwin on the arbitrage opportunity created by mark inconsistency across BDCs.

• "If you have a 2031 maturity and you're a sponsor and you think your business is really under threat of AI, you have two choices. Either take all the cash flow and invest in becoming an agentic company or take dividends." Goodwin on how sponsor incentives can convert performing loans into interest-only instruments.

• "This is going to be much bigger than either of those cycles." Goodwin comparing the AI-driven winner-loser dynamic to the internet revolution and the fracking cycle.

The Wrap

Goodwin's argument is not about whether private credit works. It is about whether the industry's loss assumptions are calibrated to the portfolio it actually holds. When a third of direct lending is exposed to software and adjacent sectors where recovery rates have historically been 10 to 30 cents, the 60-to-80-cent assumption embedded in most LP models is not conservative. It is wrong. The forced selling dynamic Goodwin describes, where bank leverage providers cap SAS exposure and BDC managers sell their best software paper at a discount, is not a distressed opportunity yet. It is an early-stage liquidity event that Diameter is positioning for with selective secondary purchases. The deeper distressed opportunity arrives later, when the maturities come due and the sponsors who chose dividends over reinvestment discover they have converted their loans into interest-only instruments backed by businesses that AI has begun to displace. That timeline runs through 2028 to 2031. Goodwin is not waiting for it passively. He is trading the dispersion now and building the positioning for what comes next.