Marks vs. Prices

• Serving each $1 of Anthropic's revenue cost about $0.31 this year, against $1.95 in 2024

  Technology and the capital cycle, October 2026

  On the numbers in Anthropic's IPO filing, serving each $1 of revenue cost roughly $0.31 in the first half of 2026 once its payments to partners are added, and $1.95 in 2024 before them. Revenue grew 14x while the compute to serve it grew about 2.3x. The open question is whether demand keeps pace with the $265.6BN of compute it has signed up for.

• Split a contractor's data center backlog by who pays for the build

  Technology and the capital cycle, October 2026

  Meta's third quarter capex ran $2.12BN past its operating cash flow, but it holds $82.45BN of cash and securities, while CoreWeave's unsecured debt yields about 15%. Both can sit in the backlog of a lower middle market contractor. Orders funded with high yield debt belong on their own line, at a lower multiple.

• Industrial borrowers default at 9.9% in private credit while software sits at 0.6%

  Private credit and lending, October 2026

  In Fitch's trailing 12-month private credit default rates through August, industrial and manufacturing borrowers nearly doubled to 9.9% while software sat at 0.6%. Most counted defaults are PIK switches, deferrals and stressed extensions, and a PIK option written at origination is not a default. For buyers of manufacturers, a PIK toggle is worth pricing at close.

• A 10% coupon on Anthropic junior debt is no benchmark for a small company's sub debt

  Private credit and lending, October 2026

  Blackstone funds were offered a 10% coupon on $9BN of Anthropic junior debt, the FT reported. A portfolio company under $20MM of EBITDA pays 13% to 15% all-in for new sub debt, and smaller borrowers default more. Lenders price the borrower and the structure, so benchmark junior debt to your size, not to the headline.

• Bank CEOs say sponsors are stuck on marks; price each 2021 deal against today's bid

  LPs, returns and liquidity, October 2026

  In the same October week, the CEOs of Morgan Stanley and Goldman Sachs described sponsors holding assets above what the market will pay. In an April poll cited by Bain, most LPs said they start to lose confidence in a GP when a full exit prices more than 5% below the last mark. Take the markdown in the quarterly letter, not in the exit memo.

• Your lender's rate curve can cost 0.37x of EBITDA in debt, even with a flat spread

  Private credit and lending, October 2026

  A third quarter pricing matrix left unitranche spreads flat while forward yields rose 82 bps. Under a 2.0x coverage test, sizing on the forward curve allows about 4.92x EBITDA and sizing on spot SOFR about 5.29x, a gap worth $3.7MM of debt on a $10MM EBITDA company. Ask which curve your lender sizes on before you negotiate the spread.

• Reuters puts OpenAI at almost half of Oracle's $638BN backlog, a share no filing gives

  Technology and the capital cycle, October 2026

  Oracle's 10-K says no customer reached 10% of revenue, but Reuters reports OpenAI at almost half of its $638BN backlog, close to $300BN. Banks quoted $18BN of Project Jupiter loans at 89 to 91 cents as the syndication stalled. Data center lenders cannot size that concentration from Oracle's filings.

• In a continuation vehicle the GP buys and most LPs sell, at the same price

  LPs, returns and liquidity, October 2026

  GP-led deals, mostly continuation vehicles, reached $115BN last year, and the average LP rolls only about 15%. The GP is expected to roll its accrued carry, so the investor closest to the company buys while the investors furthest from it sell. Before an election deadline, ask whether you would buy the company at this price today.

• Investors asked a Blue Owl fund for $1.1BN back and will get $135MM this round

  LPs, returns and liquidity, October 2026

  Holders of an estimated 39% of Blue Owl Technology Income Corp.'s shares asked to redeem, and the fund will buy back 5%, pro rata. Borrowing to pay everyone would take its debt past the legal limit for the fund. Once this round is paid, first quarter requests will be about 35% filled, and at that pace half takes about five quarters.

• Meta's Hyperion guarantee: a $46BN disclosed worst case, recorded at zero

  Technology and the capital cycle, August 2026

  Beignet Investor sold $27.3BN of bonds for the Blue Owl side of Meta's Hyperion data center campus, while committed rent runs four years from 2029. Meta's residual value guarantee is capped near $28BN and carried at zero. Above the cap and after year sixteen, losses fall first on the venture's owners, where the funds hold 80%.

• Delaware Life's restated related party book is about 4.5x its capital

  Private credit and lending, August 2026

  Delaware Life restated its related party investments from $1.4BN to at least $17BN, at least 39% of invested assets. The insurer reports $3.8BN of capital and surplus, so each point of discount on that book costs 4.5% of its capital. That arithmetic, more than the $12.5BN Lakers sale, sets the exits.

• Nvidia's $500BN financing plan is credit backed by the seller's own number

  Technology and the capital cycle, August 2026

  Six institutions signed memorandums of understanding to mobilize over $500BN of third party capital, lent to Nvidia's customers to buy Nvidia chips. The collateral is the residual value of used GPUs, and the seller stands behind that value. How far lenders haircut it is the trade.

• Accell Group's €1.56BN buyout met four market tests, and each answer was worse

  Private credit and lending, August 2026

  A KKR-led consortium took Accell Group private at €1.56BN in 2022, paying €58 a share. A recapitalization, a lender takeover, a failed sale process and a suspension of payments followed, and outside money declined the last two tests. Most buyouts of that era have met no market at all since the day they were bought.

• The exit window is open for durable assets, not for 2021 software held at 2021 marks

  LPs, returns and liquidity, August 2026

  Goldman's July outlook has global M&A up 48% in the first half, yet roughly 16,000 buyout companies have been held more than four years. The 2021 vintage has sold 19% of its investments by year four, against 30% historically. Buyers are sorting assets, and waiting does not move a manager up the queue.

• 13 buyouts over $10BN made $274BN of the $276BN gain in 2025 deal value

  Technology and the capital cycle, July 2026

  Global buyout value reached $904BN in 2025, up 44%, while deal count fell 6% to 3,018. Bain attributes $274BN of the gain to the buyouts over $10BN, with data centers the most visible share. A GP holding an ordinary company should not price its exit to that market.

• A fund up 439% was margin called four trading days after its investor letter

  Technology and the capital cycle, July 2026

  Situational Awareness reported 439% net for the first half, and four trading days later its prime brokers were selling the borrowed slice of its stock book. Its private stakes, which no lender could margin call, came through intact. The financing broke, not the forecast.

• Private credit's low correlation was partly a marking effect, and the shock found the daily price

  Private credit and lending, July 2026

  From their 2025 highs to early 2026, shares of five listed managers fell 41% to 67%. Withdrawal requests at three semi-liquid credit funds ran 7.9% to 13.9% in one quarter, and the preset caps became the gates. Redemption queues, not marks, are where the stress now shows.

• 2025 buyouts paid the 2021 record price, with more equity and the same return math

  LPs, returns and liquidity, July 2026

  The median US buyout in 2025 cleared at 12.5x EBITDA, tying the 2021 record. Sponsors now write half or more of the purchase price, up from about 35%, so a bad deal shows up as a weak return rather than a covenant breach. The price, which the return is computed against, has not come down.

Earlier posts, March to July 2026, are on LinkedIn.