Marks to Market

Solvent, But Not on Schedule

The system can afford the losses. What it may not afford is the timeline — and the same compression that broke SaaS marks is already climbing toward the trillion-dollar IPOs.

David H. Friedel Jr./ 2026-08-14
Subscribe
MarketsAIMacroeconomics

The last piece established the envelope: roughly $1 trillion of marked private equity software value on top of $500–750 billion of debt, priced for a multiple regime that ended.1 The question readers keep asking is the right one.

Can investors actually take the hit?

Split it in two, because the answers differ.

Solvency: yes

The losses land on pensions, insurers, endowments, and sovereigns. Long-dated liabilities. No depositors. No overnight funding. A pension writing down its private equity sleeve is a funding-ratio problem and a governance headache — not a run.

Even the bear case at the credit layer — $50–75 billion of impairments — sits against $1.6 trillion of Tier 1 capital at the reporting banks.2 This is not leverage stacked on leverage inside the banking perimeter. The right analogy is not 2008. It is a solvent institution forced to sell illiquid assets on someone else's timetable.

Schedule: no

The workout scenario — losses metered out over five to eight years through continuation vehicles — depends on four things: rate cuts, a functioning consumer, patient LPs, and an open exit window.

The macro is closing all four at once.

Fourth-quarter GDP was revised down to 0.7% on weak consumer spending; the Supreme Court struck the original tariffs, and the administration replaced them under different authority within days; and the Middle East engagement has added a dimension already visible in the data.3 The effective tariff rate has gone from 2.1% to roughly 11.7%, pass-through to consumers now exceeds 50%, and the named risk is stagflation — a labor market saying cut while tariff prices say hold.4

A Fed pinned by tariff inflation cannot deliver the cuts that reopen exits. A 0.7% economy compresses the same ARR the marks depend on, independent of anything AI does. And if public equities fall while private marks stay sticky, the denominator effect forces allocation-capped LPs to sell fund stakes into a 73-cent market — which sets more comps, which is the spiral.

The environment doesn't change whether investors can take the losses. It changes the price of time. And time is the entire base case.

The Cost of Time: Software Value and the Capex Migration

The capex objection

Here is the strongest pushback: if software is compressing, why are the four hyperscalers spending $725 billion this year — up 77% from $410 billion — with analysts projecting a trillion in 2027 and Goldman modeling $7.6 trillion through 2031?5 Amazon's free cash flow is projected to go negative to fund it.6

Because it is not an objection. It is the other side of the same ledger.

The capex is not a bet that application software keeps its margins. It is a bet that value migrates down the stack — out of the application layer and into compute, power, and silicon. Every dollar of that $725 billion finances the capability that compresses the layer above it. The hyperscalers are not disagreeing with the SaaS repricing. They are funding it.

The spending and the write-downs are one transaction, recorded on two different balance sheets.

The floor above

Which raises the question nobody underwriting the trillion-dollar IPOs wants asked.

Anthropic raised at a $965 billion post-money valuation. OpenAI is expected to list. SpaceX trades around $1.4 trillion.7 The AI-lab valuations assume the model layer keeps durable pricing power — that frontier capability stays scarce and margins stay wide.

Meanwhile, the two releases that matter landed within the last three weeks, and they attack the moat from both ends.

At the capability end: Moonshot's Kimi K3 — 2.8 trillion parameters, the first open-weight model in the 3-trillion class — launched July 16 with full weights published July 27.8 It is the first open model ever to top WebDev Arena, ranking first of 99 models, and on the independent Artificial Analysis index it sits fourth overall — behind only the top closed flagships from Anthropic and OpenAI, and ahead of everything else.9 The frontier gap is no longer a tier. It is a margin of a few points, and the weights are downloadable.

At the price end: DeepSeek shipped the official V4-Flash on July 31, MIT-licensed weights on Hugging Face the same day, priced at $0.14 per million input tokens and $0.28 output.10 During its preview run, it was the most-used model on OpenRouter for seven consecutive weeks — not a benchmark claim, a revealed preference.11 That is frontier-adjacent capability at a price rounding toward zero.

That is the SaaS story with the nouns changed. Frontier capability was the moat; open weights are the commoditization; the margin compresses toward the cost of compute underneath it.

The pattern is the one this series has traced from the start, which is: commoditization does not stop at a layer; it climbs. It took the application layer's multiples in 2026. The model layer is priced as if it is exempt.

The hyperscalers' $725 billion says value accrues at the bottom of the stack. The open-weight releases say margins compress toward it. Both can be right. What cannot be right is a trillion-dollar valuation on the floor between them.

artistic-process-man-s-hand-close-up-pen-in-hand-2026-03-23-23-39-09-utc.png

What this means for the private equity workout

Tie it together.

The PE software book needs time, and the macro is taking the time away. The hyperscaler capex is financed on the assumption that AI revenue eventually justifies $7.6 trillion of infrastructure — an assumption that open-weight parity pressures from below, because commodity models mean commodity inference pricing. And the IPO window that is supposed to relieve the exit backlog is being held open by exactly the valuations that the compression climbs toward next.

If the AI listings reprice the way the SaaS cohort did, the window shuts, the denominator effect hits every institutional portfolio simultaneously, and the private equity workout loses its schedule and its exit route in the same quarter.

None of this requires a crash. It requires only that the same arithmetic keep working on the next floor up — and there is no visible mechanism that stops it.

The system is solvent. The schedule is the risk. And the schedule now depends on the durability of the one asset class priced as if compression does not apply to it.

Part of the series: Marks to Market
  1. The Clearing Price
  2. Solvent, But Not on Schedule
  3. Waiting on Appointments · coming soon

Footnotes

  1. The Clearing Price — Part one of this series, which sizes the envelope referenced here: roughly $1 trillion of marked PE software equity on top of $500–750 billion of debt, and the Airtable print that forced the first number into the open.
  2. Office of Financial Research, "Measuring Counterparty Exposures to Private Credit," Brief 26-02, March 12, 2026. — Maps how private credit losses reach the banking perimeter, and supports the comparison used here between a $50–75 billion bear-case impairment and $1.6 trillion of Tier 1 capital at the reporting banks. https://www.financialresearch.gov/briefs/files/OFRBrief-26-02-measuring-counterparty-exposures-private-credit.pdf
  3. Howland Capital, "U.S. Economy Q1 2026: Jobs, Tariffs & Inflation Outlook," April 2026. — Source for the quarter that closes the workout's four conditions at once: GDP revised down to 0.7% on weak consumer spending, the tariff substitution after the Supreme Court ruling, and the labor-market softening underneath both. https://www.howlandcapital.com/insights/u-s-economy-q1-2026-jobs-tariffs-inflation-outlook/
  4. Stanford Institute for Economic Policy Research, "The U.S. economy in 2026: What to watch for." — The tariff arithmetic behind the stagflation framing: an effective rate rising from 2.1% to roughly 11.7% with consumer pass-through now above 50%, which is what pins the Fed between a labor market asking for cuts and prices arguing against them. https://siepr.stanford.edu/publications/policy-brief/us-economy-2026-what-watch
  5. Tom's Hardware / Financial Times, "Big Tech's AI spending plans reach $725 billion," April 30, 2026; and Yahoo Finance / Goldman Sachs, June 3, 2026. — The capex figures the objection rests on — $725 billion across the four hyperscalers this year, up 77% from $410 billion, with a trillion projected for 2027 and Goldman modeling $7.6 trillion of infrastructure through 2031. Goldman's longer-range model is summarised at https://finance.yahoo.com/sectors/technology/article/meta-microsoft-amazon-and-alphabet-are-about-to-spend-a-shocking-amount-of-money-to-dominate-the-ai-era-115359575.html
  6. CNBC, "Tech AI spending approaches $700 billion in 2026, cash taking big hit," February 6, 2026. — Tracks what the buildout costs the balance sheets financing it, including the projection that Amazon's free cash flow goes negative to fund its share. https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html
  7. Fortune, "In its first deal since going public, Italian unicorn Bending Spoons buys Airtable at a $9 billion discount," August 5, 2026. — Context on the listing environment the exit backlog depends on, and on the private marks — Anthropic at a $965 billion post-money, SpaceX around $1.4 trillion — that are holding the IPO window open. https://fortune.com/2026/08/05/bending-spoons-italian-unicorn-startup-airtale-ipo/
  8. Northflank, "Kimi K3: benchmarks, pricing, hardware requirements, and self-hosting," updated August 3, 2026. — Release timeline and self-hosting requirements for the first open-weight model in the 3-trillion-parameter class — launched July 16, full weights published July 27. The hardware footprint is the reason "downloadable" is not the same as "free to serve." https://northflank.com/blog/what-is-kimi-k3-self-hosting
  9. MoClaw, "What Is Kimi K3? Moonshot's 2.8T Model," updated July 30, 2026. — The benchmark placement and the pricing complication in one source: WebDev Arena first of 99 models at 1,678 Elo, Artificial Analysis Intelligence Index 57.1 for fourth overall (behind Claude Fable 5 at 59.9 and GPT-5.6 Sol Max at 58.9), and an API priced at $3 input / $15 output — a fivefold jump over its predecessor. https://moclaw.ai/blog/what-is-kimi-k3
  10. Hugging Face, "DeepSeek V4 Flash Is Now Official: What Changed in the 0731 Build," July 31, 2026. — The price end of the pincer: MIT-licensed weights published the same day as the API, at $0.14 per million input tokens (cache miss) and $0.28 output. https://huggingface.co/blog/ResterChed/deepseek-v4-flash-official-release
  11. Caixin Global, "DeepSeek Releases Official V4-Flash Model as China's AI Race Intensifies," August 1, 2026. — Reports the revealed-preference datapoint used here rather than a benchmark claim: seven consecutive weeks as the most-used model on OpenRouter during the preview run. https://www.caixinglobal.com/2026-08-01/deepseek-releases-official-v4-flash-model-as-chinas-ai-race-intensifies-102470292.html
Back to the Journal