Newsletter · · Ashutosh Agarwal

Private Credit Default Data Finally Catches Up With the Bear Case - Private Credit and Alternatives - Week of August 26, 2026

Private Credit & Alternatives for the week of August 26, 2026. Podcast synthesis on the first hard Q2 default figures (five-year highs at Blackstone, Blue Owl, Ares and Golub, a 6% market-wide rate), why a leading bull's own argument places the AI debt downstream at CoreWeave, Oracle and the special-purpose vehicles, and Nvidia's move to make itself the collateral-setter for the whole structure.

Private Credit & Alternatives

Week of August 26, 2026: Private Credit Default Data Finally Catches Up With the Bear Case

Private Credit & Alternatives, New York, August 26, 2026

For most of this cycle, the bears have had the better story and the bulls have had the better numbers. The critics could describe, in vivid detail, how a private-credit blow-up would unfold, slow, opaque, sold as safe, but when they reached for hard evidence, the data kept coming back benign. Defaults were low. Redemptions, until recently, were manageable. The people warning of trouble were narrating a fire nobody could yet smell.

This week that gap started to close, and it closed from both directions at once.

First, the numbers finally showed up. A fresh review of the biggest private-credit funds' own quarterly reports found defaults at their highest level in five years, not at the fringe, but at the marquee names everyone owns: Apollo's peers, Blackstone, Blue Owl, Ares, Golub. The industry-wide default rate, by one rating agency's count, has reached 6%. This is the "slowly, then all at once" moment made concrete: the first line of it, in real figures, from the funds' own books.

Second, and this is the subtler, more important development, the loudest bull of the week accidentally drew a map to where the risk is buried. Raoul Pal came on to make the confident case that the AI boom is nothing like the dot-com bust. Along the way he pointed out that the giant technology companies doing all the building are barely borrowing at all, debt equal to just 4% of their stock-market value, against cash flows bigger than most governments. He's right. But follow the logic: if the hyperscalers aren't the ones taking on the debt, someone downstream is. And when you ask who, the answer is the exact list of names the bears have been circling for a year, the second-tier cloud companies, the weakest of the tech giants, and the private funds and off-the-books vehicles financing them. The bull case and the bear case now agree on where the borrowed money sits. They disagree only on whether it matters.

Put those two together and the week has a clear shape. The story is no longer just narrative. It has numbers behind it (real defaults, at the biggest funds), a map (the debt lives downstream, not at the deep-pocketed builders), a new pressure point (a bipartisan political revolt against the data centers that back so much of the new lending), and, as a counterweight, a genuinely thoughtful account, from an operator actually building these products, of how to sell private markets to ordinary savers without hurting them.

One honest caveat up front, because it matters for how much weight to put on the week: the bear case this time was carried mostly by data and by macro commentators, not by the operationally-credible insiders who dominated last week. The people running actual money who spoke this week were, if anything, constructive. That's a real shift in texture, and we'll come back to it.

TL;DR

  • The default data arrived, and it's the worst in five years, at the biggest funds. On Money Metals' Weekly Market Wrap Podcast, host Mike Maharrey walked through a Wall Street Journal review of the largest players' own quarterly filings: private-credit funds run by "Aries Management, Blackstone, Blue Owl Capital, and Golub Capital reported loan defaults hitting the highest level since 2021." Blue Owl's fund default rate "hit 2.8 percent in Q2… the highest level in at least five years"; non-performing loans elsewhere hit five-year highs; and per Fitch, "U.S. private credit default rate stood at 6% at the end of May," with "14 defaults in May alone." His frame: private credit is ~$1.4 trillion, about 10% of corporate debt, the same small share subprime was of mortgages before 2008.
  • Even the loudest bull says the debt is downstream, which is exactly where private credit lives. On Tom Bilyeu's Impact Theory, macro investor Raoul Pal argued the hyperscalers carry debt worth just "4% of their market cap" and generate cash "more than God," so they can pay "almost any interest rate… ad infinitum." The risk, he conceded, sits elsewhere: "There's CoreWeave, there's maybe Oracle, there's a bunch of special purpose vehicles. Could those blow up? Yes." That is the private-credit and off-balance-sheet corner of the market, named by a bull.
  • NVIDIA is quietly making itself too big to fail, and telling everyone the depreciation rate. On DHUnplugged, the hosts detailed NVIDIA's preliminary deal with "Apollo… BlackRock… Brookfield, Goldman Sachs, KKR" to "mobilize 500… billion" for AI infrastructure, with Jensen Huang calling it the "first time that technology chips have become an investable asset class." Their sharpest point: NVIDIA controls how fast its chips lose value by choosing when to release new ones, so it is now "slow-walking" its release cycle to make the collateral behind all this debt predictable, and positioning itself so that "his failure will take down a much, much larger market than himself," the definition of too big to fail.

What's new

1. The default numbers arrive, and they're at the biggest funds, not the fringe

For a year the bears have been describing a wave of defaults that the data refused to confirm. This week the confirmation started to land. On Money Metals' Weekly Market Wrap Podcast (Aug 19), a gold-and-silver show, so treat the surrounding editorializing accordingly, but the figures below come from a Wall Street Journal review of the funds' own quarterly reports, host Mike Maharrey laid out the first hard evidence. "A review of quarterly reports issued by the biggest players in the private credit market, analyzed… by the Wall Street Journal, shows… loan health and investor returns are worsening." Specifically, funds "overseen by Aries Management, Blackstone, Blue Owl Capital, and Golub Capital reported loan defaults hitting the highest level since 2021", the depths of the pandemic.

The single number that should stick: "the default rate at Blue Owl Capital's fund hit 2.8 percent in Q2… the highest level in at least five years." That does not sound alarming on its own. The point is the direction and the company it keeps: "non-performing loans at other funds also hit five-year highs, exceeding levels seen during the spike when the Federal Reserve was tightening… in 2023." Zoom out to the whole market and, "according to Fitch Ratings, U.S. private credit default rate stood at 6% at the end of May," with "14 defaults in May alone." So far the trouble is clustered "in healthcare… industrial and manufacturing, and the business service sectors," plus companies squeezed by higher oil prices, one example being Loparex, "a firm that makes plastic film," which "recently defaulted on its private loan." The worry Maharrey flagged is where it spreads next: "In many private credit funds, 20% or more of the outstanding debt is owed by software companies," the exact sector most exposed to disruption by AI.

Why this matters beyond the headline: the managers have almost no cushion left. As he put it, quoting the Journal's reporting, "private investment firms can't afford any more write-downs after a rough 2025," because their fund performance "is already suffering, and their stock prices have only just started to recover." And the mechanism by which a slow leak becomes a flood is the funding side: "if returns stay stuck or drop, more investors may walk away. Fewer investors mean it would be harder for fund managers to raise money, shrinking the supply of capital to refinance existing corporate loans when they come due." That is the doom loop in one sentence, worse returns, fewer investors, less money to roll over loans, more defaults. His comparison to 2008 is the one worth holding: at the top of the housing bubble, subprime was only "13% to 15% of all mortgages," yet its collapse "took the whole housing market down with it." Private credit is a similar sliver, about 10% of the roughly $14.5 trillion in U.S. corporate debt, and "all you need is one big default, and that's when the whole bottom can fall out." The data is not yet a crisis. But for the first time this year, it is pointing the same direction the bears have been pointing all along.

2. The bull's best defense is also a confession: the debt is downstream

The most intellectually useful moment of the week came from someone making the optimistic case. On Tom Bilyeu's Impact Theory (Aug 22), macro investor Raoul Pal set out to reassure. His central point is genuinely reassuring, and genuinely important. Ask how indebted the giant technology companies actually are, he said, and "the technical answer would be" almost nothing: "4% of their market cap in debt. It's fuck-all. How much cash do they generate? More than God." These are "the largest cash flow generating machines on Earth," and even after going cash-flow-negative to build data centers, "they've got the revenues to pay almost any interest rate on the debt ad infinitum." On that, he is almost certainly right, and it is the single best rebuttal to the crude "AI is a debt bubble" line.

But listen to where he puts the risk instead. "Now, there's a bunch of others," he went on. "There's CoreWeave, there's maybe Oracle, there's a bunch of special purpose vehicles. Could those blow up? Yes." That sentence is the whole ballgame for this newsletter. If the deep-pocketed builders aren't the borrowers, the borrowing is happening one layer down, at the second-tier cloud companies, the weaker tech giants, and the off-the-books vehicles built to fund them. And that layer is precisely where private credit, insurance money, and the new GPU-backed lending have gone. The bull isn't disagreeing with the bear about where the leverage sits. He's confirming it, and simply betting it won't cascade.

His reasons for that bet are worth steel-manning, because they're strong. If the worst happened, say OpenAI, with "trillions of dollars of obligations," failed, "firstly, the US government steps in," because it "can't allow the US to slow down on the race for intelligence." And the physical assets don't vanish: "all compute gets bought in seconds by the fastest bidder," so "it's actually very hard to have massive economic damage. Yes, you'll have a shock, but you don't have anything long-lasting." He drew the contrast with the fiber-optic bust of 2001–02, when the telecom builders "were insanely leveraged" and "had huge amounts of negative free cash flow… they made no money." Today's builders are the opposite. "We're just not there yet." He also thinks the boom has years left simply because it's barely begun: "I think we're about 30% of where we should be on the buildout this year, because the electricity permitting is so slow," with capital-spending plans "out to about 2030." And he noted the buildout is already throwing off cash for its owners, Elon Musk is leasing out his Colossus supercomputer at "30%, 35% margins plus," and "Google are doing both sides of this equation," building compute and renting it out.

The counterweight on the same episode came from the host, Tom Bilyeu, who is more cautious and channeled Warren Buffett: the CAPE ratio (a smoothed, ten-year measure of how expensive stocks are) is "sitting at 40x" versus a long-run average "of 16 to 17x", dot-com territory, and Buffett "is sitting on more… cash. Historic levels," having "backed out of his Apple position" even as he "just did add to his Google position." Bilyeu's worry is the classic one: "every revolutionary technology ends up taking on debt to the point where it wipes out the first round of investors." Pal's answer, adjust the CAPE for years of money-printing, and remember that "all investments are dangerous", is fair, but it doesn't rebut the specific point that the danger has moved downstream. Two smart people, one bull and one bear, ended up agreeing on the location of the risk. That agreement is the takeaway.

3. NVIDIA is engineering itself into too-big-to-fail, and volunteering the depreciation schedule

The mechanics of the frontier's financing got their clearest explanation of the cycle on DHUnplugged (Aug 19). The hosts walked through NVIDIA's preliminary agreements with "Apollo… BlackRock… Brookfield, Goldman Sachs, KKR" to "mobilize 500… billion" for AI infrastructure, Wall Street's biggest lenders pooling capital to finance the data centers that buy NVIDIA's chips, at what NVIDIA calls "attractive rates for customers." Jensen Huang's own framing, they noted, was that this is "the first time that technology chips have become an investable asset class."

The problem they zeroed in on is the one that has haunted this whole structure: these loans are secured by the chips, and chips lose value fast, faster than any previous kind of computer, and NVIDIA itself controls how fast, because it decides when to launch a better product. As they put it, the financiers' complaint is: "you're having us loan money out to people building data centers at a specific depreciation rate. And then you, NVIDIA, are at any moment able to change the depreciation rate by releasing a new chip, which will [gut] the value of the entire collateral." NVIDIA's response is quietly remarkable: it is "slow-walking" its release cycle, "incrementally increasing the rate at which new chips come out", and, more importantly, promising to "tell you the rate at which these will depreciate so we can turn these into a collateralized… object that you can predict the curves on." In plain terms, the company that controls the collateral's value is now publishing a schedule for how fast that value will fall, so lenders can model it.

The hosts also captured the bull framing NVIDIA prefers: stop thinking about the hardware and think about the compute. A data center "can sell a token for… a dollar per million tokens… no matter what the underlying hardware is," and even six-to-eight-year-old chips stay profitable because "the electricity costs and the maintenance costs… [are] in the sub pennies per million tokens." Reframe the asset as a stream of sellable compute rather than a depreciating box, and you can stretch the loan timelines. The catch they flagged: "this assumes NVIDIA has a lot of control over the rate of chips… But if anybody leapfrogs NVIDIA, this entire model collapses," because a cheaper, better competitor would crush the collateral overnight.

Their most provocative read was strategic. NVIDIA's play, they argued, "is always to position himself such that his failure will take down a much, much larger market than himself. Thus causing the sort of natural bailout structure where people want to keep him alive." By wiring "all finance fully dependent on NVIDIA not going upside down," Jensen Huang "may have created his own… moat." That is the textbook definition of too big to fail, the lesson we supposedly learned from the banks in 2008 and, they noted drily, plainly did not. Their nightmare scenario is worth quoting because it names the two things happening at once: if NVIDIA ever stumbled, it would be "much more like if '99 was mixed with 2007", dot-com enthusiasm and real, strandable hardware capital, layered on top of "financialization at a much greater magnitude." Every company all-in on AI and every institution that underwrote the loans "left holding the bag" simultaneously.

For balance, the retail-facing bull version of this argument turned up on Buy Hold Rant (Aug 19), where the hosts pushed back on the fear that all this is "circular financing", companies funding each other to buy each other's products. Their analogy was the car business: NVIDIA lending to data-center builders is "no different than… car manufacturers making it easier for their customers to get loans to buy cars." When GM or Ford first did it, "there may have been lots of eyebrows raised… But ultimately, it has turned out to be perfectly fine." On the $500 billion package, their read was that "there are legitimate lenders who are going to assess your business case and lend based on that. So it's not just blind money." They also made a useful, under-appreciated point: a lot of AI revenue isn't from the poster-child startups but from "existing products that are integrating AI capabilities", Google Workspace and Microsoft's Office at higher price tiers, which is "solid and probably not going away." They conceded the core tension honestly, though: "the spend for data centers is currently outpacing revenue… that is at least a temporary problem. How temporary? I don't know."

4. How to sell private markets to Main Street, from an operator who's actually building it

Against a week thick with warnings about ordinary savers being sold illiquid assets, the most valuable operator voice explained how a serious firm tries to do it responsibly. On Alt Goes Mainstream (Aug 19), Dave Donahoo, who leads Franklin Templeton's U.S. wealth business, described the guardrails. Franklin Templeton is a $1.7 trillion firm with an "approximately $300 billion" private-markets arm, built around specialist managers left to do one thing well, Lexington Partners in secondaries ("33 years"), Clarion Partners in real estate ("40"), and Benefit Street Partners in private credit ("20").

The most instructive part was his walk-through of why the wrapper matters as much as the asset. Take secondaries (buying existing investors out of their private-fund stakes). The easiest structure for a retail buyer is an "interval fund," but Donahoo refused it, for two reasons that cut against the industry's commercial incentives. First, fees: by regulation, taking private equity into an interval fund "eliminate[s] the incentive fee," and "most managers who choose to do that increase their management fee to offset" it, a swap he wouldn't make because it would mean pricing retail clients differently from institutions. Second, cash timing: an interval fund "you have to take capital daily," but secondary deals "typically close at quarter end," so daily inflows would force him to "dilute our underwriting standards to just wave deals in, or… sit on excess cash." So he chose a non-traded structure instead (the Franklin-Lexington Private Market Fund), which also lets the investor qualification "come down from" qualified purchaser to the less-restrictive qualified client, genuinely widening access, but through the structure he judged honest rather than the one that was easiest to sell.

On the retirement channel everyone is chasing, he was matter-of-fact: Franklin Templeton already "deliver[s] private real estate today into 401k plans primarily through custom target dates" and advisor-managed accounts, "with one of the largest record keepers in the U.S.", a reminder that private assets are already inside some Americans' retirement plans, quietly, via real estate. And on infrastructure, including data centers, the collateral behind so much of this cycle's lending, he described a partnership model with three specialists (DigitalBridge for "data centers and digital infrastructure," Copenhagen Infrastructure Partners for renewables, and Actis, "part of General Atlantic," for growth economies), structured with "contractually obligated deal flow with no fee layering." His closing note is the one the whole industry should tattoo somewhere: what clients want most is "somebody on the other end of the phone that if things don't work well, they can call," plus "sober selling and being really intellectually honest on the front end." Coming from a manufacturer, in a week when the data turned, that emphasis on honest expectations lands as either reassuring or ominous, depending on your priors.

5. The macro bear frame: three bubbles, hundreds of gated funds, and pensions holding things they don't understand

The week's fullest bear thesis came from a macro pundit, and it's worth surfacing precisely because it's the mirror image of Donahoo's careful optimism. On The Competent Investor (Aug 21), Chris MacIntosh of Capitalist Exploits argued that there are "actually three bubbles… the credit bubble, private credit bubble… the AI Mag7 equity bubble. And then the big daddy, which is the sovereign debt bubble", all "tied into one pivotal structural component… interest rates." His warning on the plumbing: private credit now runs "layered leverage… up to 20 times leverage on… illiquid assets," and much of that debt "was financed at rates three, 400, 500 basis points lower than today." As those loans mature and reprice higher, funds that "don't have the equity" to refinance are "forced to sell," only to find the assets "not at par… In many instances, the equity tranche is completely wiped out and even the debt's impaired."

The consequence, in his telling, is already visible: "we have now hundreds of these funds that are all gated", "gating" meaning investors "can't get [their] money back." And his most pointed charge concerns who ends up holding this. A "rule change… under the Trump administration" that let pension funds, insurers, and individuals buy private assets was, he argued, "an offloading of risk to the peasants. It was sold as… we're letting you in to the magic room. No, they were exiting" at high valuations. He says he has "spoken with some pension fund managers who don't even know what they bought," having allocated to "structured vehicles" that yield "12%" and look liquid "until it's not." On the AI side he made the same connection everyone else did, from the debt angle: the buildout is now being financed with borrowing rather than profits "because there is no revenue behind it… they don't make any money," and the companies themselves concede they won't be profitable "for… years if not decades." Keep the source in view, this is a doom-oriented macro commentator, and claims like "20 times leverage" and "hundreds of gated funds" are his assertions, not audited figures. But the shape of his argument now rhymes exactly with the hard default data in item 1.

6. A new risk at the frontier, and it isn't financial, it's political

The collateral behind the AI-credit boom is data centers, and this week that collateral picked up a threat no spreadsheet was modeling: voters. On Tom Bilyeu's Impact Theory (Aug 21), the panel flagged that "we have both left and right governors who… signed memorandums… for pauses on" data centers, "both Texas, Greg Abbott, and Josh Shapiro in Pennsylvania", and that "across party lines, both parties are declining in support of data centers." The striking detail: polling suggests "people would rather have a nuclear plant next to them than have a data center."

Bilyeu's own view is that the backlash is misguided, data centers "are net taxpayers wherever you find them," have "lowered people's property taxes" in places like Loudoun County, and built civic infrastructure (a data center in Quincy, Washington "built a high school… lowered utility rates"). A co-host pushed back on the jobs claim, these are "4 people and a bunch of electric plugs," not a warehouse, and they settled on a range of roughly "50 to 200 jobs" per site depending on size. They also aired the genuine local costs: heavy "peak water usage," one documented case of groundwater contamination, and the way "dry cooling" solves water only by "moving the problem over to electricity," which then runs into permitting walls (Microsoft's stalled restart of Three Mile Island came up). Whatever your view on the merits, the read-through for this newsletter is direct: if moratoria spread, the physical assets underpinning hundreds of billions in new lending get harder to build and slower to fill, a political dent in the collateral, arriving the same week the credit data softened.

Two ground-level signals worth noting briefly

First, the retail reach-for-yield is no longer theoretical, you can hear it being sold. On Millionaire Mindcast (Aug 21, a replay), host Matt Aitchison spent a segment pitching his own "Imagos Income Fund", "senior secured private credit, asset backed, first position, first lien loans" paying investors "10% annually… paid out monthly", and coaching listeners to borrow against their whole-life insurance policies at "5.5%" to invest in it and "pocket the spread." It's a promotion, not analysis, and should be read as such. But it is a near-perfect field recording of the exact dynamic the critics keep describing: illiquid private loans, wrapped in a promise of smooth monthly income, marketed down-market to ordinary savers, even financed with a second loan. Second, the underwriting foundation under software lending is quietly eroding. On Disruption / Interruption (Aug 20), DealSync founder Shalom Ben Or, describing asset-based lending as "a $1.5 trillion industry that runs on emails, PDFs, and meetings", noted that revenue-based lending against software subscriptions "became a shaky asset because of what AI did to SaaS," as usage- and outcome-based pricing breaks the tidy recurring-revenue math that private lenders underwrote against. Small window, but it's the software-loan worry showing up at the ground floor, from an operator, not a pundit.

The debate

Is the data the beginning of the wave, or the normal churn of a maturing market? This is the week's central question, and for once both sides have real evidence rather than just conviction. The bear reading is straightforward: defaults at the largest funds, Blackstone, Blue Owl, Ares, Golub, are at five-year highs, a 6% market-wide default rate is "pretty high," and the managers, per the reporting Money Metals cited, "can't afford any more write-downs" after a rough 2025. When the biggest, best-diligenced books start showing strain, the fringe is presumably worse. The bull steel-man, though, is not nothing. A 2.8% default rate at a flagship fund is still low in absolute terms; the trouble is concentrated in identifiable, non-AI corners (healthcare, industrials, oil-squeezed names); and rising defaults off a historically low base is exactly what a normalizing credit market looks like after a decade of near-zero losses. Franklin Templeton's Donahoo, building products for the long haul, plainly isn't behaving like a man who expects a collapse. The honest synthesis: the direction has turned, unambiguously, and it's turned at the names that matter, but the level is still early-cycle, not crisis. What resolves the debate is the software line item. Everyone agrees roughly a fifth of these loan books is software, everyone agrees AI is the wildcard there, and no one yet has the default data for that slice. That's the number to wait for.

Where does the AI debt actually live, and does it cascade? Here the remarkable thing is the agreement. Raoul Pal, the bull, says the hyperscalers are "4% of market cap in debt" and fine; the leverage is at "CoreWeave… Oracle… a bunch of special purpose vehicles." The bears would not change a word of that sentence, they'd just append "…and that's where private credit and insurance money went." So the factual map is settled: the borrowing is downstream of the cash-rich giants, in the second-tier clouds and the off-balance-sheet vehicles. The disagreement is purely about contagion. Pal's case that a blow-up stays contained is genuinely strong, the government won't let the AI race stall, the physical compute gets bought "in seconds," and unlike the 2001 telecoms, today's builders gush cash. The bears' case is that DHUnplugged's "'99 mixed with 2007" is the right frame: stranded hardware and financial engineering failing together, with NVIDIA wired in as the single point of failure the whole system is now built around. The swing factor is whether the downstream borrowers can keep servicing debt long enough for the buildout's revenue to arrive, a question of timing, and Pal himself says we're only "30% of where we should be on the buildout," years from the payoff.

Can private markets be sold to ordinary people safely? This debate got its two purest data points in the same week, at opposite ends. At the responsible end, Donahoo's Franklin Templeton refuses the fee-inflating interval-fund structure, matches the wrapper to the asset's true liquidity, and preaches "sober selling." At the other end, a personal-finance podcast is coaching listeners to take a loan against their life insurance to buy a 10%-yield private-credit fund. Both are real; both are the industry. The bull case is that the sophisticated manufacturers are genuinely trying to get the structure right, and that widening access to assets that now make up most of the economy is a real service. The bear case, unchanged and only reinforced this week, is that the marketing runs far ahead of the guardrails, and that the promise of smooth, high monthly income sold to skittish savers is exactly the mismatch that breaks in a downturn. Donahoo's own standard is the test: is there "somebody on the other end of the phone" when it goes wrong? For the Imagos-style products reaching down-market, the answer is unlikely to be yes.

What the silence still says

The most telling feature of the week is who spoke and who didn't. The insurers, the balance sheets holding the most opaque private-credit risk, were silent for a fourth straight week; MacIntosh's charge that pension and insurance buyers "don't even know what they bought" went, once again, unanswered by anyone actually running that money. Just as notable: this week the bear case was carried by hard data (the WSJ/Fitch figures) and by macro pundits (Money Metals, MacIntosh, DHUnplugged), while the operators who spoke, Donahoo building carefully, Pal reassuring, leaned constructive. That's the reverse of last week, when the most credible voices warning of trouble were insiders (an ICG portfolio manager, a hedge-fund manager actively shorting the trade, a veteran FT journalist). Read that two ways. Optimistically: the insiders have calmed down because the panic really has passed. Pessimistically: the data got worse while the people closest to it went quiet, and the warnings are now coming from the outside looking in. Either way, the asymmetry from last week persists, the sellers of the new story are everywhere, and the holders of the existing risk still aren't talking.

The names in play

  • NVIDIA (NVDA): the anchor of the entire AI-credit market, now actively managing its own systemic importance. Its ~$500 billion financing package with Apollo, BlackRock, Brookfield, Goldman Sachs and KKR turns chips into "an investable asset class," and its decision to slow-walk new chip releases and publish depreciation schedules is an attempt to make its own products' aging predictable enough to lend against. Per DHUnplugged, the strategy makes it "too big to fail", the single point whose failure would take everything down with it.
  • CoreWeave (CRWV) and Oracle (ORCL): named by a bull, Raoul Pal, as where the AI debt actually sits, the second-tier cloud and the weakest of the tech giants, alongside "a bunch of special purpose vehicles." If contagion comes, this is the layer to watch, not the hyperscalers.
  • Blue Owl (OWL), Blackstone (BX), Ares (ARES), Golub (private): the four funds whose own quarterly reports, per the Wall Street Journal review aired on Money Metals, showed defaults at their highest since 2021, Blue Owl's fund specifically at a five-year-high 2.8% default rate in Q2. Blackstone, Brookfield and Apollo also sit inside the NVIDIA lending consortium, putting several of them on both the stress side and the growth side of the same story.
  • Apollo (APO), BlackRock (BLK), Brookfield (BAM/BN), KKR (KKR), Goldman Sachs (GS): the named members of NVIDIA's financing consortium, the same roster of alternative-lending and Wall Street giants underwriting the frontier's debt.
  • Franklin Templeton (BEN): the operator counterpoint on distribution. Via Dave Donahoo, it detailed how it builds retail private-market products, choosing a non-traded structure over an interval fund to avoid inflating fees, delivering private real estate into 401(k)s through custom target-date funds, and partnering with DigitalBridge, Copenhagen Infrastructure Partners and Actis on infrastructure, the "do it the right way" case, from a $1.7 trillion firm with a ~$300 billion private-markets arm.
  • DigitalBridge (DBRG), Copenhagen Infrastructure Partners & Actis / General Atlantic (private): Franklin Templeton's chosen infrastructure specialists, the first focused squarely on data centers and digital infrastructure, i.e. the AI-buildout collateral, now being packaged for wealth clients.
  • The hyperscalers, Alphabet/Google (GOOGL), Meta (META), Microsoft (MSFT): the builders who, per Pal, are barely borrowing ("4% of market cap in debt") and generating enormous cash, leasing out spare compute at "30-35% margins." The point isn't that they're at risk; it's that their strength is what pushes the debt downstream to those who are.
  • The insurers, Athene, Corebridge, Global Atlantic, F&G, Brookfield Reinsurance (various): silent for a fourth straight week, even as a macro commentator publicly alleged that pension and insurance buyers "don't even know what they bought."

Read-throughs

  • The map of the AI debt is now agreed on, trade it. A bull (Raoul Pal) and the bears now place the leverage in the same spot: not at Google, Meta and Microsoft, but at CoreWeave, Oracle, and the special-purpose vehicles and private funds financing the buildout. Stop stress-testing the hyperscalers' balance sheets; they're the strongest part of the chain. Watch the second-tier cloud names, the weaker tech giants, and the off-balance-sheet vehicles, that's where a private-credit problem would actually surface.
  • The default data has turned; the software slice is the number that matters next. Defaults at the biggest funds are at five-year highs and the market-wide rate is ~6%, but the trouble is still in healthcare, industrials and oil-squeezed names, not yet software, which is ~20% of these loan books and the sector most exposed to AI disruption. When the software default figures start climbing in the quarterly reports, that's the confirmation of contagion. Until then, the data is early-cycle, not crisis.
  • NVIDIA's credit standing is now the master variable for AI lending. By making itself the collateral-value-setter and the too-big-to-fail node for the whole structure, NVIDIA has concentrated the system's risk onto one company. Two things to monitor: any sign a competitor could leapfrog its chips (which would collapse the "predictable depreciation" premise overnight), and any widening in the cost of insuring NVIDIA's own debt. The whole edifice rests on NVIDIA staying both dominant and pristine.
  • Political risk to data centers is a real, and newly bipartisan, threat to the collateral. Governors in both Texas and Pennsylvania have signed pause memoranda, and public support is falling across party lines. Moratoria don't hit a loan book directly, but they slow the construction and lease-up of the physical assets backing hundreds of billions in new lending. Track state-level data-center legislation as a leading indicator for the AI-credit story, not just an ESG sidebar.
  • The distribution gap is the tell on discipline. In one week you can hear a $1.7 trillion manager refusing a fee-inflating structure on principle, and a personal-finance host coaching listeners to borrow against their life insurance to chase a 10% private-credit yield. That spread, between how the best manufacturers build these products and how they're actually marketed down-market, is the clearest gauge of whether this democratization ends well. The further the marketing gets from Donahoo's "sober selling," the more the bears' liquidity-mismatch warning applies.

What changed

Last week the skeptics fused the two big stories, AI financing and private credit, into a single narrative: same money, same illiquid-and-marked-to-model machinery, same danger of risky things dressed as safe. It was a powerful story, but it was still a story, carried by insiders describing a fire that hadn't yet shown up in the numbers.

This week the story acquired evidence. The first hard default data arrived, and it pointed the bears' way: five-year-high defaults at Blackstone, Blue Owl, Ares and Golub, a 6% market-wide rate, and managers with no cushion left for write-downs. At the same time, the cycle's loudest bull inadvertently drew the map, Raoul Pal's insistence that the hyperscalers are barely levered is also a confession that the debt sits downstream, at CoreWeave, Oracle and the special-purpose vehicles, exactly where private credit and insurance money have gone. The bull and the bear now agree on where the risk is; they differ only on whether it spreads.

Three other things shifted the picture. NVIDIA's role deepened from lender to systemic linchpin, controlling its collateral's value, publishing depreciation schedules to keep the loans financeable, and, per DHUnplugged, wiring itself in as the too-big-to-fail node the whole market now depends on. A genuinely new risk appeared at the frontier, and it's political rather than financial: bipartisan governors pausing data-center approvals, threatening the physical collateral behind the lending. And the retail-distribution debate got its cleanest split yet, an operator (Franklin Templeton) explaining how to sell private markets responsibly, in the same week a podcast ad showed how recklessly they can be sold.

The one thing that didn't change is who's talking. The insurers stayed silent for a fourth week. And in a reversal from last week, the voices warning of trouble this time were mostly data and macro pundits, while the operators running money leaned constructive. That reversal is genuinely ambiguous: either the insiders have relaxed because the danger has passed, or the numbers got worse while the people closest to them went quiet. The cycle's central question, are the values real? didn't get answered this week. But it did, for the first time, get its first real numbers. They were worse than the year's headlines, and they came from the biggest funds' own books.