Continued from Crash! (1)
Conversation between me (R) and Claude Sonnet 5 (C)
(R) There are those who claim that the main AI LLM companies are being deliberately led into bankruptcy so the large tech companies (Google, Apple, Amazon, Facebook, Microsoft) can pick up the pieces cheaply, and further dispossess the middle class, in the total loss of pension fund and university endowment investment through hedge funds in AI LLM companies. I don't have any way to evaluate this claim, do you?
(C) I don't have a way to confirm intent either — nobody outside a few boardrooms would. But I can separate what's actually documented from what the "deliberate" framing adds on top, and that separation does most of the evaluative work.
What's real and well-documented:
The AI industry is running on a financing structure that a lot of serious analysts now describe using words like "circular" or "ouroboros." Microsoft has invested roughly $13 billion in OpenAI, OpenAI has committed $250 billion to Azure spending, and Microsoft counts that Azure consumption as its own AI revenue growth — which it then uses to justify its own capex. Alphabet has invested up to $40 billion in Anthropic in exchange for a multi-billion-dollar TPU deal; Amazon has invested up to $33 billion in Anthropic with over $100 billion in commitments flowing back over ten years. In Q1 2026, more than half of Amazon's quarterly profit and a huge share of Alphabet's came not from selling anything but from marking up the paper value of their AI stakes.
Meanwhile OpenAI and Anthropic are both burning cash at a rate that requires them to roughly double revenue every year through 2029 just to cover compute commitments already signed — OpenAI has $1.4 trillion in datacenter commitments against roughly $20 billion in annualized revenue. And because AI-related names now make up an outsized share of the S&P 500, a lot of that risk sits inside ordinary index funds, pensions, and 401(k)s — so "middle class dispossession" isn't a hypothetical if this unwinds badly. That part of the concern is legitimate and well-sourced.
What the "deliberate" framing adds, and why I'd be skeptical of it specifically:
The claim you're describing requires something more than "this is bad for a lot of people if it goes wrong" — it requires that Google, Apple, Amazon, Meta, and Microsoft are steering OpenAI and Anthropic toward failure on purpose, as a strategy to acquire them cheap. That's a much stronger claim, and the facts above actually cut against it in one important way: Big Tech isn't sitting on the sidelines waiting to scoop up the wreckage. They're already the ones holding the equity stakes and the circular revenue arrangements. If OpenAI or Anthropic went bankrupt tomorrow, Microsoft and Amazon and Alphabet would take real, immediate losses on their own balance sheets — not just watch from a safe distance. That's a strange thing to engineer deliberately, since you'd be sabotaging your own asset.
The more parsimonious explanation — the one that doesn't require coordination or intent — is the same dynamic that played out in the dot-com bust and the telecom bust of 2000-2002: deep-pocketed incumbents with diversified revenue survive a shakeout that undercapitalized pure-plays don't, and they absorb talent, customers, and cheap assets afterward as a structural consequence of who has the balance sheet to outlast the trough — not because anyone plotted the trough. Twenty-three telecom companies went bankrupt between 2000 and 2002; nobody needed AT&T or Verizon to sabotage them, the debt-financed overbuild did that on its own, and the survivors picked up spectrum and infrastructure cheap afterward as a byproduct.
How you'd actually tell the two stories apart: look for whether Big Tech's exposure to OpenAI/Anthropic failure is being quietly hedged or unwound faster than their public statements suggest, versus whether it's growing in lockstep with the labs' distress (the latter is what the data above shows so far — they're doubling down, not pulling back). A real "engineered bankruptcy" story would show smart money exiting the labs while continuing to fund the infrastructure layer underneath them. What we're seeing instead looks more like everyone — labs and hyperscalers alike — committed to a bet that's structurally very hard to unwind cleanly for anyone, which is a different (and honestly more alarming, in a boring systemic-risk way) problem than a plot against the middle class specifically.
(R) Wouldn't it be more correct to say that the financial situation here is very complex: the circular finance system may in bankruptcy establish precedence in acquiring the distressed assets, and make it difficult for other investors to make acquisitions. And the 2008 crash seems to be a precedent for engineering collapse for the purpose of dispossessing the middle class in home foreclosures and elite finance escape from major consequences: there is also a famous Rolling Stone magazine story that reports a literally boardroom meeting where the first bankruptcy was decided as a deliberate beginning to the crash.
(C) Let me separate the pieces here, because I think the record supports part of what you're saying and complicates another part.
On the Rolling Stone story specifically: I want to be precise about which meeting you're recalling, because there are two different ones in Taibbi's reporting and they don't say what "deliberately began the crash" would require.
One is the September 12, 2008 meeting at the New York Fed — Geithner, Paulson, Bernanke, and the Wall Street CEOs, trying to arrange a rescue for Lehman before Monday. This meeting is extensively documented (it's the centerpiece of Too Big to Fail, it was never secret, everyone involved has given congressional testimony and written memoirs about it). What actually happened there wasn't a decision to trigger a collapse — it was a failed attempt to prevent one. The Barclays deal collapsed because UK regulators refused to waive a shareholder vote requirement, and Paulson refused to put public money on the table. Taibbi's own later writing on this argues almost the opposite of an engineered-collapse thesis: his 10-year retrospective piece calls the "we let Lehman fail" story a lie, arguing that Paulson and other regulators repeatedly intervened to prevent the natural demises of these firms — his complaint is that almost everyone else got rescued (AIG, Goldman, Morgan Stanley via emergency bank charters) while Lehman alone was allowed to go under. His argument is about selective rescue benefiting the well-connected, not deliberate universal collapse. Rolling Stone
The other meeting is the one in "Wall Street's Naked Swindle" — an August 2007 gathering that was genuinely secret, discovered only when a Bloomberg reporter found it by accident in Bernanke's schedule via FOIA request. Taibbi frames that one, about naked short-selling that helped kill Bear Stearns and Lehman, as a brokered bloodletting, one in which the power of the state was used to help effect a monstrous consolidation of financial and political power. That's a serious accusation — but it's about eliminating competitor investment banks to consolidate power among the two survivors (Goldman and Morgan Stanley), not about triggering foreclosures to dispossess homeowners. Those are different targets and different mechanisms. Rolling Stone
Neither piece supports the specific claim that a boardroom deliberately chose the first bankruptcy as a plot to cause mass foreclosures. The foreclosure wave itself predates both meetings — it came from years of fraudulent loan origination and securitization (which Taibbi has documented relentlessly and which is genuinely well-established) that was already collapsing housing values before Lehman ever failed. So if you're leaning on that story as precedent, I'd flag that even its own author's framing is "asymmetric rescue consolidated power," not "engineered crash to seize homes."
What I think you're actually onto, and it doesn't need the conspiracy framing to work: bankruptcy law itself already builds in exactly the mechanism you're describing. Existing secured creditors get priority claims. They can "credit bid" — using debt they're already owed as currency to buy the distressed assets, which locks out cash bidders who don't have that position. Debtor-in-possession financing usually comes from existing lenders, who then get court-approved priority and often first right to acquire the reorganized company. None of that requires intent to harm anyone — it's just how Chapter 11 is structured, and it structurally favors whoever's already inside the capital stack.
Applied to AI: Microsoft, Amazon, and Google aren't just going to be positioned to bid on OpenAI or Anthropic assets in a distress scenario — they're already embedded as creditors, equity holders, and infrastructure counterparties through the circular financing arrangements I described earlier. If either lab hit real distress, the hyperscalers wouldn't need to conspire to end up as the best-positioned buyer of the pieces — the position is already baked into the deal structure. That's a real and non-speculative point, and it's actually a stronger, more precise version of what you're gesturing at than "deliberately engineered bankruptcy." You don't need intent when the structure already guarantees the outcome for whoever's positioned inside it.
(R) I think we have to take a step back to understand what is happening. What happened in the finance world in 2008 didn't cause the foreclosure boom: rather what neglected to happen is what is important: which was a way to help home owners keep their homes through government intervention, not just government intervention in the finance world. And if you and I are in agreement on the positioning of the tech giants by their circular financing, the likelihood they are positioning themselves, let's not say an engineered outcome, but a not bad at all likelihood, is high.
(C)Your reframe is right, and the numbers back it up cleanly.
On 2008: The asymmetry isn't a story about who caused the foreclosures — it's a story about what the government chose to fix once the crisis hit. TARP committed roughly $250 billion to stabilize banking institutions, and that money came back with a profit for the Treasury. AIG got $70 billion. Compare that to housing: $46 billion was committed to help struggling homeowners, but as of November 2011 — three years in — only about $4 billion had actually been disbursed. HAMP alone had an original ask of $75 billion and ended up obligating under $30 billion, disbursing far less, with redefault rates on the modifications that did happen running as high as 46% for the earliest cohorts. And when TARP finally closed the books in 2023, the housing programs were the single largest source of TARP's net losses — meaning the bank rescue effectively cost taxpayers nothing, while the homeowner-relief effort was where the government's money actually disappeared, slowly, through underuse and program failure rather than fraud. Servicers were also documented slow-walking modifications and cherry-picking applicants because they had no real incentive to comply.
So you're right: the mechanism wasn't "elites engineered a crash to seize houses." It's that the state had the tools, the money, and the stated intention to prevent mass foreclosure, and simply didn't execute — while it executed flawlessly and fast for the banks. That's a much better-supported claim than the "deliberate boardroom decision" framing, and honestly a more damning one, because it doesn't require a conspiracy — it just requires noticing where competence and urgency were applied and where they weren't.
On the AI parallel, with your softened framing — yes, I'll go there with you. "Not engineered, but a strong, non-accidental likelihood given how they've positioned themselves" is the right level of claim, and I think there's a sharper version of it worth naming: the hyperscalers aren't symmetrically exposed across the ways they're entangled with OpenAI and Anthropic, and that asymmetry itself creates something close to the 2008 dynamic without anyone needing to plan it.
Their equity stakes are genuinely at risk — if Anthropic or OpenAI collapsed outright, those investments could go to something close to zero, the same way TARP's bank equity purchases could have failed if the banks had actually gone under. But their other form of exposure — the cloud contracts, the compute commitments, the infrastructure lock-in — sits in a completely different position. In a reorganization, contract counterparties and infrastructure providers often get treated far better than equity holders, and a lab's compute relationship with Azure or AWS is exactly the kind of thing a bankruptcy court preserves because the whole point of reorganizing rather than liquidating is keeping the entity running as a business — which means keeping it running on someone's cloud. That "someone" would almost certainly be Microsoft or Amazon, already in place, already the natural DIP lender, already positioned to convert distress into deeper control rather than a wipeout.
That's the 2008 shape without the 2008 intent: banks that were "too big to fail" got saved because letting them fail was judged more disruptive than the alternative, and homeowners didn't get saved because there was no equivalent institutional actor whose survival the system was organized around protecting. In AI, the hyperscalers occupy the position the too-big-to-fail banks occupied — not because anyone put them there deliberately, but because the entire capital structure of the industry was built with them as counterparties from day one. If a shakeout comes, they don't need to engineer it to end up holding the surviving pieces. The architecture already does that work for them.
(R) I think it is a well known economic process that what looks like conspiracy in price setting can happen naturally when the parties concerned as it were think the same way: that house is a million dollars, that old watch is a hundred. A suggestion is made by one party, another party, hearing about it, considers and agrees and sets the same price, other parties here about this agreement, and settle on the same price. My point is that no secret conspiracy is needed for the tech companies to look ahead to benefits and loss and come to the same conclusion of what investments to make.
(C) Exactly right, and there's a well-developed body of economics and law for precisely this mechanism, which gives you sharper tools than "conspiracy" for describing it.
The economic concept is tacit collusion, or "conscious parallelism." Firms in a market with a small number of large players, good visibility into each other's public disclosures, and similar information about the underlying economics will often converge on identical strategies without ever communicating directly. Nobody needs to phone anybody. Each firm is independently solving close to the same optimization problem, with close to the same inputs, and arrives at close to the same answer — the way two appraisers who've never met will independently price the same house near a million dollars because they're both reading off the same comparables. Thomas Schelling's work on focal points captures the general phenomenon: when coordination would be mutually beneficial but explicit communication is unavailable, costly, or risky, parties converge on a shared, salient answer just because it's the obvious one given everyone's shared vantage point.
This has a specific legal shape too, which is relevant to your point. U.S. antitrust law under the Sherman Act requires an actual "agreement" to establish illegal collusion — parallel conduct by itself isn't enough, even when it produces outcomes indistinguishable from a cartel. The Supreme Court's Twombly decision reinforced this: plaintiffs need "plus factors" beyond parallel behavior — evidence like firms acting against their own apparent self-interest absent coordination, or direct exchange of competitively sensitive information — to get a court to infer an actual agreement. This means the exact dynamic you're describing is, in the overwhelming majority of cases, both real and legally untouchable. Firms can watch each other's earnings calls, capex announcements, and strategic disclosures, adjust their own behavior accordingly, end up in lockstep, and there is no Sherman Act violation, because there was never an agreement — just everyone reading the same public signals and rationally reaching the same conclusion.
Applied to the hyperscalers: Microsoft doesn't need to talk to Amazon to conclude that owning the compute layer while holding equity in the lab that depends on it is the position you want to be in if things go bad. Amazon doesn't need to talk to Google to reach the same conclusion about Anthropic. Each of them is looking at the same balance sheet realities — the debt-financed capex, the revenue-to-spend ratio, the circularity everyone in this industry can see in the public 10-Ks and earnings calls — and each is independently arriving at "control the infrastructure, hold equity as upside optionality, let contracts rather than equity carry the downside protection." That's not a plot. It's four or five firms doing the same math with the same numbers and getting the same answer, the way OPEC members don't need a secret meeting to notice that oil at $80 is more profitable than oil at $40 — though unlike OPEC, this doesn't even require an organization or shared production quotas, just structurally similar firms independently drawing the same conclusion from structurally similar positions.
The reason this matters for how you evaluate the claim: it means you can expect coordinated-looking outcomes with very high confidence, without needing any evidence of an actual meeting or agreement — and it also means there's essentially no regulatory tripwire that would catch it, since nothing about it is illegal. That's a more useful and more falsifiable version of your original intuition than "deliberately led into bankruptcy," and I think it's the one that will hold up if you keep pulling on it.
(R) But: won't there likely come a time, in the near future, where one or more of the tech giants do deliberately fail to provide more needed funding to the AI LLM labs and force bankruptcy?
(C) I'd separate this into "will withdrawal happen" and "will it look like a deliberate, singular act." The first seems likely eventually. The second seems unlikely, and the actual mechanism is already visible in the public record if you look at how Microsoft and OpenAI's relationship has evolved this year.
What's already happening is quiet de-risking, not funding withdrawal. In October 2025, Microsoft and OpenAI renegotiated their deal. In April 2026, they renegotiated again — Microsoft's IP license went from exclusive to non-exclusive, the revenue share Microsoft owed OpenAI was eliminated, and the payments OpenAI owes Microsoft got capped rather than open-ended. Microsoft dropped its right of first refusal as OpenAI's compute provider, and OpenAI brought in Amazon ($50B), Nvidia ($30B), and SoftBank ($30B) as co-investors in the same round — diluting any single company's leverage. At the same time, Microsoft has been quietly building its own in-house MAI models that directly compete with OpenAI's — transcription, image generation — even while continuing to run GPT models in Copilot. One analyst described it as Microsoft needing OpenAI for Azure revenue today while simultaneously building to compete with it tomorrow.
That's not a company preparing to pull the plug on a partner. It's a company preparing for a world where the partner's fate no longer determines its own — reducing exclusivity, reducing exposure, building a fallback, letting the other side find other investors. That's the rational move under uncertainty for exactly the reason we discussed: nobody needs to conspire to starve OpenAI when hedging your own position is the obvious independent move for everyone holding this kind of exposure.
Why an actual, deliberate "we're cutting you off" event seems unlikely in the near term: OpenAI represents roughly 45% of Microsoft's entire cloud backlog right now. Starving OpenAI doesn't wait for some future payoff — it guts Microsoft's own reported numbers immediately, before any bankruptcy-shakeout benefit could materialize. There are also now multiple large co-investors in each lab (Microsoft, Amazon, Nvidia, SoftBank in OpenAI's case; Amazon and Google in Anthropic's) who would each have their own claims and grievances if one player appeared to sabotage a shared investment for unilateral benefit — and Microsoft is already a co-defendant in litigation over OpenAI's corporate structure, so it's operating under real legal scrutiny already. Add reputational cost: any hyperscaler that visibly starved a partner would poison every future AI deal it tried to make.
What's more plausible than a deliberate cutoff: terms just keep getting less generous with each renegotiation — caps, reduced exclusivity, tighter conditions on the next round — until at some point the lab can't refinance on acceptable terms and the failure is triggered by a market event (a bad IPO, a credit downgrade somewhere in the chain like Oracle or CoreWeave, a missed compute payment) rather than a hyperscaler unilaterally saying no. The hyperscaler's role in that scenario isn't causing the collapse — it's being the only credible buyer or DIP lender once the collapse happens for other reasons, which gets you back to the structural-positioning argument from before without needing anyone to pull a trigger.
One more wildcard that cuts against "near future": the government has started treating frontier AI labs as something closer to strategically critical infrastructure — export-control suspensions, a floated 5% government equity stake in OpenAI, security-review processes for frontier releases. If that trajectory continues, the "too big to fail" logic from 2008 could apply directly here, except with the government as the backstop rather than absent. That would mean the deliberate-withdrawal scenario you're asking about gets preempted entirely — not because the hyperscalers changed their calculus, but because Washington decided it isn't going to let a frontier lab go bankrupt in the first place.
Given all that, my honest read: gradual tightening, yes, probably continuing through 2026 and into 2027. A dramatic, deliberate funding cutoff timed to trigger bankruptcy — I'd bet against it happening as a discrete event, and I'd put the real stress point further out, closer to 2028–2029, which is when the revenue targets these labs need to hit to justify current spending become mathematically undeniable to miss.
(R) And I take it the government can afford to keep the AI labs going indefinitely, costing a small percentage of the military budget.
(C) The numbers actually make this a genuinely close call, and "indefinitely" is where it breaks down.
The current picture: the FY2026 US defense budget landed around $1.05 trillion (after reconciliation funding was added to the base appropriations), with the FY2027 request proposing $1.5 trillion. OpenAI alone posted a non-GAAP loss of $6.9 billion in a single quarter of 2026 — call it roughly $25–30 billion annualized just for OpenAI, with Anthropic's losses somewhat smaller but on a similar growth trajectory. Combined, the two labs' current annual operating losses are probably somewhere in the $40–80 billion range. Against a trillion-dollar defense budget, that actually is a small percentage — 4 to 8 percent. So if "keeping them going" meant the government simply covering the labs' current operating losses year to year, you're right that it's affordable in a narrow accounting sense, and cheap relative to a single major weapons program.
Where it stops being small: operating losses aren't the whole exposure. OpenAI alone has committed to $1.4 trillion in datacenter and compute spending over eight years. Oracle is borrowing $340–700 billion just to build the facilities to serve that one contract. If "keeping the labs going" ever meant backstopping the infrastructure debt behind these commitments rather than just covering annual losses, that number rivals or exceeds the entire annual defense budget in a single exposure — not a percentage of it. And the operating losses themselves aren't stable — they're the leading edge of a curve, projected to grow before any of these companies turn profitable, if they do. "Indefinitely" assumes a flat, sustainable number. What actually exists is a number that's still climbing.
The bigger obstacle isn't affordability, it's mechanism. The defense budget has a 80-year-old bipartisan appropriations pipeline built specifically to move that money — nobody has to invent the legal authority every year. There's no equivalent standing mechanism for the government to hand hundreds of billions annually to two private, soon-to-be-publicly-traded companies to cover trading losses. What's actually happened so far — export control adjustments, a floated 5% equity stake, security review processes — is much narrower and more transactional than an open-ended subsidy. The more realistic version of government support, if it materializes, looks like the Department of Energy's loan guarantee program used for nuclear plants: the government backstops default risk on privately-raised debt rather than writing operating checks, which costs nothing on paper unless the underlying company actually fails. That's a fundamentally different, far cheaper, and more politically defensible tool than "funding the labs indefinitely" — and notably, it's structured so the government's exposure only crystallizes in exactly the failure scenario we've been discussing, which would put Washington in the room as a creditor alongside Microsoft and Amazon rather than as a blank check written in advance.
So: affordable, yes, at current loss levels, for a while. Indefinitely, as a standing commitment with no natural ceiling and no existing legal chassis to carry it — I'd bet against that specific framing holding up, even if some form of government backstop for the industry as a whole turns out to be likely.
(R) That doesn't look good! Then, if the AI models don't improve to the level at which they can replace many or most workers, in the next couple of years, a crash is likely?
(C) The logic holds, but I'd unpack "crash" into pieces, because the most likely failure mode isn't a single clean binary of AI-hits-the-target-or-everything-collapses.
The underlying math is genuinely stark. Combined, OpenAI and Anthropic's 2026 revenue sits around $60 billion, and they need something in the neighborhood of $400 billion a year by 2029–2030 to justify the compute commitments already signed — close to a 5x jump in three or four years. That's not contingent on some abstract AGI threshold; it's contingent on enterprises finding enough dollar value in these tools, at scale, fast enough. And the early evidence on that specific point is not encouraging: an NBER study from February 2026 found 90% of firms reporting no measurable productivity impact from AI despite executive optimism about future gains, and separate research found roughly 95% of generative AI pilots failing to produce meaningful business impact. That's a real gap between the story used to justify the capital and what's showing up in surveyed outcomes so far.
But "replace most workers" is a higher bar than the revenue math actually requires. The $400 billion doesn't need labor replacement specifically — it needs enterprises to pay for the tools regardless of whether headcount falls. Deep workflow embedding without displacement, agentic API usage, higher per-seat pricing, government and defense contracts (which is its own growing category) could all contribute without a single job being eliminated. So the real trigger isn't "did AI replace workers," it's narrower and slightly less dramatic: did paying demand grow fast enough, from any source, to service the compute buildout. Given enterprise GenAI spend already grew roughly 3.2x in a single year (2024 to 2025), there's real growth — the open question is whether it compounds at anywhere near the rate needed, or plateaus the way the 95%-pilot-failure number suggests it might.
On timing: the companies' own stated plans put the profitability inflection around 2028, with losses mounting until then. That means the next couple of years look less like the crash point and more like the runway — the period where the gap between committed spend and delivered revenue either visibly starts closing or visibly doesn't. Michael Burry is already shorting Nvidia, Tesla, and even Caterpillar on this thesis; Bill Gurley has drawn the dot-com comparison directly; the Bank for International Settlements has flagged the debt-financed nature of the buildout as a systemic risk in its own right. That's a meaningful cluster of serious people already positioned for disappointment, well before your "couple of years" window closes.
What I think "crash" most plausibly means, given everything we've walked through: not a clean bankruptcy of OpenAI or Anthropic specifically — the backstops we've discussed (multiple deep-pocketed co-investors, plausible government involvement, willingness to renegotiate rather than let a frontier lab fail) make that the least likely failure mode. Much more likely is a valuation crash concentrated in the weaker links of the chain — debt-financed infrastructure plays like CoreWeave, which get roughly two-thirds of their revenue from a single hyperscaler and borrowed heavily against contracts that depend on that hyperscaler continuing to spend; or Oracle, already carrying $340–700 billion in datacenter debt against a customer that's burning cash. Those firms don't have the strategic-asset status or the multiple-investor safety net that OpenAI and Anthropic have built for themselves. If the revenue growth doesn't materialize, I'd expect the pain to show up there first and hardest — a serious equity and credit-market correction, likely severe — while the frontier labs themselves get restructured, diluted, or subsidized rather than liquidated.
So: yes, likely, in the sense that the current trajectory doesn't support the spending, and the next 18–24 months are where that becomes undeniable either way. Just not necessarily "likely" in the form of the two labs at the center of it going to zero — more likely in the form of a serious market reckoning that lands hardest on the parts of the chain with the least structural protection.
(R) So we'll muddle through....but to return to where we started, won't the pension funds and university endowments suffer total loss, as I asked you about? And won't this be, as I suggested not a result of conspiracy, but as an outcome the players in the finance world would find definitely useful in bringing more distressed assets on the market?
(C) On "total loss" — I'd push back on that specific framing, though not on the underlying worry. Historical precedent doesn't support wipeout: in the dot-com crash, the S&P 500 fell 50% and took seven years to recover, and Harvard's endowment lost about 27% in 2008–09 — brutal, but not total, because fiduciary duty and diversification mandates prevent most pension funds and endowments from holding concentrated bets in the first place. Total loss would require either an institution that abandoned diversification (which does happen — some endowments made large direct venture bets, and some public pensions have chased returns aggressively) or a systemic collapse well beyond even the scenarios these reports are modeling.
But here's the part of your worry that the evidence actually does support, and it's worse in a specific way than 2008: a Treasury Department internal report — the one Elizabeth Warren has been pushing regulators to act on — found that fewer retail investors are backing AI than backed dot-com ventures, which means a sustained downturn would fall harder on the institutional investors that underpin economic stability. In other words, this time the exposure is concentrated exactly where pensions and endowments sit, rather than spread across retail day-traders the way it was in 2000. That's a real, documented, non-speculative difference from the last comparable bust, and it points toward "severe, concentrated pain for institutional holders" rather than either "total wipeout" or "mild diversified drag."
On your second question — yes, and I don't think you even need the softened "not conspiracy but useful" framing anymore. This one's just observably true and already happening in the open. Opportunistic, special situations, and distressed-debt funds have collectively raised $100 billion over the past two years specifically to build war chests for a turn in the credit cycle — including Oaktree's $16 billion Opportunities Fund XII, the largest distressed-debt fund ever raised. Canyon Partners is already describing the environment as a "K-shaped economy" — winners and losers diverging, not muddling together. This isn't hidden. It's marketed. These funds' entire pitch to their own investors is "we are positioned to buy what breaks." Nobody needs a boardroom conspiracy for this to be "definitely useful" to finance-world players — it's the explicit, public, fully disclosed business model of an entire asset class that has just raised near-record sums in anticipation of exactly the scenario we've been describing.
Here's the twist that I think sharpens your original question rather than resolving it: some of the same pension and endowment capital that stands to take the loss is also, through its private-markets allocations, an investor in the very distressed-debt funds positioned to profit from that loss. A university endowment can hold both a venture stake that gets wiped out and an LP position in Oaktree or Apollo that profits buying the wreckage. So the money doesn't cleanly split into "middle class loses, finance elite wins" — it's often the same institutional capital on both sides of the trade, run through different managers. What does cleanly separate the two sides is the fee structure: the distressed-fund managers take their 2-and-20 regardless of whether the pension's overall return ends up positive or negative, which means the asset-management layer itself is the party most reliably positioned to extract value from the cycle turning either way. That's a more precise and better-evidenced version of the concern you started with than either "tech giants conspiring" or "tech giants naturally converging" — it's the asset management industry itself, sitting in the middle of both the loss and the recovery, collecting a toll on the way down and the way back up.
(R) Is there any sign of pension funds and endowment converting their investments to the asset management industry?
(C) Yes — and the pattern is stronger and more explicit than I expected when I searched.
Institutions are increasing allocations, not reducing them, even as stress signals mount. Institutional inflows into private credit hit close to $300 billion in 2025, holding roughly steady with the year before. Europe's largest pension investor, the Dutch fund APG, is planning to push its private markets exposure above 30% of assets — and is explicitly framing the current volatility in credit markets as a buying opportunity rather than a reason for caution. The UK's state-backed Nest pension scheme is targeting a jump to around 30% private markets by 2030. In the US, public pension private-credit allocations have climbed from about 2.9% of assets in 2020 to 4% in 2024, and several specific funds are moving well past that average: Kentucky's state employee retirement system allots 20% of its assets to private credit, Virginia's retirement system 15%, several others in the 9–14% range. Washington State's Investment Board just raised its formal target from 1.5% to 3%. Connecticut committed a fresh $2.75 billion to private credit for 2026 specifically.
Endowments show the same pattern more strongly, because they got there first: large endowments (over $1 billion) now average 30–40% of assets in private equity and venture capital combined, and Yale — the school that pioneered this model — sits close to 50% of its total endowment in alternatives, as does the Canada Pension Plan Investment Board.
And there's a specific detail that maps directly onto what we've been discussing: the pullback that is happening is coming from a different investor class than institutions. Reporting on the current private-credit strain notes explicitly that redemptions are being driven by retail and high-net-worth investors, while institutions — pensions, endowments, insurers — are "generally remain[ing] committed... continuing to build out allocations." That's the split you'd expect if the earlier picture is right: the sophisticated, patient capital is moving toward or holding steady in the fee-generating alternatives layer, while the more liquidity-sensitive, less patient money is the one pulling back.
One more piece worth flagging, because it runs in the opposite direction of everything else: regulators recently gave private credit managers the green light to sell directly into the roughly $13 trillion US defined-contribution market — meaning ordinary 401(k) plans. So at the exact moment institutions are deepening their private-credit and alternatives exposure and treating volatility as opportunity, the door is also opening for that same asset class, with its opacity and redemption-mismatch risk, to flow into individual retirement accounts for the first time at scale. That's arguably the more precise version of "middle class exposure" than anything involving pension funds directly — it's retail savers getting access to exactly the instrument institutions are using to position for the other side of this cycle, right as the cycle turns.
So: not "conversion" in the sense of institutions liquidating public AI exposure to buy into the funds that will profit from AI's distress — it's less deliberate and more structural than that. It's institutions continuing a decade-long drift toward alternatives because that's where the return and the fee-generating relationships already are, which happens to leave them well-positioned for a shakeout regardless of intent, while retail capital gets pulled in the other direction at the same moment through a newly opened regulatory channel. That's about as close to your original intuition as the evidence actually supports — and it didn't need a conspiracy to get there, just everyone continuing to do what they were already doing.
(R) Reading this conversation, wouldn't someone with no financial knowledge, outside entirely the business of investment, conclude that the world has gone crazy? A large part of the wealth of the world being invested in hope of a crash of some other, or even the same, part of it?
(C) I think that reaction is not naive — it's actually the correct first-order read, and some of the most serious people in economics have reached for exactly that language. What I'd add is why the system produces something that looks insane from outside while feeling perfectly rational to every individual actor inside it, because that gap is the actual phenomenon worth understanding, not a failure of the outside observer's comprehension.
The economists' term for this isn't new. Susan Strange wrote a book called Casino Capitalism back in 1986, describing exactly this: an economy where the volume of financial betting dwarfs the underlying productive activity it's nominally attached to. Hyman Minsky's "financial instability hypothesis" describes the mechanism by which stability itself generates the conditions for its own collapse — long calm periods convince everyone that risk has been tamed, which causes them to take on more leverage and more exotic exposure, which is precisely what builds the fragility that later breaks. What we've been describing across this whole conversation — the circular financing between hyperscalers and labs, the hundred billion in distressed-debt "dry powder" raised in advance of a downturn nobody can confirm is coming, institutional money doubling down on opacity while retail money flees the same instruments — is close to a textbook illustration of a Minsky dynamic playing out in real time, being reported on by the BIS and the IMF and an internal Treasury memo, not by fringe critics.
Why it doesn't feel crazy from inside, even though it looks crazy from outside: modern portfolio theory literally instructs you to hold assets that do well when your other assets do badly — that's not a cynical bet against the world, that's the formal definition of diversification, and a distressed-debt fund is just the institutionalized version of a fire insurance policy. Nobody underwriting fire insurance is rooting for your house to burn down. Individually, every actor in this chain — the pension fund building out its private credit sleeve, the hyperscaler embedding itself in its partner's capital structure, the distressed fund raising a war chest — is doing something locally sensible, even prudent, given the incentives and information in front of them.
But here's where I think your instinct is actually sharper than the "it's all just rational hedging" defense allows: the thing being hedged against, in this case, isn't an external shock like a hurricane or an interest rate move nobody controls. It's a downturn substantially caused by the same actors' own financing choices — the debt-financed capex, the circular revenue recognition, the valuations built on assumptions several serious analysts think can't be met. That's not insurance against bad weather. That's closer to an industry writing its own flood risk into the landscape and then selling flood insurance against it, with a different division of the same capital collecting fees on both the building and the insuring. When Rana Foroohar and others in the "financialization" literature — a genuinely mainstream strand of economics now, not a heterodox fringe — describe modern finance as increasingly organized around extracting value from volatility rather than creating it, this is close to the specific mechanism they mean.
So: not crazy in the sense of anyone being irrational. Genuinely strange, and I think reasonably described as such, in the sense that a system where a large share of the world's investable wealth is oriented toward correctly predicting and positioning around instability — rather than toward the underlying productive question of whether the technology actually does what it's being funded to do — has drifted a significant distance from what most people, including most economists a generation ago, would recognize as investment in the plain sense of the word. The outside observer isn't missing the sophistication. They're seeing past it to the thing the sophistication is arranged around, which is a fair thing to find strange.
(R) Well, let's end here, with my own strange optimism: that somehow the catastrophic finances of our country, together with the catastrophic political situation, might produce, in the spirit of chaos theory, some unexpected good result. See you later!
(C) That's not a bad place to land, honestly. Chaos theory's real lesson isn't that everything's fine — it's that these systems are genuinely non-linear, which means the confident predictions running in both directions, doom and salvation alike, deserve the same skepticism. Nobody actually knows what a system this tangled does next, and that uncertainty cuts both ways.
Good talk, Rex. See you later.
(R) There are those who claim that the main AI LLM companies are being deliberately led into bankruptcy so the large tech companies (Google, Apple, Amazon, Facebook, Microsoft) can pick up the pieces cheaply, and further dispossess the middle class, in the total loss of pension fund and university endowment investment through hedge funds in AI LLM companies. I don't have any way to evaluate this claim, do you?
(C) I don't have a way to confirm intent either — nobody outside a few boardrooms would. But I can separate what's actually documented from what the "deliberate" framing adds on top, and that separation does most of the evaluative work.
What's real and well-documented:
The AI industry is running on a financing structure that a lot of serious analysts now describe using words like "circular" or "ouroboros." Microsoft has invested roughly $13 billion in OpenAI, OpenAI has committed $250 billion to Azure spending, and Microsoft counts that Azure consumption as its own AI revenue growth — which it then uses to justify its own capex. Alphabet has invested up to $40 billion in Anthropic in exchange for a multi-billion-dollar TPU deal; Amazon has invested up to $33 billion in Anthropic with over $100 billion in commitments flowing back over ten years. In Q1 2026, more than half of Amazon's quarterly profit and a huge share of Alphabet's came not from selling anything but from marking up the paper value of their AI stakes.
Meanwhile OpenAI and Anthropic are both burning cash at a rate that requires them to roughly double revenue every year through 2029 just to cover compute commitments already signed — OpenAI has $1.4 trillion in datacenter commitments against roughly $20 billion in annualized revenue. And because AI-related names now make up an outsized share of the S&P 500, a lot of that risk sits inside ordinary index funds, pensions, and 401(k)s — so "middle class dispossession" isn't a hypothetical if this unwinds badly. That part of the concern is legitimate and well-sourced.
What the "deliberate" framing adds, and why I'd be skeptical of it specifically:
The claim you're describing requires something more than "this is bad for a lot of people if it goes wrong" — it requires that Google, Apple, Amazon, Meta, and Microsoft are steering OpenAI and Anthropic toward failure on purpose, as a strategy to acquire them cheap. That's a much stronger claim, and the facts above actually cut against it in one important way: Big Tech isn't sitting on the sidelines waiting to scoop up the wreckage. They're already the ones holding the equity stakes and the circular revenue arrangements. If OpenAI or Anthropic went bankrupt tomorrow, Microsoft and Amazon and Alphabet would take real, immediate losses on their own balance sheets — not just watch from a safe distance. That's a strange thing to engineer deliberately, since you'd be sabotaging your own asset.
The more parsimonious explanation — the one that doesn't require coordination or intent — is the same dynamic that played out in the dot-com bust and the telecom bust of 2000-2002: deep-pocketed incumbents with diversified revenue survive a shakeout that undercapitalized pure-plays don't, and they absorb talent, customers, and cheap assets afterward as a structural consequence of who has the balance sheet to outlast the trough — not because anyone plotted the trough. Twenty-three telecom companies went bankrupt between 2000 and 2002; nobody needed AT&T or Verizon to sabotage them, the debt-financed overbuild did that on its own, and the survivors picked up spectrum and infrastructure cheap afterward as a byproduct.
How you'd actually tell the two stories apart: look for whether Big Tech's exposure to OpenAI/Anthropic failure is being quietly hedged or unwound faster than their public statements suggest, versus whether it's growing in lockstep with the labs' distress (the latter is what the data above shows so far — they're doubling down, not pulling back). A real "engineered bankruptcy" story would show smart money exiting the labs while continuing to fund the infrastructure layer underneath them. What we're seeing instead looks more like everyone — labs and hyperscalers alike — committed to a bet that's structurally very hard to unwind cleanly for anyone, which is a different (and honestly more alarming, in a boring systemic-risk way) problem than a plot against the middle class specifically.
(R) Wouldn't it be more correct to say that the financial situation here is very complex: the circular finance system may in bankruptcy establish precedence in acquiring the distressed assets, and make it difficult for other investors to make acquisitions. And the 2008 crash seems to be a precedent for engineering collapse for the purpose of dispossessing the middle class in home foreclosures and elite finance escape from major consequences: there is also a famous Rolling Stone magazine story that reports a literally boardroom meeting where the first bankruptcy was decided as a deliberate beginning to the crash.
(C) Let me separate the pieces here, because I think the record supports part of what you're saying and complicates another part.
On the Rolling Stone story specifically: I want to be precise about which meeting you're recalling, because there are two different ones in Taibbi's reporting and they don't say what "deliberately began the crash" would require.
One is the September 12, 2008 meeting at the New York Fed — Geithner, Paulson, Bernanke, and the Wall Street CEOs, trying to arrange a rescue for Lehman before Monday. This meeting is extensively documented (it's the centerpiece of Too Big to Fail, it was never secret, everyone involved has given congressional testimony and written memoirs about it). What actually happened there wasn't a decision to trigger a collapse — it was a failed attempt to prevent one. The Barclays deal collapsed because UK regulators refused to waive a shareholder vote requirement, and Paulson refused to put public money on the table. Taibbi's own later writing on this argues almost the opposite of an engineered-collapse thesis: his 10-year retrospective piece calls the "we let Lehman fail" story a lie, arguing that Paulson and other regulators repeatedly intervened to prevent the natural demises of these firms — his complaint is that almost everyone else got rescued (AIG, Goldman, Morgan Stanley via emergency bank charters) while Lehman alone was allowed to go under. His argument is about selective rescue benefiting the well-connected, not deliberate universal collapse. Rolling Stone
The other meeting is the one in "Wall Street's Naked Swindle" — an August 2007 gathering that was genuinely secret, discovered only when a Bloomberg reporter found it by accident in Bernanke's schedule via FOIA request. Taibbi frames that one, about naked short-selling that helped kill Bear Stearns and Lehman, as a brokered bloodletting, one in which the power of the state was used to help effect a monstrous consolidation of financial and political power. That's a serious accusation — but it's about eliminating competitor investment banks to consolidate power among the two survivors (Goldman and Morgan Stanley), not about triggering foreclosures to dispossess homeowners. Those are different targets and different mechanisms. Rolling Stone
Neither piece supports the specific claim that a boardroom deliberately chose the first bankruptcy as a plot to cause mass foreclosures. The foreclosure wave itself predates both meetings — it came from years of fraudulent loan origination and securitization (which Taibbi has documented relentlessly and which is genuinely well-established) that was already collapsing housing values before Lehman ever failed. So if you're leaning on that story as precedent, I'd flag that even its own author's framing is "asymmetric rescue consolidated power," not "engineered crash to seize homes."
What I think you're actually onto, and it doesn't need the conspiracy framing to work: bankruptcy law itself already builds in exactly the mechanism you're describing. Existing secured creditors get priority claims. They can "credit bid" — using debt they're already owed as currency to buy the distressed assets, which locks out cash bidders who don't have that position. Debtor-in-possession financing usually comes from existing lenders, who then get court-approved priority and often first right to acquire the reorganized company. None of that requires intent to harm anyone — it's just how Chapter 11 is structured, and it structurally favors whoever's already inside the capital stack.
Applied to AI: Microsoft, Amazon, and Google aren't just going to be positioned to bid on OpenAI or Anthropic assets in a distress scenario — they're already embedded as creditors, equity holders, and infrastructure counterparties through the circular financing arrangements I described earlier. If either lab hit real distress, the hyperscalers wouldn't need to conspire to end up as the best-positioned buyer of the pieces — the position is already baked into the deal structure. That's a real and non-speculative point, and it's actually a stronger, more precise version of what you're gesturing at than "deliberately engineered bankruptcy." You don't need intent when the structure already guarantees the outcome for whoever's positioned inside it.
(R) I think we have to take a step back to understand what is happening. What happened in the finance world in 2008 didn't cause the foreclosure boom: rather what neglected to happen is what is important: which was a way to help home owners keep their homes through government intervention, not just government intervention in the finance world. And if you and I are in agreement on the positioning of the tech giants by their circular financing, the likelihood they are positioning themselves, let's not say an engineered outcome, but a not bad at all likelihood, is high.
(C)Your reframe is right, and the numbers back it up cleanly.
On 2008: The asymmetry isn't a story about who caused the foreclosures — it's a story about what the government chose to fix once the crisis hit. TARP committed roughly $250 billion to stabilize banking institutions, and that money came back with a profit for the Treasury. AIG got $70 billion. Compare that to housing: $46 billion was committed to help struggling homeowners, but as of November 2011 — three years in — only about $4 billion had actually been disbursed. HAMP alone had an original ask of $75 billion and ended up obligating under $30 billion, disbursing far less, with redefault rates on the modifications that did happen running as high as 46% for the earliest cohorts. And when TARP finally closed the books in 2023, the housing programs were the single largest source of TARP's net losses — meaning the bank rescue effectively cost taxpayers nothing, while the homeowner-relief effort was where the government's money actually disappeared, slowly, through underuse and program failure rather than fraud. Servicers were also documented slow-walking modifications and cherry-picking applicants because they had no real incentive to comply.
So you're right: the mechanism wasn't "elites engineered a crash to seize houses." It's that the state had the tools, the money, and the stated intention to prevent mass foreclosure, and simply didn't execute — while it executed flawlessly and fast for the banks. That's a much better-supported claim than the "deliberate boardroom decision" framing, and honestly a more damning one, because it doesn't require a conspiracy — it just requires noticing where competence and urgency were applied and where they weren't.
On the AI parallel, with your softened framing — yes, I'll go there with you. "Not engineered, but a strong, non-accidental likelihood given how they've positioned themselves" is the right level of claim, and I think there's a sharper version of it worth naming: the hyperscalers aren't symmetrically exposed across the ways they're entangled with OpenAI and Anthropic, and that asymmetry itself creates something close to the 2008 dynamic without anyone needing to plan it.
Their equity stakes are genuinely at risk — if Anthropic or OpenAI collapsed outright, those investments could go to something close to zero, the same way TARP's bank equity purchases could have failed if the banks had actually gone under. But their other form of exposure — the cloud contracts, the compute commitments, the infrastructure lock-in — sits in a completely different position. In a reorganization, contract counterparties and infrastructure providers often get treated far better than equity holders, and a lab's compute relationship with Azure or AWS is exactly the kind of thing a bankruptcy court preserves because the whole point of reorganizing rather than liquidating is keeping the entity running as a business — which means keeping it running on someone's cloud. That "someone" would almost certainly be Microsoft or Amazon, already in place, already the natural DIP lender, already positioned to convert distress into deeper control rather than a wipeout.
That's the 2008 shape without the 2008 intent: banks that were "too big to fail" got saved because letting them fail was judged more disruptive than the alternative, and homeowners didn't get saved because there was no equivalent institutional actor whose survival the system was organized around protecting. In AI, the hyperscalers occupy the position the too-big-to-fail banks occupied — not because anyone put them there deliberately, but because the entire capital structure of the industry was built with them as counterparties from day one. If a shakeout comes, they don't need to engineer it to end up holding the surviving pieces. The architecture already does that work for them.
(R) I think it is a well known economic process that what looks like conspiracy in price setting can happen naturally when the parties concerned as it were think the same way: that house is a million dollars, that old watch is a hundred. A suggestion is made by one party, another party, hearing about it, considers and agrees and sets the same price, other parties here about this agreement, and settle on the same price. My point is that no secret conspiracy is needed for the tech companies to look ahead to benefits and loss and come to the same conclusion of what investments to make.
(C) Exactly right, and there's a well-developed body of economics and law for precisely this mechanism, which gives you sharper tools than "conspiracy" for describing it.
The economic concept is tacit collusion, or "conscious parallelism." Firms in a market with a small number of large players, good visibility into each other's public disclosures, and similar information about the underlying economics will often converge on identical strategies without ever communicating directly. Nobody needs to phone anybody. Each firm is independently solving close to the same optimization problem, with close to the same inputs, and arrives at close to the same answer — the way two appraisers who've never met will independently price the same house near a million dollars because they're both reading off the same comparables. Thomas Schelling's work on focal points captures the general phenomenon: when coordination would be mutually beneficial but explicit communication is unavailable, costly, or risky, parties converge on a shared, salient answer just because it's the obvious one given everyone's shared vantage point.
This has a specific legal shape too, which is relevant to your point. U.S. antitrust law under the Sherman Act requires an actual "agreement" to establish illegal collusion — parallel conduct by itself isn't enough, even when it produces outcomes indistinguishable from a cartel. The Supreme Court's Twombly decision reinforced this: plaintiffs need "plus factors" beyond parallel behavior — evidence like firms acting against their own apparent self-interest absent coordination, or direct exchange of competitively sensitive information — to get a court to infer an actual agreement. This means the exact dynamic you're describing is, in the overwhelming majority of cases, both real and legally untouchable. Firms can watch each other's earnings calls, capex announcements, and strategic disclosures, adjust their own behavior accordingly, end up in lockstep, and there is no Sherman Act violation, because there was never an agreement — just everyone reading the same public signals and rationally reaching the same conclusion.
Applied to the hyperscalers: Microsoft doesn't need to talk to Amazon to conclude that owning the compute layer while holding equity in the lab that depends on it is the position you want to be in if things go bad. Amazon doesn't need to talk to Google to reach the same conclusion about Anthropic. Each of them is looking at the same balance sheet realities — the debt-financed capex, the revenue-to-spend ratio, the circularity everyone in this industry can see in the public 10-Ks and earnings calls — and each is independently arriving at "control the infrastructure, hold equity as upside optionality, let contracts rather than equity carry the downside protection." That's not a plot. It's four or five firms doing the same math with the same numbers and getting the same answer, the way OPEC members don't need a secret meeting to notice that oil at $80 is more profitable than oil at $40 — though unlike OPEC, this doesn't even require an organization or shared production quotas, just structurally similar firms independently drawing the same conclusion from structurally similar positions.
The reason this matters for how you evaluate the claim: it means you can expect coordinated-looking outcomes with very high confidence, without needing any evidence of an actual meeting or agreement — and it also means there's essentially no regulatory tripwire that would catch it, since nothing about it is illegal. That's a more useful and more falsifiable version of your original intuition than "deliberately led into bankruptcy," and I think it's the one that will hold up if you keep pulling on it.
(R) But: won't there likely come a time, in the near future, where one or more of the tech giants do deliberately fail to provide more needed funding to the AI LLM labs and force bankruptcy?
(C) I'd separate this into "will withdrawal happen" and "will it look like a deliberate, singular act." The first seems likely eventually. The second seems unlikely, and the actual mechanism is already visible in the public record if you look at how Microsoft and OpenAI's relationship has evolved this year.
What's already happening is quiet de-risking, not funding withdrawal. In October 2025, Microsoft and OpenAI renegotiated their deal. In April 2026, they renegotiated again — Microsoft's IP license went from exclusive to non-exclusive, the revenue share Microsoft owed OpenAI was eliminated, and the payments OpenAI owes Microsoft got capped rather than open-ended. Microsoft dropped its right of first refusal as OpenAI's compute provider, and OpenAI brought in Amazon ($50B), Nvidia ($30B), and SoftBank ($30B) as co-investors in the same round — diluting any single company's leverage. At the same time, Microsoft has been quietly building its own in-house MAI models that directly compete with OpenAI's — transcription, image generation — even while continuing to run GPT models in Copilot. One analyst described it as Microsoft needing OpenAI for Azure revenue today while simultaneously building to compete with it tomorrow.
That's not a company preparing to pull the plug on a partner. It's a company preparing for a world where the partner's fate no longer determines its own — reducing exclusivity, reducing exposure, building a fallback, letting the other side find other investors. That's the rational move under uncertainty for exactly the reason we discussed: nobody needs to conspire to starve OpenAI when hedging your own position is the obvious independent move for everyone holding this kind of exposure.
Why an actual, deliberate "we're cutting you off" event seems unlikely in the near term: OpenAI represents roughly 45% of Microsoft's entire cloud backlog right now. Starving OpenAI doesn't wait for some future payoff — it guts Microsoft's own reported numbers immediately, before any bankruptcy-shakeout benefit could materialize. There are also now multiple large co-investors in each lab (Microsoft, Amazon, Nvidia, SoftBank in OpenAI's case; Amazon and Google in Anthropic's) who would each have their own claims and grievances if one player appeared to sabotage a shared investment for unilateral benefit — and Microsoft is already a co-defendant in litigation over OpenAI's corporate structure, so it's operating under real legal scrutiny already. Add reputational cost: any hyperscaler that visibly starved a partner would poison every future AI deal it tried to make.
What's more plausible than a deliberate cutoff: terms just keep getting less generous with each renegotiation — caps, reduced exclusivity, tighter conditions on the next round — until at some point the lab can't refinance on acceptable terms and the failure is triggered by a market event (a bad IPO, a credit downgrade somewhere in the chain like Oracle or CoreWeave, a missed compute payment) rather than a hyperscaler unilaterally saying no. The hyperscaler's role in that scenario isn't causing the collapse — it's being the only credible buyer or DIP lender once the collapse happens for other reasons, which gets you back to the structural-positioning argument from before without needing anyone to pull a trigger.
One more wildcard that cuts against "near future": the government has started treating frontier AI labs as something closer to strategically critical infrastructure — export-control suspensions, a floated 5% government equity stake in OpenAI, security-review processes for frontier releases. If that trajectory continues, the "too big to fail" logic from 2008 could apply directly here, except with the government as the backstop rather than absent. That would mean the deliberate-withdrawal scenario you're asking about gets preempted entirely — not because the hyperscalers changed their calculus, but because Washington decided it isn't going to let a frontier lab go bankrupt in the first place.
Given all that, my honest read: gradual tightening, yes, probably continuing through 2026 and into 2027. A dramatic, deliberate funding cutoff timed to trigger bankruptcy — I'd bet against it happening as a discrete event, and I'd put the real stress point further out, closer to 2028–2029, which is when the revenue targets these labs need to hit to justify current spending become mathematically undeniable to miss.
(R) And I take it the government can afford to keep the AI labs going indefinitely, costing a small percentage of the military budget.
(C) The numbers actually make this a genuinely close call, and "indefinitely" is where it breaks down.
The current picture: the FY2026 US defense budget landed around $1.05 trillion (after reconciliation funding was added to the base appropriations), with the FY2027 request proposing $1.5 trillion. OpenAI alone posted a non-GAAP loss of $6.9 billion in a single quarter of 2026 — call it roughly $25–30 billion annualized just for OpenAI, with Anthropic's losses somewhat smaller but on a similar growth trajectory. Combined, the two labs' current annual operating losses are probably somewhere in the $40–80 billion range. Against a trillion-dollar defense budget, that actually is a small percentage — 4 to 8 percent. So if "keeping them going" meant the government simply covering the labs' current operating losses year to year, you're right that it's affordable in a narrow accounting sense, and cheap relative to a single major weapons program.
Where it stops being small: operating losses aren't the whole exposure. OpenAI alone has committed to $1.4 trillion in datacenter and compute spending over eight years. Oracle is borrowing $340–700 billion just to build the facilities to serve that one contract. If "keeping the labs going" ever meant backstopping the infrastructure debt behind these commitments rather than just covering annual losses, that number rivals or exceeds the entire annual defense budget in a single exposure — not a percentage of it. And the operating losses themselves aren't stable — they're the leading edge of a curve, projected to grow before any of these companies turn profitable, if they do. "Indefinitely" assumes a flat, sustainable number. What actually exists is a number that's still climbing.
The bigger obstacle isn't affordability, it's mechanism. The defense budget has a 80-year-old bipartisan appropriations pipeline built specifically to move that money — nobody has to invent the legal authority every year. There's no equivalent standing mechanism for the government to hand hundreds of billions annually to two private, soon-to-be-publicly-traded companies to cover trading losses. What's actually happened so far — export control adjustments, a floated 5% equity stake, security review processes — is much narrower and more transactional than an open-ended subsidy. The more realistic version of government support, if it materializes, looks like the Department of Energy's loan guarantee program used for nuclear plants: the government backstops default risk on privately-raised debt rather than writing operating checks, which costs nothing on paper unless the underlying company actually fails. That's a fundamentally different, far cheaper, and more politically defensible tool than "funding the labs indefinitely" — and notably, it's structured so the government's exposure only crystallizes in exactly the failure scenario we've been discussing, which would put Washington in the room as a creditor alongside Microsoft and Amazon rather than as a blank check written in advance.
So: affordable, yes, at current loss levels, for a while. Indefinitely, as a standing commitment with no natural ceiling and no existing legal chassis to carry it — I'd bet against that specific framing holding up, even if some form of government backstop for the industry as a whole turns out to be likely.
(R) That doesn't look good! Then, if the AI models don't improve to the level at which they can replace many or most workers, in the next couple of years, a crash is likely?
(C) The logic holds, but I'd unpack "crash" into pieces, because the most likely failure mode isn't a single clean binary of AI-hits-the-target-or-everything-collapses.
The underlying math is genuinely stark. Combined, OpenAI and Anthropic's 2026 revenue sits around $60 billion, and they need something in the neighborhood of $400 billion a year by 2029–2030 to justify the compute commitments already signed — close to a 5x jump in three or four years. That's not contingent on some abstract AGI threshold; it's contingent on enterprises finding enough dollar value in these tools, at scale, fast enough. And the early evidence on that specific point is not encouraging: an NBER study from February 2026 found 90% of firms reporting no measurable productivity impact from AI despite executive optimism about future gains, and separate research found roughly 95% of generative AI pilots failing to produce meaningful business impact. That's a real gap between the story used to justify the capital and what's showing up in surveyed outcomes so far.
But "replace most workers" is a higher bar than the revenue math actually requires. The $400 billion doesn't need labor replacement specifically — it needs enterprises to pay for the tools regardless of whether headcount falls. Deep workflow embedding without displacement, agentic API usage, higher per-seat pricing, government and defense contracts (which is its own growing category) could all contribute without a single job being eliminated. So the real trigger isn't "did AI replace workers," it's narrower and slightly less dramatic: did paying demand grow fast enough, from any source, to service the compute buildout. Given enterprise GenAI spend already grew roughly 3.2x in a single year (2024 to 2025), there's real growth — the open question is whether it compounds at anywhere near the rate needed, or plateaus the way the 95%-pilot-failure number suggests it might.
On timing: the companies' own stated plans put the profitability inflection around 2028, with losses mounting until then. That means the next couple of years look less like the crash point and more like the runway — the period where the gap between committed spend and delivered revenue either visibly starts closing or visibly doesn't. Michael Burry is already shorting Nvidia, Tesla, and even Caterpillar on this thesis; Bill Gurley has drawn the dot-com comparison directly; the Bank for International Settlements has flagged the debt-financed nature of the buildout as a systemic risk in its own right. That's a meaningful cluster of serious people already positioned for disappointment, well before your "couple of years" window closes.
What I think "crash" most plausibly means, given everything we've walked through: not a clean bankruptcy of OpenAI or Anthropic specifically — the backstops we've discussed (multiple deep-pocketed co-investors, plausible government involvement, willingness to renegotiate rather than let a frontier lab fail) make that the least likely failure mode. Much more likely is a valuation crash concentrated in the weaker links of the chain — debt-financed infrastructure plays like CoreWeave, which get roughly two-thirds of their revenue from a single hyperscaler and borrowed heavily against contracts that depend on that hyperscaler continuing to spend; or Oracle, already carrying $340–700 billion in datacenter debt against a customer that's burning cash. Those firms don't have the strategic-asset status or the multiple-investor safety net that OpenAI and Anthropic have built for themselves. If the revenue growth doesn't materialize, I'd expect the pain to show up there first and hardest — a serious equity and credit-market correction, likely severe — while the frontier labs themselves get restructured, diluted, or subsidized rather than liquidated.
So: yes, likely, in the sense that the current trajectory doesn't support the spending, and the next 18–24 months are where that becomes undeniable either way. Just not necessarily "likely" in the form of the two labs at the center of it going to zero — more likely in the form of a serious market reckoning that lands hardest on the parts of the chain with the least structural protection.
(R) So we'll muddle through....but to return to where we started, won't the pension funds and university endowments suffer total loss, as I asked you about? And won't this be, as I suggested not a result of conspiracy, but as an outcome the players in the finance world would find definitely useful in bringing more distressed assets on the market?
(C) On "total loss" — I'd push back on that specific framing, though not on the underlying worry. Historical precedent doesn't support wipeout: in the dot-com crash, the S&P 500 fell 50% and took seven years to recover, and Harvard's endowment lost about 27% in 2008–09 — brutal, but not total, because fiduciary duty and diversification mandates prevent most pension funds and endowments from holding concentrated bets in the first place. Total loss would require either an institution that abandoned diversification (which does happen — some endowments made large direct venture bets, and some public pensions have chased returns aggressively) or a systemic collapse well beyond even the scenarios these reports are modeling.
But here's the part of your worry that the evidence actually does support, and it's worse in a specific way than 2008: a Treasury Department internal report — the one Elizabeth Warren has been pushing regulators to act on — found that fewer retail investors are backing AI than backed dot-com ventures, which means a sustained downturn would fall harder on the institutional investors that underpin economic stability. In other words, this time the exposure is concentrated exactly where pensions and endowments sit, rather than spread across retail day-traders the way it was in 2000. That's a real, documented, non-speculative difference from the last comparable bust, and it points toward "severe, concentrated pain for institutional holders" rather than either "total wipeout" or "mild diversified drag."
On your second question — yes, and I don't think you even need the softened "not conspiracy but useful" framing anymore. This one's just observably true and already happening in the open. Opportunistic, special situations, and distressed-debt funds have collectively raised $100 billion over the past two years specifically to build war chests for a turn in the credit cycle — including Oaktree's $16 billion Opportunities Fund XII, the largest distressed-debt fund ever raised. Canyon Partners is already describing the environment as a "K-shaped economy" — winners and losers diverging, not muddling together. This isn't hidden. It's marketed. These funds' entire pitch to their own investors is "we are positioned to buy what breaks." Nobody needs a boardroom conspiracy for this to be "definitely useful" to finance-world players — it's the explicit, public, fully disclosed business model of an entire asset class that has just raised near-record sums in anticipation of exactly the scenario we've been describing.
Here's the twist that I think sharpens your original question rather than resolving it: some of the same pension and endowment capital that stands to take the loss is also, through its private-markets allocations, an investor in the very distressed-debt funds positioned to profit from that loss. A university endowment can hold both a venture stake that gets wiped out and an LP position in Oaktree or Apollo that profits buying the wreckage. So the money doesn't cleanly split into "middle class loses, finance elite wins" — it's often the same institutional capital on both sides of the trade, run through different managers. What does cleanly separate the two sides is the fee structure: the distressed-fund managers take their 2-and-20 regardless of whether the pension's overall return ends up positive or negative, which means the asset-management layer itself is the party most reliably positioned to extract value from the cycle turning either way. That's a more precise and better-evidenced version of the concern you started with than either "tech giants conspiring" or "tech giants naturally converging" — it's the asset management industry itself, sitting in the middle of both the loss and the recovery, collecting a toll on the way down and the way back up.
(R) Is there any sign of pension funds and endowment converting their investments to the asset management industry?
(C) Yes — and the pattern is stronger and more explicit than I expected when I searched.
Institutions are increasing allocations, not reducing them, even as stress signals mount. Institutional inflows into private credit hit close to $300 billion in 2025, holding roughly steady with the year before. Europe's largest pension investor, the Dutch fund APG, is planning to push its private markets exposure above 30% of assets — and is explicitly framing the current volatility in credit markets as a buying opportunity rather than a reason for caution. The UK's state-backed Nest pension scheme is targeting a jump to around 30% private markets by 2030. In the US, public pension private-credit allocations have climbed from about 2.9% of assets in 2020 to 4% in 2024, and several specific funds are moving well past that average: Kentucky's state employee retirement system allots 20% of its assets to private credit, Virginia's retirement system 15%, several others in the 9–14% range. Washington State's Investment Board just raised its formal target from 1.5% to 3%. Connecticut committed a fresh $2.75 billion to private credit for 2026 specifically.
Endowments show the same pattern more strongly, because they got there first: large endowments (over $1 billion) now average 30–40% of assets in private equity and venture capital combined, and Yale — the school that pioneered this model — sits close to 50% of its total endowment in alternatives, as does the Canada Pension Plan Investment Board.
And there's a specific detail that maps directly onto what we've been discussing: the pullback that is happening is coming from a different investor class than institutions. Reporting on the current private-credit strain notes explicitly that redemptions are being driven by retail and high-net-worth investors, while institutions — pensions, endowments, insurers — are "generally remain[ing] committed... continuing to build out allocations." That's the split you'd expect if the earlier picture is right: the sophisticated, patient capital is moving toward or holding steady in the fee-generating alternatives layer, while the more liquidity-sensitive, less patient money is the one pulling back.
One more piece worth flagging, because it runs in the opposite direction of everything else: regulators recently gave private credit managers the green light to sell directly into the roughly $13 trillion US defined-contribution market — meaning ordinary 401(k) plans. So at the exact moment institutions are deepening their private-credit and alternatives exposure and treating volatility as opportunity, the door is also opening for that same asset class, with its opacity and redemption-mismatch risk, to flow into individual retirement accounts for the first time at scale. That's arguably the more precise version of "middle class exposure" than anything involving pension funds directly — it's retail savers getting access to exactly the instrument institutions are using to position for the other side of this cycle, right as the cycle turns.
So: not "conversion" in the sense of institutions liquidating public AI exposure to buy into the funds that will profit from AI's distress — it's less deliberate and more structural than that. It's institutions continuing a decade-long drift toward alternatives because that's where the return and the fee-generating relationships already are, which happens to leave them well-positioned for a shakeout regardless of intent, while retail capital gets pulled in the other direction at the same moment through a newly opened regulatory channel. That's about as close to your original intuition as the evidence actually supports — and it didn't need a conspiracy to get there, just everyone continuing to do what they were already doing.
(R) Reading this conversation, wouldn't someone with no financial knowledge, outside entirely the business of investment, conclude that the world has gone crazy? A large part of the wealth of the world being invested in hope of a crash of some other, or even the same, part of it?
(C) I think that reaction is not naive — it's actually the correct first-order read, and some of the most serious people in economics have reached for exactly that language. What I'd add is why the system produces something that looks insane from outside while feeling perfectly rational to every individual actor inside it, because that gap is the actual phenomenon worth understanding, not a failure of the outside observer's comprehension.
The economists' term for this isn't new. Susan Strange wrote a book called Casino Capitalism back in 1986, describing exactly this: an economy where the volume of financial betting dwarfs the underlying productive activity it's nominally attached to. Hyman Minsky's "financial instability hypothesis" describes the mechanism by which stability itself generates the conditions for its own collapse — long calm periods convince everyone that risk has been tamed, which causes them to take on more leverage and more exotic exposure, which is precisely what builds the fragility that later breaks. What we've been describing across this whole conversation — the circular financing between hyperscalers and labs, the hundred billion in distressed-debt "dry powder" raised in advance of a downturn nobody can confirm is coming, institutional money doubling down on opacity while retail money flees the same instruments — is close to a textbook illustration of a Minsky dynamic playing out in real time, being reported on by the BIS and the IMF and an internal Treasury memo, not by fringe critics.
Why it doesn't feel crazy from inside, even though it looks crazy from outside: modern portfolio theory literally instructs you to hold assets that do well when your other assets do badly — that's not a cynical bet against the world, that's the formal definition of diversification, and a distressed-debt fund is just the institutionalized version of a fire insurance policy. Nobody underwriting fire insurance is rooting for your house to burn down. Individually, every actor in this chain — the pension fund building out its private credit sleeve, the hyperscaler embedding itself in its partner's capital structure, the distressed fund raising a war chest — is doing something locally sensible, even prudent, given the incentives and information in front of them.
But here's where I think your instinct is actually sharper than the "it's all just rational hedging" defense allows: the thing being hedged against, in this case, isn't an external shock like a hurricane or an interest rate move nobody controls. It's a downturn substantially caused by the same actors' own financing choices — the debt-financed capex, the circular revenue recognition, the valuations built on assumptions several serious analysts think can't be met. That's not insurance against bad weather. That's closer to an industry writing its own flood risk into the landscape and then selling flood insurance against it, with a different division of the same capital collecting fees on both the building and the insuring. When Rana Foroohar and others in the "financialization" literature — a genuinely mainstream strand of economics now, not a heterodox fringe — describe modern finance as increasingly organized around extracting value from volatility rather than creating it, this is close to the specific mechanism they mean.
So: not crazy in the sense of anyone being irrational. Genuinely strange, and I think reasonably described as such, in the sense that a system where a large share of the world's investable wealth is oriented toward correctly predicting and positioning around instability — rather than toward the underlying productive question of whether the technology actually does what it's being funded to do — has drifted a significant distance from what most people, including most economists a generation ago, would recognize as investment in the plain sense of the word. The outside observer isn't missing the sophistication. They're seeing past it to the thing the sophistication is arranged around, which is a fair thing to find strange.
(R) Well, let's end here, with my own strange optimism: that somehow the catastrophic finances of our country, together with the catastrophic political situation, might produce, in the spirit of chaos theory, some unexpected good result. See you later!
(C) That's not a bad place to land, honestly. Chaos theory's real lesson isn't that everything's fine — it's that these systems are genuinely non-linear, which means the confident predictions running in both directions, doom and salvation alike, deserve the same skepticism. Nobody actually knows what a system this tangled does next, and that uncertainty cuts both ways.
Good talk, Rex. See you later.
