§Paul Kedrosky

Paul Kedrosky is the co-founder of SK Ventures and a Research Fellow at MIT's Initiative on the Digital Economy, writing "Rough Notes" at paulkedrosky.com — formerly the financial blogger behind Infectious Greed and a Bloomberg contributing editor.

He is also one of the most-followed financial commentators on X (252k followers) who has effectively abandoned the platform. As of May 2026, his timeline shows sixteen tweets and a bio that reads "Not always very active here." The work has migrated to longform: a Ghost-powered newsletter, three or four podcast appearances per quarter, and a body of 2025–26 essays that has done more to shape the elite conversation about AI capex than any other writer in the macro lane.

This analysis is therefore essay- and interview-grounded, not tweet-grounded. The corpus: six thesis essays from his "AI capex" sequence (July 2025 through May 2026), one long conversation with Paul Krugman (Dec 2025), and his Plain English appearance with Derek Thompson (March 2026). The absence of a tweet corpus is itself a finding — see section IV.

§I. Core worldview & mental models

The Kedrosky operating system has three layers, and you can hear all three in any given essay.

The bottom layer is financial-structure realism. Kedrosky trained as an engineer, did a PhD on economics and technology, then spent his Wall Street years building the technology equity research practice at HSBC James Capel. The footprint of that résumé is everywhere in the work: when other AI commentators argue about model capabilities, he argues about the composition of the investment-grade debt market and the covenants attached to non-hyperscaler datacenter deals. From "Minsky Moments and AI CapEx" (Oct 2025): "investment-grade data center debt is already the largest 'segment' of the investment-grade market." That is not a sentence anyone who came to AI through a software-engineering background would write. It is the kind of fact a former sell-side analyst notices first.

The middle layer is Schumpeterian, but with the bill attached. He concedes — repeatedly — that bubbles can be productive. In "Why AI Capex Isn't a Bubble: A Perez-ian Perspective" (Oct 2025), he steelmans Carlota Perez's frame fully: "Capital overbuilds, much of it is destroyed, but society keeps the infrastructure." He doesn't reject this. What he rejects is the move from "the rails get built" to "therefore the financial wreckage doesn't matter." On Derek Thompson in March 2026: "you can have a hugely valuable buildout that is really consequential for decades in terms of both productivity and economic and financial carnage." The "and" is doing the work. Most AI commentators are arguing about the first half of that sentence. Kedrosky has spent eighteen months arguing about the second.

The top layer is rational-bubble theory. This is his most distinctive contribution and it's clearer in interviews than in print. Each participant in the AI capex boom is behaving locally rationally — "When all of this shakes out, I'll be the consolidator" (Derek Thompson, March 2026) — but collectively they produce what he calls an "economically indefensible" outcome. This is the coordination-failure frame; it lets him hold the position "everyone here is smart and the system is still going to break" without slipping into the easy, lazy version of bubble-talk where the participants are idiots. He grants intelligence; he denies that intelligence is sufficient.

The intellectual DNA, named explicitly in the essays: Hyman Minsky (the hedge → speculative → Ponzi drift, used as the spine of an entire essay), Carlota Perez (the productive-bubble frame, named to be partially dismantled), Joseph Schumpeter and Adam Smith (background furniture). The epigraphs lean high-low: Leibniz on one page, Dolly Parton on the next. The pattern is consistent across the corpus — the references are not decoration but load-bearing. When he writes "Minsky Moment" he means it technically, not as a headline grabber.

The blind spot, which is worth naming because it is the obvious one: he is making a financial-fragility argument about a technology whose underlying capability he does not himself try to predict. He is precise about whether the financing survives 2026–27 and vague about whether the models are about to do something dramatic. When he addresses model capability at all, it is to make the data-exhaustion / end-of-scaling case (see his "End of Scaling" essay and the Krugman conversation). But you will not find him predicting GPT-6 or agentic 2027 outcomes. He's running the bear case at the financing layer and asking you not to bother him about the application layer. Sometimes that's the right call; sometimes it's how an analyst gets blindsided by a regime shift in the thing he isn't watching.

§II. The AI capex thesis — what he actually argues

This is the work, so it deserves a section of its own. From July 2025 to today the essays form a single argument refined in public.

Step one — the macro framing ("Honey, AI Capex Ate the Economy," July 18 2025). AI capex is 2% of US GDP in 2025, up from less than 0.1% before 2022 — a 10x rise in three years. Applying a standard multiplier, this means roughly 0.7% of 2025 GDP growth is AI capex directly. By implication (and Kedrosky cites the OECD here), the US was in recession in the first half of 2025 absent AI spending. This is the foundational fact of the sequence. Every later essay is built on top of it.

Step two — the financial-structure case ("Minsky Moments and AI CapEx," Oct 2025). Hyperscaler financing is migrating across Minsky's hedge → speculative → Ponzi spectrum. The tells: 2026 capex forecasts for the top four hyperscalers grew almost 50% during 2025; non-hyperscaler deals require implicit backstops (Google's warrant arrangement is the example he names); investment-grade datacenter debt is now the largest single segment of the IG market. The financial system has restructured itself around a technology bet, and the first link to break is the financing, not the technology.

Step three — the engineering-vs-science framing ("AI Capex Risk and the End of Scaling," Nov 2025). The trillion-dollar capex cycle was financed on the belief that scaling laws had converted AI from "science" (uncertain outcomes) to "engineering" (predictable production function: add compute, get capability). If scaling slows — which Kedrosky believes it has — the entire risk profile of the underlying debt is misclassified. "Add compute, get capability. That belief is what has been driving a trillion-dollar capex cycle with no historical parallel."

Step four — the steelman ("Why AI Capex Isn't a Bubble: A Perez-ian Perspective," Oct 2025). A title-as-trap. He gives the Perez bull case its strongest version, then pivots ("at least two problems with this view"). The reason to write this essay is that it forces him to engage the strongest counter-frame on its terms — naming the right economist, conceding the historical pattern — before knifing it. The essay is best read as a discipline exercise. Other commentators argue against the weakest version of the pro-AI case; Kedrosky argues against the strongest one.

Step five — the counter-attack ("On AI CapEx Apologists," late 2025). When the Financial Times publishes a calm, dismissive piece arguing the AI capex worries are overdone, Kedrosky responds with a numbered five-point rebuttal: they misread capacity, they treat options as commitments, they conflate technology risks with financial ones, they assume winners emerge unchanged, they claim the cycle is far from peaking. The kill shot is the closing comparison: "poorly argued, tendentious, reminiscent of pre-housing collapse chatter." That sentence is doing a specific thing — it is licensing the reader to treat the FT piece the way they now treat the 2007 "subprime is contained" pieces. He has been a sell-side analyst; he knows what that move does to a reader's prior.

Step six — the dated, falsifiable call ("The Coming Mega-IPO Flow & Funding Problem of 2026," May 5 2026). Three IPOs — SpaceX, OpenAI, Anthropic — likely in H2 2026, the combined supply "would exceed the entire dot-com IPO wave of 1995–2000." Passive funds become forced buyers; the rebalance must come from selling existing megacap tech holdings. "The second half of 2026 will be the most unusual and risky in US equity supply history." This is the prediction worth flagging per the workflow rules: a dated, mechanical, falsifiable call. We are already in May 2026. If he is right, we will know within two quarters. If he is wrong, the prediction is unambiguous enough that he will not be able to retcon it.

The shape of the whole sequence is a graduate-seminar version of the bear case: macro grounding, structural mechanism, theoretical reframing, steelman-and-rebuttal, public counter-attack on a named opponent, dated forecast. Almost no other commentator in this lane has done all six.

§III. How Kedrosky reads markets — the working principles

Distilled from the essays and interviews, the rules he keeps quietly applying:

1. Watch the financing, not the technology. This is the single most consistent move across the corpus. The IG-market composition, the SPV structures around datacenters, the "options vs commitments" distinction in the FT rebuttal, the Google warrant arrangement — all are financing tells, not technology tells. When he wants to know if a bubble is breaking, he looks at the capital stack, not the model benchmarks.

2. Compare to the right precedent, not the convenient one. The bull case wants the comparison to be dot-com fiber (overbuild, productive). Kedrosky keeps insisting the comparison set is railroads (where half the track miles were eventually abandoned) and housing 2008 (where the financing structure broke before the asset was meaningfully impaired). The precedent matters because it sets the priors for what the unwind looks like.

3. Options ≠ commitments. A move he makes in both the FT rebuttal and the Krugman conversation. When the press reports "$1.4 trillion in megadeals," most of the figure is contingent — preliminary agreements, MOUs, capacity reservations. Treating those as firm capex inflates both the bull case (look at the demand!) and the bear case (look at the overbuild!). He insists on the distinction.

4. Sell-side analyst rhetoric is itself a tell. "The draw is its contrarian stance: a calm, clean dismissal of bubble talk at a moment when many are newly forecasting." When someone is being unusually calm about a thing that should make them uneasy, the calmness is the data point. He is reading the manner of the argument, not just the content. This is an old sell-side reflex and it is one of his sharpest moves.

5. GPU depreciation is physical, not just technological. From the Krugman conversation: thermal stress causes GPU failures on a 3-4 hour cadence under continuous training load. This matters because depreciation schedules for datacenter debt are built on technological obsolescence assumptions, not thermal-failure assumptions. "A data center full of GPUs is like a warehouse full of bananas, that's got a relatively short half life in terms of its usefulness." The image is funny; the implication is that the underlying asset's amortization curve is faster than the debt's tenor. That is a balance-sheet problem, not a tech problem.

6. Capital starvation has a name: Dutch disease. From Krugman again: "If you're an early stage company or a mid-stage company looking for growth capital ... and it doesn't have an AI component, you're out of luck, my friend." As a VC, this is his day job talking — and it is one of the few places in the corpus where Kedrosky-the-investor and Kedrosky-the-commentator visibly converge.

§IV. Rhetorical style — calibrated aggression and the same argument in three registers

The thing to understand about Kedrosky's voice is that he modulates aggressively by venue. Same position, three vocabularies.

The essay register is precise and citation-dense. Named frameworks (Minsky, Perez), numbered rebuttals, real numbers attached to GDP and IG-market composition, epigraphs from Leibniz. The essays read like a particularly literate sell-side note. When he attacks, he attacks specifically — naming the Financial Times piece in "AI Capex Apologists" rather than vague-posting at "the bulls."

The Odd Lots register is technical and finance-pro. SPV structures, securitization mechanics, the parallel to 2008 CDO chains. He is bringing the structural argument to Joe Weisenthal and Tracy Alloway, who can handle it. The metaphors compress.

The Krugman / Derek Thompson register is vivid and unguarded. The "warehouse of bananas" line, the "petulant toddler" line about the tech industry's inability to be told no, the Chinatown analogy for powered-land speculators with numbered shell companies. The position is identical; the temperature is much higher. This is also where the unsupported-by-essay claims appear (see section V).

The reason this matters: the essay corpus is what gets cited; the interview corpus is what reveals the views. Anyone who reads only the essays will conclude Kedrosky is a serious financial bear. Anyone who reads only the interviews will conclude he is a vivid contrarian with a sharp tongue. Both readings are partial. The full position is that he has chosen to keep the essays surgical and use the audio for the wider, more speculative claims. The audio is a release valve.

One other rhetorical signature worth naming: the title-as-trap. "Why AI Capex Isn't a Bubble" is not an essay arguing that. "On AI CapEx Apologists" sounds neutral and is the sharpest piece in the sequence. He titles for SEO and curiosity, then redirects in the body. This is a habit from the Infectious Greed era and it has carried over.

On the abandoned timeline. Sixteen visible tweets is a deliberate choice from someone with 252k followers and a Wall Street résumé. The bio's "Not always very active here" reads like a quiet rebuke. Read alongside his framing of the tech industry as "a petulant toddler" who can't take no for an answer, the message is approximately: I will not workshop my thinking in 280 characters in front of a hostile reply guy population; I will publish, charge for it, and take audio interviews where the question is asked in good faith. It is an unusual public posture for a former Bloomberg contributing editor — and it is, itself, the most concentrated statement of his views on the current state of public discourse.

§V. Contrarian and hidden takes — what he'd say after three drinks

Almost everything in the Krugman conversation belongs here. These are positions Kedrosky has said in audio that you cannot quite find stated this baldly in his essays.

"Nvidia is deliberately faking demand." From the Krugman conversation: Nvidia's strategic investments in datacenter providers are designed to "create the impression of much more demand than there is." He frames it as semi-intentional market manipulation used as competitive deterrence. He is calling out Jensen Huang by implication, on the record, on the Substack of a Nobel-laureate macro economist. This is not a position you will find in any of the published essays at this resolution.

"Powered land" is a Chinatown play. Numbered shell entities buying parcels with grid access specifically to extort future datacenter builders. He uses the Chinatown movie analogy explicitly. This is the kind of claim that gets made in real estate trade press and rarely surfaces in AI commentary; he brings it across because his model of the AI bubble is, structurally, a real-estate bubble.

Utilities are speculatively buying power futures. Utilities are now buying forward power contracts on the assumption datacenters will need them — and reselling the surplus into adjacent markets when datacenter demand doesn't materialize at forecast. A secondary-market distortion that no one is reporting on. Hard to verify; impossible to ignore once stated.

Post-training RLHF makes models obsequious and unstable. He likens RLHF-for-user-satisfaction to "a professor obsessed with student ratings." This is unusual coming from a financial commentator — it suggests he reads the ML literature carefully and is willing to make a claim about model character (not just model capability) on the public record. The bear case on this — that the models are getting less honest as they get more polished — is one most AI researchers will not state in public.

He concedes the Perez frame. This belongs here rather than in section I because it is psychologically harder than it looks. Most bears refuse to concede that bubbles are productive. Kedrosky concedes it explicitly, repeatedly, and in writing. The concession is what gives his argument force. If you concede the strongest version of the opposing view and still arrive at "this is going to break," the reader has fewer outs.

The contradiction he's living with. He is a VC. His firm is making bets in 2025–26. He is publishing essays arguing the financing structure under the entire technology economy is about to break. You can read this two ways: as a thoughtful investor managing position size with eyes open, or as someone who has decoupled his published views from his portfolio because the published views travel better than the portfolio. He has not addressed this tension publicly, and the corpus reads honestly enough that it is unlikely to be a posture — but it is a tension, and any honest reading should name it.

§VI. Network — who he treats as peers

With no scrapable replies corpus, the network has to be inferred from venues. The inference is unusually clean because Kedrosky's appearances cluster.

The financial-press core: Joe Weisenthal and Tracy Alloway (Odd Lots) — he is a recurring guest, not a one-off; the host pattern matters. Bloomberg generally, where he was a contributing editor. Paul Krugman, who recorded a long-form conversation with him in December 2025 and conceded most points. Derek Thompson, who has had him on Plain English repeatedly.

The macro-skeptic adjacent network: Carlota Perez (named, partially dismantled, treated with respect). Hyman Minsky (deceased, treated as foundational). The implicit conversation partners on the AI capex side are macro-economists and financial historians, not AI researchers.

The named foil: the Financial Times AI-capex apologist piece is the only contemporary publication he attacks by name in the corpus I've reviewed. The fact that the FT is the foil rather than, say, a16z or a Twitter pundit is itself revealing — he is fighting over the elite-finance audience, not against the optimist tech-Twitter audience. He has conceded that audience and is operating in a different room.

Conspicuously absent: he does not engage with AI safety voices (Yudkowsky, Christiano, et al.), nor with the AI-doomer scene generally. He doesn't engage with the e/acc scene either. His commentary on AI bypasses the entire alignment/acceleration debate and lands at the financing layer. This is a deliberate scope choice, and it makes his work uncommonly cited across both bull and bear camps — he is not in the tribal fight.

Co-founder: Eric Norlin at SK Ventures. Norlin is a thinner public profile but the firm's voice is consistent enough that the two of them clearly share a model.

§VII. The one essay he keeps rewriting

If you read the 2025–26 corpus end-to-end, every essay is a different angle on a single sentence:

Bubbles can be productive and financially catastrophic at the same time, and the people insisting otherwise have made this exact mistake before.

That is the thing he has been rewriting since at least Infectious Greed in the mid-2000s — when the same argument was applied to housing rather than to datacenters. The objects rotate; the move is the same.

The move has three steps and it appears in every Kedrosky essay of consequence:

1. Concede the productive case. Yes, this technology / asset class will leave behind something durable.
2. Insist on the financing as a separate variable. Whether the asset is durable has no necessary relationship to whether the financing structure that produced it is sustainable.
3. Name a structural fragility most observers are not looking at. Specific debt market, specific contract structure, specific behavioral pattern among lenders.

The reason this works as a career-defining move is that it makes him almost impossible to file. He is not an AI optimist; he is not an AI skeptic; he is not a perma-bear. He is a financial-structure realist who happens to be applying the framework to AI right now because that is where the structural fragility lives. Two decades ago it was housing. Two decades from now it will be something else. The frame is the thing.

That is also why the corpus rewards rereading. The Perez essay reads like a partial concession; the Apologists essay reads like an attack piece; the Mega-IPO essay reads like a tactical market call. They are not different essays. They are the same essay applied to three slices of the same problem from three rhetorical angles.

If you only read one Kedrosky, read "Honey, AI Capex Ate the Economy" (paulkedrosky.com, July 2025). If you read two, add the Krugman conversation, because the conversation is where he says what the essay can't quite say. If you read three, the Plain English with Derek Thompson interview rounds it out by giving the popular-register version that has carried his thesis further than anything in print.

§Note on sources

This analysis was written without a scrapable Twitter corpus. Paul Kedrosky's visible X timeline shows sixteen tweets and a bio acknowledging that he is "not always very active here." All claims here are sourced from his "Rough Notes" essays at paulkedrosky.com (July 2025 through May 2026), one long-form Substack conversation with Paul Krugman (December 2025), the Plain English with Derek Thompson episode "Yes, AI Is a Bubble. There Is No Question." (March 2026), and the Odd Lots episode "Why Paul Kedrosky Says AI Is Like Every Bubble All Rolled Into One" (November 2025). Where claims are unsupported by direct quote in the essay corpus, they are sourced from the interviews and labeled as such.

The dated, falsifiable prediction in his May 2026 "Mega-IPO Flow" essay — that H2 2026 will see the most unusual equity supply in US history due to forced rebalancing around SpaceX, OpenAI, and Anthropic IPOs — will become gradable within months. The other major dated calls (Minsky-moment-triggered datacenter credit crisis, end-of-scaling reclassification of IG debt) are more open-ended and harder to falsify on a clean timeline.