Paul Kedrosky (partner, SK Ventures; research fellow, MIT Initiative on the Digital Economy; ex-sell-side analyst) · AI capex, data-center financing, bubbles and system-scale economics — running synthesis of his appearances, with per-item breakdowns and a stock index.
The downstream beneficiary of look-through lending — financing that never asks what happens inside the data center "leads to much more data center construction and far more GPUs and a lot more SK Hynix high bandwidth memory." Real demand, but created by a credit decision rather than end use.
The yardstick for model commoditization — on composite measures a frontier Anthropic model, Qwen and DeepSeek land in the same place, and in his blind Pepsi-Coke tests behind a harness "inevitably no one can tell the difference."
Edged (private; inference-ASIC startup, per a Wall Street Journal story — spelling as spoken)
Not a pick but a signal: the inference-ASIC startup cleared design verification in 42 days versus a normal six or seven months and its first design worked — evidence that AI-designed silicon ("vibe chipping") is dissolving the chip industry's tribal-knowledge moat, which "will help prick the bubble itself."
Named only inside the host's question about whether one model winner emerges the way Google beat Ask Jeeves; Kedrosky's answer is that the models have converged, so the search-era winner-take-all analogy doesn't apply this cycle.
The worked example of overdetermined failure, not a forward call: down 75% with every post-hoc explanation missing that the P/E went from ~70 to ~20 — at high valuations there are so many low-probability ways to fail that failure becomes predictable, and the stock "got pecked to death by ducks."
Raised as the listed-company marker of the supply regime change — big tech shifting from buybacks to share issuance and dilution, the same direction as the coming mega-IPO wave. No view on the business itself.
Context for the IPO-supply mechanism: the fund was long exactly the liquid, best-performing names managers must sell months ahead to fund mega-IPO allocations, so "anticipation of the upcoming flood of new issues became a pin" under its implosion.
Discussed purely as issuance supply: with Anthropic and a couple of peers, the listings would be "bigger than all the IPOs from the '90s combined — not just that, it's all post-World War II combined," an estimate he has raised from $4T to ~$5.5T.
Bearish on both sides — as issuance supply (part of the $4–5.5T IPO wave funds must sell liquid winners to absorb) and as product, since "very little difference" separates its frontier model from Qwen or DeepSeek, so multi-billion-dollar training runs can no longer be justified.
His stand-in for how the semiconductor "super cycle" ends — record cash into Taiwanese and Chinese fabs implies "a tsunami of supply in early 2028," and in the most capital-intensive boom-bust industry on earth "once you lock in supply, prices are going to zero." In ten years: "remember when Micron was at such and such a price."
The scarcity story doesn't survive the usage data: GPUs run at only 35–40% utilization even at peak load while used A100 prices are bid up — so the marginal buyer is hoarding and double/triple-ordering, and the idle fleet is "a potential flux of product into the market" waiting to unwind.
Frontier-model economics, not technology, are the problem: with token prices deflating 70–80% a year it takes 400% unit growth just to stand still, so 18% quarter-on-quarter "was deemed a disappointment" — Wile E. Coyote over thin air, now with external debt service on top.
In one line: AI is "the first [bubble] that sits at the intersection of all of the forces that created the largest bubbles in US history" — loose credit, a genuine technology story, a real-estate component and a policy angle, all at once — and the datable tell is that data-center financing crossed from internal cash flow to external credit in the first half of 2026, while the product it funds (tokens) deflates 70–80% a year. Kedrosky is not an AI sceptic: the technology is "wildly useful." He is a scale analyst arguing the financial structure built on it has become divorced from what the machines actually do.
The Minsky crossover is the whole argument (2026-AUG-28). The bull rebuttal used to be they're spending their own cash flow, who are you to object. As of H1 2026 that flipped: more than half of data-center financing is external — ABS, private credit, sovereigns, SPVs. Lenders underwrite the 12-year renewable lease and the prime credit behind it, not the asset: "there could be hide-and-go-seek competitions going on inside the data centers and they wouldn't give a…" A JP Morgan figure puts data-center-related paper at 15–18% of the investment-grade market — larger than financial services. Tech's defining virtue (no debt, pristine balance sheet, cash-flow monsters) has been reversed into leveraged assets with perpetual maintenance capex. So: "who gets re-rated first? Do utilities all of a sudden see a huge spike in valuation, or do technology companies… get re-rated to look more like utilities? I obviously think the latter."
Tokens are "the first hyper-deflationary commodity in the history of modern economies" — and the math is brutal. Prices fall 70–80% a year, so a frontier model company needs 400% unit growth just to stand still, before pleasing Wall Street and before servicing the new external debt. Read the headline numbers through that bar: OpenAI's 18% quarter-on-quarter "was deemed a disappointment." Great for consumers, "toxic to the frontier companies" — and every adjacent industry tokens "brush up against" gets a deflation wave of its own.
The original sin: the whole build-out was sized off the wrong customers. LLMs first worked brilliantly on software because code has a strict grammar, a tight gradient descent and an expansive output shape (one prompt in, a million lines out). Most white-collar work is the opposite — compressive (40 pages in, five bullets out) with no error feedback. "The first domain where AI was applied could hardly be less representative of AI's future if you tried, and yet that's the domain from which we're extrapolating our futures." Capacity plans, hurdle rates and syndicated debt were all sized off that curve.
Failure is overdetermined at high valuations. Stop looking for the catalyst: twenty independent ways to break, each ≤5% likely, compounds to a >60% chance of failure — "what looks unpredictable is actually highly predictable." Nike is the worked case (P/E ~70 → ~20; every explanation blames shoes or politics), and the mechanism is being "pecked to death by ducks." His candidate ducks for AI: sovereign funding stress, the credit market, public backlash, and a chip supply tsunami in early 2028 from record Taiwanese/Chinese fab spending — "once you lock in supply, prices are going to zero."
The models have stopped separating; the harnesses hide it. Composite year-over-year gains have flatlined (10–12% → 1–2%), variance across vendors has collapsed — "very little difference between a frontier model from Anthropic and a frontier model from Qwen or from DeepSeek" — and his blind Pepsi-Coke tests confirm nobody can tell. What creates the illusion of progress is the harness (Claude Code, the codexes) wrapping the model. Conclusion: "the game is almost over in terms of pretending that you can justify multi-billion dollar training runs," and "the most successful frontier AI company will be the first one to stop pretending they can train new AI models." Meanwhile AI-designed silicon ("vibe chipping" — Edged's 42-day design verification) is dissolving the chip incumbents' tribal-knowledge moat from the inside.
$4–5.5 trillion of IPO supply is a market-wide event, not a single-name one. SpaceX, Anthropic and a few peers listing would exceed all post-WWII US IPOs combined. Funds hold no spare cash, so allocations are funded by selling — and what gets sold is predictable: the most liquid, the most overlapping, and perversely the best-performing holdings, starting weeks-to-months ahead so as not to look "bigfooted." He modelled it in March/April and thinks the resulting pressure was the pin under the Situational Awareness implosion; the same regime shift shows up in listed names (Oracle moving from buybacks to issuance).
The macro stakes: AI is more than half of US GDP growth, six quarters running. For "probably the sixth time in Western history" a non-governmental force is large enough to move the tide of GDP growth — which makes policy credit-taking a causality error ("the dog barks, the mailman leaves, the dog takes credit"). It also means the economy's growth path is hostage to one capex cycle. And the public has turned: 75% don't want a data center in their county (worse than nuclear), driven by lost agency and, uniquely in the US, the fact that a threat to employment is a threat to health care and therefore to personal solvency.
Appearances
One dated page per item — each has its stock table, talking points, and the saved transcript. Newest first.