Alpha Exchange · · 51 min
Paul Kedrosky, Managing Partner, SK Ventures
With Paul Kedrosky
ChorusBring the market's conversation to your AI. Learn more
Summary
Paul Kedrosky of SK Ventures argues that today's AI boom combines the ingredients of past bubbles across technology, real estate, credit and government policy, making it a rational but fragile bubble. He frames Nvidia as a quasi-central bank of AI with GPUs as currency, pointing to its GPU price backstop vehicle, vendor financing, engineered scarcity behind its margins, and export-control leakage as sources of risk. He also warns that token deflation, bounded demand, debt-funded neoclouds and Oracle's leveraged pivot, and a narrow early-adopter base in software engineering create overbuild and credit tightening risks, with Texas data center policy and Treasury market spillovers adding pressure. He concludes that the eventual cost of overbuild could fall on taxpayers, and that markets are watching the wrong signals such as CDS spreads and token rental prices.
Key takeaways
- Paul Kedrosky (bearish on Oracle): Oracle's debt-funded AI pivot leaves it near a ratings downgrade, with CDS and spreads widening.
- Paul Kedrosky (bearish on NVIDIA): Nvidia acts as a central bank for AI, creating shadow-banking-like credit that could fuel fragility.
- Paul Kedrosky (bearish on NVIDIA): Nvidia's revenue-growth pressure led it to stretch export rules, letting GPUs reach China via transshipment.
- Paul Kedrosky (bearish on NVIDIA): Nvidia's announced GPU price backstop was an unbacked currency peg that would break, and it never materialized.
- Paul Kedrosky (uncertain on NVIDIA): Nvidia's margins rest on engineered scarcity, a pattern it has repeated across gaming, crypto and now AI.
What was said, by company and theme
23 views from 1 speaker. Every quote is checked against the episode's transcript (how).
NVIDIA
Across the shows: 177 speakers in other episodes, 183 bullish and 175 bearish calls, past 30 days
Where everyone stands on NVIDIA
Paul Kedrosky · SK VenturesBearish
Nvidia acts as a central bank for AI, creating shadow-banking-like credit that could fuel fragility.
Kedrosky calls Nvidia the central bank of AI. He says it runs a quasi-fractional reserve system by expanding its balance sheet in many ways. These include unsecured lending and repayment schedules that stretch years into the future and require no cash exchange. He argues this creates a currency-like claim that becomes its own shadow banking system, analogous to what built up before the GFC. This is why he sees Nvidia as a dispenser of fragility.
“I often call NVIDIA the central bank of AI, that they're running a quasi fractional reserve system where they're vastly expanding the scope and consequence of their balance sheet in a host of different ways, including unsecured loan, unsecured lending”at 34:07
Paul Kedrosky · SK VenturesBearish
Nvidia's revenue-growth pressure led it to stretch export rules, letting GPUs reach China via transshipment.
Kedrosky argues Nvidia was under heavy pressure to keep growing revenue. That pressure led it to do things Jensen Huang believed were in the company's and the nation's interest. Those things included stretching export policy and tolerating transshipment of GPUs that served China, through cloud services in Southeast Asia or black-market chips that ended up in Asia. He presents this as a consequence of the fragility built up in the system, not a one-off. The implication is regulatory and geopolitical risk for Nvidia.
“the pressure that Nvidia was under to continue to grow revenues, which led it to do things ... really stretching the rules with respect to export policy and transshipment of GPUs that were servicing China”at 34:37
Paul Kedrosky · SK VenturesBearish
Nvidia's announced GPU price backstop was an unbacked currency peg that would break, and it never materialized.
Kedrosky frames Nvidia as a kind of central bank of AI, with GPUs acting as the currency. He cites Jensen Huang's backstop of GPU prices through a roughly $500 billion vehicle with Blackstone, Goldman or KKR. He argues it was an attempt to peg the currency, which the market quickly saw was not sufficiently backed. He says a run and a break of the peg were inevitable. He adds that the vehicle appears to have never materialized, so it may not be in prices at all. He uses it as evidence that GPUs function as a commodity and that the system has points of stress.
“this was an attempt at a currency peg. ... the market realized fairly quickly wasn't sufficiently [backed]. There was inevitably gonna be a run on that.”at 40:26
Paul Kedrosky · SK VenturesUncertain
Nvidia's margins rest on engineered scarcity, a pattern it has repeated across gaming, crypto and now AI.
Kedrosky traces Nvidia's history from graphics boards to gaming consoles, which pushed it into premium gaming. He says it protected margins by becoming a premium provider with engineered scarcity. He notes gross margins rose in that period, then rose again in the crypto boom, which he calls the same phenomenon. He argues the AI GPU cost signal is therefore shaped by deliberately engineered scarcity, not just by disparate market forces. He does not say the margins will fall, but the repeated reliance on scarcity implies the margin profile is contingent on supply constraints that he sees as artificial.
“And to his credit, he realized that the best way to protect his margins was to position yourself as a premium provider with engineered scarcity. ... But consciously engineered scarcity.”at 44:44
Oracle
Across the shows: 32 speakers in other episodes, 10 bullish and 30 bearish calls, past 30 days
Where everyone stands on Oracle
Paul Kedrosky · SK VenturesBearish
Oracle's debt-funded AI pivot leaves it near a ratings downgrade, with CDS and spreads widening.
Oracle is trying to move away from its mature, possibly declining database business into AI infrastructure. Doing so is heavily debt-assisted, and the company is using large amounts of debt. Kedrosky says it is teetering on another ratings downgrade, and its credit default swaps and bond spreads trade accordingly. He calls Oracle a destination of fragility: a company that has to borrow heavily to escape its exposed position in the AI chain.
“it's heavily debt assisted. They're using prodigious amounts of debt to the point that they're teetering on the edge of yet another ratings downgrade and their credit default swaps trade accordingly as do the spreads on much of their debt.”at 32:44
AI infrastructure & data centers
Where everyone stands on AI infrastructure & data centers
Paul Kedrosky · SK Ventures
US executive orders and permit changes accelerate AI buildout, including gas turbines that avoid the grid.
Kedrosky argues the US government sees AI as an existential battle with China and is actively encouraging the buildout, as opposed to only failing to supervise it. He cites a bevy of executive orders over the last two and a half years aimed at accelerating AI infrastructure. One example is EPA changes that permit broader sulfur dioxide emissions for behind-the-meter natural gas turbines, which lets data centers avoid grid interconnection. He says these measures are cumulative and create a tailwind that drives the boom.
“EPA changes that permit even broader sulfur dioxide emissions for behind the meter power, natural gas turbines so that you can avoid grid interconnection.”at 10:16
Paul Kedrosky · SK Ventures
AI technology works, but the capex overbuild will be calamitous and leave assets stranded.
Kedrosky argues that the historical pattern holds: capex booms in railroads, electricity and fiber delivered important technologies but also overbuilt. Calling the technology important does not refute the investment risk. Carlotta Perez's view is that overbuild eventually yields valuable assets. Kedrosky says this time the overbuild is being justified ex post by the claim that it always works out, which accelerates the speed of the boom and the eventual failures. Many assets may be stranded or useless, and the question is who pays.
“the technology works, but the damage with respect to the investment is calamitous and wide, and it's just a question of who pays the price.”at 18:22
Paul Kedrosky · SK Ventures
Industry assumes AI demand elasticity is infinite, so it is overbuilding compute on bounded demand.
Token volume must double roughly every 95 days just to stand still, given the price collapse. Kedrosky says the industry naively assumes people's marginal propensity to consume tokens is infinite, so that falling prices will always find more uses. He argues that demand is bounded, as with cars at zero price. He says the industry extrapolates from the last few years of exponential growth in both demand and price collapse. That leads to overbuilding and fragility.
“And underneath it all is this kind of naive assumption that people's marginal propensity to consume, the elasticity is infinite.”at 22:42
Paul Kedrosky · SK Ventures
Debt-dependent neoclouds, formerly crypto miners, face fragility from borrowing to build GPU capacity.
Many neoclouds were crypto miners. As mining stopped being profitable, they pivoted to using their large GPU farms for AI training and inference. That pivot required vast amounts of debt to build out capacity. Kedrosky calls them recipients of fragility: they are trying to solve a business problem, but they have to borrow heavily to do it.
“And so this looked like a great place to be, but it required vast amounts of debt to build it out such that you could be useful for doing it. So again, a source of fragility”at 33:37
Paul Kedrosky · SK Ventures
Texas's go-slow stance on data centers after the midterms is a real tightening of AI buildout.
Kedrosky says the midterm results mattered a lot, and the initial market reaction was too positive. He treats Texas as a de facto sovereign because it is the third-largest AI data center builder globally, after the US excluding Texas and China. With a governor-elect signalling a go-slow policy on data centers, he calls it a tightening moment. He concedes some construction may get locked in over the next week or two. He also rejects the bullish argument that fewer data centers protect margins, because hundreds of data centers are under construction in China and inference can be served from anywhere.
“Very significant... This was a tightening moment. There was no question about it.”at 37:09
Paul Kedrosky · SK Ventures
Nvidia's new leadership has begun tightening credit to neoclouds, which drive much of token demand.
On Nvidia's earnings call, executives said they were huge supporters of neoclouds, then said they would be careful to choose the right recipients of lending. Kedrosky dismisses the first part and says the tightening cycle has begun for credit extended to neoclouds. He notes neoclouds account for something like 60% of token sales, which turn into compute sales tied to neoclouds. He concludes that neoclouds have entered a tightening cycle, which threatens a key part of the AI demand chain.
“The reality is that the tightening cycle has begun with respect to the extension of credit to NeoClouds who represent ... something on the order of 60% of token sales”at 39:30
AI bubble & tech valuations
Where everyone stands on AI bubble & tech valuations
Paul Kedrosky · SK Ventures
AI boom is a rational bubble because it combines technology, real estate, credit and government forces at once.
Kedrosky uses a four-pillar framework (technology, real estate, credit, government policy) that historically underlies major booms such as the South Sea Bubble, railroads, rural electrification, fiber and the GFC. He argues today's AI boom has all four at once, with each participant seeing its own activity as rational. The fragility comes from the intersection of these silos, which makes the boom a rational bubble in financial-economic terms. Real estate: many treat data centres like commercial real estate, as yield instruments full of GPUs. Credit: a loose credit story. Government: the US sees an existential race with China and is actively accelerating the buildout.
“one of the things that makes it really unusual and almost a rational bubble in financial in a financial economic sense is that it combines each of those components, each of those pillars,”at 0:00
Paul Kedrosky · SK Ventures
Private valuations of SpaceX, Anthropic and OpenAI rival the dot-com Nasdaq peak in nominal terms.
Kedrosky compares the private-market valuations of SpaceX, Anthropic and OpenAI, in nominal terms, with the peak valuation of the Nasdaq during the dot-com bubble. His point is that the froth has moved out of public equities, where the number of listed stocks has declined over 25 years, and into private markets. That is why the usual search for dot-com-style frothy public stocks comes up empty. The comparison is his main evidence that the current boom is large by historical standards.
“in nominal terms, the valuations in the private markets of those three firms are roughly equivalent to the peak valuation of the NASDAQ at the peak of the dot com bubble.”at 13:34
AI pricing & unit economics
Where everyone stands on AI pricing & unit economics
Paul Kedrosky · SK Ventures
AI tokens are a hyperdeflationary commodity, falling about 13-fold per year and still deflating.
Kedrosky describes tokens, the atomic unit of AI, as the first hyperdeflationary industrial commodity. By his measure, token prices have been falling about 13-fold year over year and are structurally set to keep deflating. This price collapse drives his broader argument about exponential volume needs and overbuild risk.
“Think of them as the first hyper deflationary industrial commodity that is structurally a deflationary commodity that has been deflating and will continue to by some measures about 13 fold year over year.”at 21:12
Paul Kedrosky · SK Ventures
GPU rental price indices are nearly useless as predictors because spot prices reflect scarcity, not long-term commitments.
Kedrosky says people treat GPU rental price indices as predictive, but they are nearly useless. The indices average across very different token and GPU providers, which hides large dispersion, a Simpson's paradox problem. Marginal buyers pay spot prices for compute, like a hotel room price during a big conference. Those prices do not reflect what hyperscalers pay under long-term commitments. He argues Jensen Huang's claim that flat or rising rental prices show everything is fine is a profound misunderstanding, and that the price signal partly reflects engineered scarcity.
“But it's not reflective of what's happening in the broader market in terms of what the hyperscalers and others are paying in terms of their long term commitments.”at 43:10
Credit markets
Where everyone stands on Credit markets
Paul Kedrosky · SK Ventures
After AI capex, hyperscalers and levered firms face a prolonged deleveraging insensitive to rate cuts.
Kedrosky argues that AI-era firms are levering up dramatically because their pristine pre-AI balance sheets let them borrow. He expects the aftermath to be a prolonged deleveraging, similar to Richard Koo's balance sheet recession in Japan. In that regime, firms want to pay down debt rather than borrow, so further interest-rate declines do little to revive borrowing. He draws on the Japan episode, which has parallels to both the dot-com and GFC episodes.
“levering up dramatically in part because they can because they have such pristine balance sheets. The hyperscalers pre AI CapEx had such pristine balance sheets that what we will see on the other side of this is a prolonged deleveraging”at 14:36
AI: overall outlook
Where everyone stands on AI: overall outlook
Paul Kedrosky · SK Ventures
AI is probably among the most important IT technologies of the last hundred years.
Kedrosky rejects the argument that AI does not work or is all hyperbole. He calls it probably among the most important technologies of the last hundred years from an IT standpoint, while granting that air conditioning and some medical advances matter more. He uses this to argue that a strong technology story and a financial bubble can coexist.
“It's probably among the most important technologies of the last hundred years from an IT standpoint.”at 17:07
US Treasuries & yields
Where everyone stands on US Treasuries & yields
Paul Kedrosky · SK Ventures
AI fundraising is at the margin putting pressure on Treasury markets as capex leaks into them.
Kedrosky describes capex-boom leakage into other markets. In late summer, AI fundraising was increasingly causing marginal pressure in Treasury markets. He calls this remarkable and treats it as a sign of how circular and gigantic the capex phenomenon has become. It parallels earlier booms, where the spending spilled into adjacent markets.
“fundraising around AI was increasingly causing at the margin pressure in treasury markets, which is remarkable this idea, right?”at 18:01
Cybersecurity
Where everyone stands on Cybersecurity
Paul Kedrosky · SK Ventures
AI-driven vulnerability discovery is outpacing defenders, making critical systems increasingly fragile.
The attack surface of information-technology systems is expanding much faster than defenders can protect it, and AI systems are growing more capable at attacking. About 8% of all critical software vulnerabilities ever found since the 1960s were found in the first three weeks of October of that year. Roughly half of all critical vulnerabilities in installed software were found since January. Kedrosky says defenders must be lucky every time while attackers need to be lucky only once. Any of airlines, banks or natural gas supply could be hit.
“about 8% of all critical software vulnerabilities ever found in the history of the industry back to the nineteen sixties were found in the first three weeks of October of this year.”at 24:06
Venture capital & startups
Where everyone stands on Venture capital & startups
Paul Kedrosky · SK Ventures
AI spending is concentrated in unprofitable venture-backed firms and a few software engineers.
Ramp's data on organizational AI spending per employee shows venture-backed companies are by far the largest spenders. That spending is concentrated in young, unprofitable companies funded by venture capital. Kedrosky points to concentration in two ways: dependence on fragile young companies, and dependence on a small group of technology firms and their software engineers. That concentration is risky because of the early-adopter problem, since these buyers are unrepresentative of the long-term market.
“the largest spenders on a per employee basis by far are venture backed companies.”at 26:47
Enterprise AI adoption & AI software
Where everyone stands on Enterprise AI adoption & AI software
Paul Kedrosky · SK Ventures
AI demand projections overestimate the TAM by extrapolating from software engineers to white-collar work.
The early adopters of large language models were software engineers. Software suits these models because it is rules-based and gives a clear ground-truth signal, so the models can improve through gradient descent and run semi-autonomously. Most other white-collar work lacks that signal. A memo or PowerPoint gets worse with iteration, and whether a boss likes it is unknown until the boss says so. Extrapolating from these early users therefore overstates the total addressable market and the ability to let models run autonomously in other white-collar work.
“There is a huge overestimation of the TAM if for many reasons, but that is I think one of the most important ones is this naive extrapolation”at 30:45
Frontier AI labs & model race
Where everyone stands on Frontier AI labs & model race
Paul Kedrosky · SK Ventures
Frontier AI labs are stuck between cloud providers, unable to be true clouds or stand alone.
Kedrosky says he does not hold much hope for the frontier AI companies over time. He argues they are not the source or recipient of the fragility in the system but sit in the middle. Their position is awkward: they are neither true cloud service providers nor able to stand on their own without cloud service providers. That dependence, in his view, will make their position increasingly difficult over time.
“I don't actually hold a lot of hope for the frontier companies over time. I think increasingly, they're in this awkward position of neither being true cloud service providers nor able to stand on their own without cloud service providers.”at 35:09
AI & jobs
Where everyone stands on AI & jobs
Paul Kedrosky · SK Ventures
AI capex built on capturing white-collar labor revenue would reprice the labor market, a systemic credit risk.
Kedrosky takes AI providers at their word that their target is a material share of the global white-collar labor TAM. He says they need roughly $5 trillion in revenue by 2031-2032 to justify current AI capex with a hurdle margin. In commodity markets, repricing a slice of a market reprices the whole market. So capturing a slice of white-collar labor would reprice much of it, shrinking the base. In aging societies with high dependency ratios, that could cut the system's ability to fund itself. The tail risk is that taxpayers, rather than banks or AI labs, end up bearing the cost. He notes markets are focused on credit default swaps and token rental prices instead.
“And so they're saying to themselves, I need to get to $5,000,000,000,000 in twenty thirty one, twenty thirty two revenues to justify with appropriate hurdle in margin, the current AI CapEx spending.”at 47:02
Across the shows
What other episodes Chorus covered said about the same companies in the 30 days up to this one.
Topics
- AI adoption beyond software
- AI bubble framing
- AI compute demand elasticity
- AI financing pressure on Treasuries
- AI spending concentration
- AI technology importance
- AI-driven vulnerability discovery
- GPU price backstop
- GPU rental price indices
- corporate deleveraging after AI capex
- data center buildout
- debt-funded AI transition
- export control compliance
- frontier lab business model
- government support for AI buildout
- gross margins and pricing power
- neocloud credit
- neocloud debt-funded GPU buildout
- overbuild risk
- private AI company valuations
- token prices
- vendor financing and balance sheet
- white-collar labor repricing
More from Alpha Exchange
- Meredith Whitney, Founder and CEO, Meredith Whitney Advisors · November 9, 2026
- Lori Calvasina, Managing Director, Head of US Equity Strategy, RBC Capital Markets · November 2, 2026
- Randy Cohen, Professor of Entrepreneurship & Finance Harvard Business School Co-Founder, PEO Partners · October 28, 2026
Ask your AI about this episode
Chorus gives Claude, ChatGPT and other assistants every view from this episode and every other show Chorus follows. Ask who agrees, who changed their mind, and what else was said about the same companies.
Chorus reports what speakers said. It isn't investment advice. How Chorus works · About Chorus