Odd Lots · · 1h 21m
This Is The Problem With AI's Most Bullish Chart
With Paul Kedrosky, Stanley Lee
ChorusBring the market's conversation to your AI. Learn more
Summary
Stanley Lee and Paul Kedrosky examine whether the rising GPU rental price index is genuine evidence of AI demand, and conclude it is substantially shaped by financing conditions. Lee argues that spot compute prices are propped up by venture equity, that VC deployment fell about 25% from Q1 to Q3, and that neocloud lenders have cut loan-to-value ratios from roughly 80% to 50-60%, limited lending to a handful of inference companies, and are moving toward hedges on B300 residual values. Kedrosky takes a broader skeptical view, treating the buildout like project finance that would need $3-6 trillion in annual revenue by 2032, questioning consumer willingness to pay, and describing Nvidia's vendor financing as a fragile, highly levered loop that could unwind geometrically. The discussion also covers Anthropic's enterprise revenue prospects and cost of capital, with both speakers seeing concerns about its vertical expansion and training spend, and touches on AI-driven cybersecurity demand as a relative beneficiary.
Key takeaways
- Stanley Lee (bullish on CoreWeave): CoreWeave's move to higher prepayments and investment-grade offtakers is a credit-positive step public markets rightly rewarded.
- Stanley Lee (uncertain on NVIDIA): Nvidia has been central to financing the AI buildout, so a more conservative Nvidia would stress the ecosystem.
- Paul Kedrosky (uncertain on Anthropic): Anthropic's revenue case needs deep enterprise penetration and mass substitution; only coding shows that so far.
- Paul Kedrosky (bearish on NVIDIA): Nvidia's balance-sheet lending to AI customers is a highly levered, circular system that props up demand.
- Paul Kedrosky (bearish on NVIDIA): Nvidia's circular financing of neoclouds is a fractional banking system that will unwind geometrically
What was said, by company and theme
24 views from 2 speakers. Every quote is checked against the episode's transcript (how).
Anthropic
Across the shows: 133 speakers in other episodes, 229 bullish and 110 bearish calls, past 30 days
Where everyone stands on Anthropic
Stanley Lee · Liquid ComputeUncertain
Anthropic is rated high double-B, so its compute financing costs depend on middle-market sentiment.
Lee argues that even though Anthropic is a large buyer that could absorb much of the compute, it is rated as a high double-B company. That will affect its cost of capital, so the price it pays for compute depends on sentiment in the middle market. He also points to Anthropic as an example of a company shifting from equity-funded compute purchases toward revenue-funded ones, which he says is what the sector needs. Lee is cautious that large-lab demand will not offset the negative sentiment in the middle market.
“Anthropic is being rated as, you know, a high double b company. This will affect their cost of capital.”at 30:14
Paul Kedrosky · SK VenturesUncertain
Anthropic's revenue case needs deep enterprise penetration and mass substitution; only coding shows that so far.
Kedrosky says the bull case for Anthropic is that it is the enterprise player, so the revenue required can only come from deep enterprise penetration. Revenue can come from substitution of existing work or from complementing work and raising productivity so the pie grows. So far the only clear substitution and productivity gains are in software engineering. He adds that early adopters, such as obsessed coders paying any price, look nothing like mainstream users, so projecting from early adopters is a mistake.
“The only way to get there is to assume deep penetration of the enterprise with Anthropic and mass substitution”at 47:48
Paul Kedrosky · SK VenturesUncertain
Anthropic's pivot to inference and vertical apps helps unit economics but raises its capex obligations.
Kedrosky has long argued that the sooner Anthropic stops aggressively training frontier models and focuses on inference, the faster its unit economics improve. Over the last 18-24 months it has shipped more vertical products (legal, finance, health care) than model updates, and benchmark progress has flattened. But selling directly into those markets competes with its distributors and investors, including Microsoft marketplace sales that account for almost 48%. The S-1 suggests Anthropic will move away from cloud rentals and build more data centers itself, with future obligations he estimates at about $800 billion versus lease-based commitments before.
“So this is my argument all the way along on Anthropic has been the sooner they stop aggressively training models to be at the frontier and focus largely on inference, the faster the unit economics get better.”at 57:18
Paul Kedrosky · SK VenturesBearish
Anthropic's push into verticals is making customers reluctant to share data, weakening its model
Kedrosky argues that Anthropic is only as good as the data it gets when it enters verticals. By publicly signalling that it wants to be in customers' verticals, Anthropic has made firms wary. He says he hears new reticence from people to share data with Anthropic or partner with it in many ways, because they fear it will climb up the value chain and absorb their businesses, like a shark eating everything as it climbs. He compares this to Microsoft's history of building competing products on its platform, and says the effect is much larger for Anthropic.
“anthropic is only as good as the data it gets as it enters these verticals... now they've told people that we want to be in your vertical. And so as I talk to people more and more, there is this new reticence”at 1:14:28
NVIDIA
Across the shows: 172 speakers in other episodes, 184 bullish and 161 bearish calls, past 30 days
Where everyone stands on NVIDIA
Stanley Lee · Liquid ComputeUncertain
Nvidia has been central to financing the AI buildout, so a more conservative Nvidia would stress the ecosystem.
Lee says Nvidia has been very relevant in financing both sides of the equation. It has backstopped and invested in many neolab equity raises, and provided backstops to creditors. Because GPU residual values have been so high, these backstops have not been hit yet, but Lee sees that as a possibility. If financial conditions tighten and Nvidia simultaneously gets more conservative, that would be a major stressor on the ecosystem. He calls the new management's posture a trillion-dollar question. He argues Nvidia should now stabilize the middle market through incentives to work with distressed neocloud lenders, not by putting its own balance sheet down. Lee's concern is a total free-for-all with no control over remarketed capacity.
“NVIDIA has been very relevant in helping to finance both sides of the equation. NVIDIA has backstopped and invested in many of the Neolab equity raises.”at 27:48
Paul Kedrosky · SK VenturesBearish
Nvidia's balance-sheet lending to AI customers is a highly levered, circular system that props up demand.
Kedrosky describes Nvidia as the 'central bank' of AI because it uses its own balance sheet to extend credit, creating money on the other side as the Fed does. Customers, such as neoclouds and their banks, often get credit with no payment terms and then spend the money buying back Nvidia chips, often on delayed payment terms. That recycling attracts venture capital and more neocloud construction, and the loop is highly levered, so it works like a fractional banking system rather than a one-for-one recycling of cash.
“So it goes round and round and round, not just because it's one for one, because it's highly levered. It's a fractional banking system.”at 1:11:12
Paul Kedrosky · SK VenturesBearish
Nvidia's circular financing of neoclouds is a fractional banking system that will unwind geometrically
Circular financing, where Nvidia's deposits and investments fund neoclouds that then buy Nvidia GPUs, is inherently a highly levered fractional banking system, like the shadow banking system and private credit before it. Kedrosky says that in 2026 Nvidia was effectively that system's sole creator of money through circular deposits that generated incremental spend. Fractional banking systems do not unwind linearly when they reverse; they unwind geometrically because everything pulls back to the core. He points to the Nvidia call where the acting CEO said Nvidia would be more discerning about neoclouds, which he reads as the start of that reversal.
“the fractional reserve banking system of 2026 was Nvidia. It was no longer anybody else... when they begin to go in reverse, they don't go in reverse linearly either. They go in reverse geometrically”at 1:11:25
CoreWeave
Across the shows: 75 speakers in other episodes, 22 bullish and 78 bearish calls, past 30 days
Where everyone stands on CoreWeave
Stanley Lee · Liquid ComputeBullish
CoreWeave's move to higher prepayments and investment-grade offtakers is a credit-positive step public markets rightly rewarded.
Lee says CoreWeave did two things last week. It required higher prepayments from customers to push upfront cost onto them, and it went back upmarket. Its delayed draw term loan is now marketed to investment-grade offtakers, whereas the previous one combined Anthropic, HRT and Jane Street, which he calls more high double-B operators. Lee says the market rewarded this, which it should. He presents it as a way to reduce financing risk, though he notes that the trickle-down of this approach to middle-market lenders is what squeezes smaller neoclouds.
“I mean, just look at what CoreWeave did last week, which public markets rewarded them for, which they should.”at 9:09
CrowdStrike Holdings
Across the shows: 5 speakers in other episodes, 8 bullish and 1 bearish calls, past 30 days
Where everyone stands on CrowdStrike Holdings
Paul Kedrosky · SK VenturesBullish
AI-driven surge in critical vulnerabilities is a bonanza for CrowdStrike.
Kedrosky says about 8% of all critical software vulnerabilities ever found were discovered in the first three weeks of October, and about 50% of all critical bugs ever found were found in the first half of this year, with the pace accelerating. Since AI tools let people find vulnerabilities in other systems quickly, he argues cybersecurity vendors benefit, singling out CrowdStrike and Palo Alto. He adds that this is not good for productivity elsewhere.
“this is a fantastic bonanza for cybersecurity guys, Palo Alto, for CrowdStrike, and lots of others, but it's not particularly good for anyone else in terms of producing productivity.”at 51:05
Palo Alto Networks
Across the shows: 8 speakers in other episodes, 7 bullish and 1 bearish calls, past 30 days
Where everyone stands on Palo Alto Networks
Paul Kedrosky · SK VenturesBullish
AI-driven surge in critical vulnerabilities is a bonanza for Palo Alto Networks.
Kedrosky says the surge in critical software vulnerabilities, with about 8% found in the first three weeks of October and 50% of all critical bugs ever found in the first half of this year, is accelerating. This makes AI-era cyber threats a major business for cybersecurity vendors, and he names Palo Alto alongside CrowdStrike as a beneficiary. Productivity gains from the same tools mostly accrue to attackers and defenders rather than to other sectors.
“this is a fantastic bonanza for cybersecurity guys, Palo Alto, for CrowdStrike, and lots of others”at 51:05
AI infrastructure & data centers
Where everyone stands on AI infrastructure & data centers
Stanley Lee · Liquid Compute
Mid-market neocloud operators face a financing squeeze, so consolidation and failures are coming, and some will fail.
Lenders now underwrite the operator, not just the offtaker, because take-or-pay contracts are only as good as the neocloud's uptime and SLA performance. Private credit lenders now see the same few offtakers in 5 to 10 loans a week and want experienced operators, not first or second deployments. Lee says the gap is bridged only if neoclouds raise holdco equity or get bigger customer prepayments, and the smallest neoclouds, often converted crypto miners, will struggle to fund the roughly 30% equity cushion. Lee expects a shakeout and consolidation of middle-market neoclouds. Some will fail, which he calls healthy, but he is concerned about a free-for-all if Nvidia pulls back entirely. Access to debt in the middle market will equal access to equity.
“lenders are going to go back to these NeoClouds and say, Okay, you know what? I will do your loan. I'll do it in the mid high teens, but I want a 50% LTV structure.”at 13:15
Stanley Lee · Liquid Compute
Rising GPU rental rates mostly reflect buyers' cost of capital, not underlying demand, so the chart is less bullish than it looks.
Lee says the rising Silicon Data B200 rental index reflects demand, but also that 'a huge part' of it is the cost of capital. Spot and short-term GPU prices are high because buyers who want term compute at fixed prices cannot get financed, since their neocloud sellers cannot take the offtake contract to a lender. VC deployment fell about 25% from Q1 to Q3, so the demand is funded by venture equity rather than revenue. Lee argues the market is thin, like oil, with much of the compute traded in spot, and that a few bids could move the reference price. Residual values depend on about 30 variables, and the demand side depends on buyers' ability to pay, which currently comes from venture equity. Hyperscaler pricing is typically excluded from these charts. Lee also says demand from large labs like Anthropic will not dwarf negative sentiment in the middle market, since those labs' own financing costs depend on that sentiment.
“A lot of it is the cost of capital. I think a huge part of this is the cost of capital.”at 16:55
Stanley Lee · Liquid Compute
Private credit wants GPU residual-value hedges, likely puts on the B300 index, because compute recovery is reletting capacity.
Lee says credit is negatively convex, so lenders need compensation for writing puts. Value accrual from spot prices going higher goes to equity holders, which makes spot-heavy players hard to finance. A hedge that floors recovery is compelling and is effectively a synthetic NVIDIA backstop, though it carries basis risk, as with early CDS. Lee says he gets a lot of demand from private credit for hedges, likely a strip of put options on the B300 index. If recovery comes from reletting capacity rather than selling it, that is the recovery to hedge, while insurance on scrap value is less useful. He reports two to three inquiries a day from large private credit firms about indicative pricing on B300 puts over the last two weeks, after months when nobody focused on these hedges. Lee compares this to aircraft finance, where a liquid relet market and regular appraisals made the asset class financeable after 2008.
“I do get a lot of demand from private credit for these hedges. And what it looks like will be a strip of put options on our index, usually the b 300 index.”at 26:15
Paul Kedrosky · SK Ventures
Rising GPU rental price indices are a misleading proxy and do not prove the viability of AI capex.
Kedrosky says people confuse a rising 'number goes up' chart with the viability of the business and its ability to repay the financing behind it. The underlying economics are hard to model, so analysts grab proxy measures. GPU rental price indices are not falling and are rising, but he argues there are several reasons this could be happening besides real economic demand. He calls it an availability bias problem, like searching for keys under the lamppost. He agrees the technology can be groundbreaking and still not economic at current valuations.
“number goes up into the right with the viability of the business and your ability to pay off the underlying financing.”at 42:03
Paul Kedrosky · SK Ventures
AI capex only pays off if it generates roughly $3-6 trillion in annual revenue by 2032.
Kedrosky treats the buildout like project finance. Take projected capex through about 2030-2032, assume a hurdle rate, gross margins and target margins, then work backwards to the revenue needed to justify it. Depending on the assumptions, that comes to roughly $3-6 trillion in annual revenue by 2032, which he says is very large and not obviously attainable. He notes that the TAM framing, such as this being around 18% of global white-collar work or bigger than the global software industry, can justify almost anything, so the arithmetic is the real test.
“something like three to six trillion in terms of annual revenues by, say, the 2032 time frame.”at 44:22
Paul Kedrosky · SK Ventures
A growing share of pre-2026 data center loans are impaired, and new KYC terms will make neocloud financing costlier.
Kedrosky points to a rising share of the pre-2026 data center loan book trading at mildly impaired levels, just as Congress advances know-your-customer rules. Those rules are now showing up in loan terms outside the US, including loans in Southeast Asia. He argues lenders can't fully know their customers in a cloud world, so they will avoid these loans. He expects loans to become harder to get and more expensive, with wider spreads, which he calls obviously bad for neocloud providers.
“an increasing fraction of the pre 2026 loan book out there is trading at at least mildly impaired levels.”at 1:01:40
Paul Kedrosky · SK Ventures
Federal policy has acted as a de facto stimulus for data centers, and post-midterm state slowdowns could reverse it.
Kedrosky argues the government has effectively run a private-sector stimulus through AI capex. Over a dozen executive orders have made it easier to build data centers, connect to the grid, and work around EPA sulfur dioxide rules. He thinks the White House knows this and is afraid of ending it, because AI capex supports markets and 401(k)s. He says the midterms change this: Texas would be the third-largest data center builder if it were a country, and a go-slow governor-elect could materially slow construction.
“So I've argued that essentially we're running a private sector stimulus program aided and embedded by the federal government.”at 1:05:55
Venture capital & startups
Where everyone stands on Venture capital & startups
Stanley Lee · Liquid Compute
VC equity is funding AI compute demand, but VC deployment is slowing and the cost of venture equity is rising.
Lee says VCs deployed about 25% less in Q3 than in Q1 of this year, which he links to stress in the part of the market that depends on venture funding. He says much of the liquidity is expected to return with large IPOs. The middle-market inference companies have been raising at a blistering pace, so many VCs are over-deployed, and the cost of venture equity is getting higher. Lee says the ability of compute buyers to pay currently comes from venture equity, not revenue. He describes the AI compute boom as highly correlated with VC money, and says what matters is how much AI productivity reaches the broader economy, which he says is 'very little' so far, and whether companies can generate enough revenue to buy compute without equity.
“That being said, VCs deployed less in q three than q one by about 25%.”at 16:55
Paul Kedrosky · SK Ventures
Venture-backed companies drive AI spending, but much of that token spend won't turn into products worth keeping.
Kedrosky cites RAMP data showing that the biggest per-employee spenders are venture-backed companies, with venture money turning into spending on the most expensive frontier model tokens. He cites NBER data showing a sharp rise in the number of apps but a collapse in ratings per app. He argues this means many startup token-maxing efforts are not producing products anyone cares about, and that much of the capital behind them will probably have to be written off even if some value is created. He uses Cursor as a larger example of this dynamic.
“the largest spenders in... On a per capita basis are uniformly venture backed companies.”at 1:03:46
Private credit
Where everyone stands on Private credit
Stanley Lee · Liquid Compute
Private credit is tightening GPU loan terms sharply, cutting LTVs and demanding larger prepayments.
Lee describes typical GPU loans of about $100M, three-year fully amortizing, at roughly 70% LTV priced around 12% over the summer. He says LTVs have already fallen: 80% loans priced over the summer, then about 70% after Labor Day, and now lenders ask for 50 to 60%. Lenders are also requiring higher upfront payments and lower LTV structures, and are limiting lending to a list of five or six inference companies worth about $20 billion instead of smaller neoclouds. Emergency bridge loans to neoclouds carry tight covenants, and if there is no takeout in three to six months, lenders will foreclose and seize the equipment. Lee says the covenants that matter most are access to the hardware and a tri-party agreement with the data center to relet contracts.
“we're already seeing LTVs come down. We were seeing 80% LTV loans get priced over the summer. I think post Labor Day, it was more 70%. And now people are asking for 50 to 60%.”at 32:14
Consumer AI products
Where everyone stands on Consumer AI products
Paul Kedrosky · SK Ventures
Consumers will never pay enough for AI, so the revenue case must rest on enterprise.
Joe Weisenthal relays the argument that consumers would need to pay around $200 a month for something like ChatGPT Pro to reach the required revenue, but consumers are content with free features like Gemini summaries and search. Kedrosky agrees and says that paid consumer adoption at that scale will never happen. He therefore argues that any path to the numbers needs an enterprise story, since he doesn't believe the required revenue can be reached without it.
“Which is never gonna happen. So the whole story has to be the enterprise story. Right? ... I don't even believe that there's any way to get to these kinds of numbers without an enterprise story.”at 47:31
AI & jobs
Where everyone stands on AI & jobs
Paul Kedrosky · SK Ventures
AI productivity gains are real mainly in coding, since white-collar work lacks the ground truth loops need.
Kedrosky says software is unusual because it has ground truth: the code works or it doesn't, so models can run undirected feedback loops and improve, including recursive self-improvement. Most white-collar work has no equivalent test. Letting a model work on a PowerPoint deck for twelve hours won't produce something better, because nobody can define what better means. Looping models on tasks without ground truth yields exploding errors rather than better products, which is the core problem as use moves from coders to ordinary users.
“I defy you to give a PowerPoint presentation to a model, let it loose for twelve hours and say, I bet what it comes back with is better. It won't be.”at 49:51
IPOs, M&A & deal activity
Where everyone stands on IPOs, M&A & deal activity
Paul Kedrosky · SK Ventures
IPO oversubscription figures are meaningless, since they are non-binding indications that bankers leak.
Kedrosky says oversubscription means only that underwriters did their marketing job. Investors submit non-binding indications of how much stock they would like, knowing they will get only a fraction, and bankers then leak huge multiples to create a perception of demand. He cites Chinese IPOs at the summer 2025 bubble peak that were 8,000 times oversubscribed and then fell 50%. Tracy Alloway agrees and says the media repeats these claims uncritically.
“Oversubscription doesn't mean anything because it's a non binding commitment from institutional investors saying, In a perfect world, here's how much stock I would buy.”at 59:31
AI bubble & tech valuations
Where everyone stands on AI bubble & tech valuations
Paul Kedrosky · SK Ventures
AI capex money recycles into US stocks, making that flow fundamental to capital markets.
Responding to a Deutsche Bank note, Kedrosky accepts the argument that compute spending is recycled into US stocks, much as petrodollars once flowed into Treasuries. He points out that global trade in semiconductors is now larger than oil and gas, which he says shows the scale involved. He concludes that this recycling is now fundamental to what is happening in capital markets.
“Well, global trade in semiconductors is actually now larger than oil and gas. ... The point that that recycling is is now fundamental to what's happening in capital markets.”at 1:08:32
Across the shows
What other episodes Chorus covered said about the same companies in the 30 days up to this one.
- NVIDIA172 speakers · 184 bullish, 161 bearish calls
- Anthropic133 speakers · 229 bullish, 110 bearish calls
- CoreWeave75 speakers · 22 bullish, 78 bearish calls
- Palo Alto Networks8 speakers · 7 bullish, 1 bearish calls
- CrowdStrike Holdings5 speakers · 8 bullish, 1 bearish calls
Topics
- AI capex recycling into US equities
- AI productivity beyond coding
- AI startup spending and returns
- AI-driven cyber threat demand
- Anthropic compute financing costs
- CoreWeave financing structure
- GPU rental prices
- GPU rental rates
- GPU residual value hedging
- IPO oversubscription signals
- Nvidia financing backstops
- VC funding for AI compute buyers
- circular financing
- consumer willingness to pay for AI
- data center capex
- enterprise AI revenue
- federal policy and data center buildout
- model training vs inference strategy
- neocloud and data center loan credit
- neocloud financing and consolidation
- neocloud lending terms
- vendor financing of AI buyers
- vertical expansion
More from Odd Lots
- How GTA VI Became the Most Expensive Video Game Ever Made · November 18, 2026
- The Impossible Trilemma That Hangs Over the Nvidia Crisis · November 16, 2026
- How Wall Street Is Bracing For the Post-Jensen Huang Nvidia Era · November 14, 2026
- Why a $2 Billion Law Firm Suddenly Collapsed · November 13, 2026
- What Nvidia's Megaspeed Scandal Is Really All About · November 12, 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