Newsletter · · Ashutosh Agarwal

A Bear Pulls the AI Debt Reckoning Forward as Goldman Counters With Data - The AI Capex Tracker - Week of September 24, 2026

The AI Capex Tracker for the week of September 24, 2026, synthesizing investor podcasts published September 23 and 24. SK Ventures’ Paul Kedrosky pulled his AI-debt reckoning forward to six to twelve months as issuance bleeds into the Treasury market, Goldman Sachs credit strategist Amanda Lynam took the other side with the issuance math, the chip-depreciation bear case took a data-driven hit as chips run six to nine years, and Meta capex and the power buildout got pinned down.

The AI Capex Tracker

Week of September 24, 2026: A Bear Pulls the AI Debt Reckoning Forward as Goldman Counters With Data

Issue: Thursday, September 24, 2026

(Covers podcasts published in the last day, dated Tuesday, September 23 and Wednesday, September 24. After a week where the whole trade turned on a single question, who's paying for the build, and does it pay off, this cycle finally gave us both halves of the argument from people who actually run the money. Paul Kedrosky is back, on a different show, and he has moved his own doomsday forward: from a "2029 maturity wall" last week to a break "within six to 12 months" now. And for the first time in a while, the bull side isn't a stock chart or a hopeful CEO, it's Goldman Sachs' chief credit strategist, walking through the issuance math line by line and saying, in effect, calm down. Underneath that fight, the other great debate of this build, how fast the chips wear out, got its most rigorous, data-driven airing yet. Six substantive podcasts; the two credit voices are the ones that move a book.)

TL;DR

  • The most rigorous bear on this trade just pulled his crash forward. Paul Kedrosky, the economist who called the plumbing of the 2008 crisis, now says the AI build breaks "within six to 12 months on the outside," not at the 2029 maturity wall he described last week. His new mechanism: AI debt (now "approaching in excess of a trillion dollars," more than 60% debt-financed, up from "15% to 20% a year ago") has grown so large it caused a "treasury market freak out" a month ago, pushing the 10-year "over 5%." Rising rates lift data-center hurdle rates to a level, "11 and 12%" at the stressed end, that "there's no economics that does" (Better Offline, Sept 23).
  • Goldman's chief credit strategist answered him, point for point. Amanda Lynam of Goldman Sachs put hyperscaler capex at "upwards of $6 trillion" for 2026–2030 and forecast "$400 billion of global hyperscaler direct debt issuance" in 2027 (plus "$300 billion" more for data centers and chips). But she sees "very little evidence of a crowding out," argues the treasury move is more about commodities and mechanical hedging than AI supply, and expects the debt-financed share of capex to peak at just "35%" in 2027. Her verdict on the whole complex: constructive (Alpha Exchange, Sept 23).
  • The "chips wear out too fast" bear case took a real, data-driven hit. Azeem Azhar's research finds the math works at a six-year chip life and only "doesn't look great" at four, and the field data keeps coming in long: Amazon "seeing chips in use six, six and a half years," Google running AI chips "eight or nine years old," CoreWeave's six-year chips "100% used" and re-contracted for three more. GPU rental prices are "holding up," and long-term contracts now price above spot, the opposite of a glut (The Most Interesting Thing in AI, Sept 23; The Business Brew, Sept 23).

What's new

A note on the window and the order. This is a Thursday issue covering the podcasts dated Tuesday and Wednesday, September 23–24. The story this cycle is the financing and returns debate finally becoming a real, two-sided fight between heavyweights, so I've ranked by what actually reprices a book: the sharpened bear timeline first, its institutional rebuttal second (these two together are the whole trade right now), then the depreciation debate that decides whether either side is even doing the math right, then the Meta capex detail and the power read.

A quick plain-English key. Hurdle rate: the annual return an investment has to earn to be worth doing; when interest rates rise, the hurdle rises with them. Spread: the extra interest a borrower pays over the safe government rate, it widens when lenders get nervous. Depreciation: spreading the cost of an expensive asset (here, an AI chip) across the years you expect to use it; assume a shorter life and each year's cost looks bigger, which makes the profits look worse. Residual value: what a used asset is still worth. Off-balance-sheet: obligations a company has taken on (leases, purchase commitments) that don't show up as debt on the headline balance sheet. Investment grade / high yield: the safe tier of borrowers versus the riskier ("junk") tier that pays more.

1. Kedrosky moved his own doomsday forward, from 2029 to "six to 12 months." Better Offline, Sept 23, Paul Kedrosky (economist and writer, of SK Ventures), in conversation with host Ed Zitron. (Specialist-investor analysis, not operator testimony, but the most systematic bear framework in this letter, and it just got more urgent.)

Last week Kedrosky gave us a calendar: a "2029 maturity wall," five years out, when a wave of GPU-backed leases all come due at once. This week, on a different show, he brought the date much closer, and a different mechanism to get there.

The new mechanism is the bond market itself. Kedrosky says AI-related debt issuance has become so enormous, "approaching in excess of a trillion dollars," and "more than 60% of AI financing is debt financed up from something like 15% to 20% a year ago", that it is now "the largest piece of the investment-grade marketplace" and "the largest piece of the high-yield marketplace" at the same time. It got big enough, he argues, to "bleed into the treasury market": a month ago there was a "treasury market freak out" that pushed the 10-year "up over 5% now," in part because investors, "given the choice between owning 10-year treasuries backed by the full faith and credit of the United States, or owning 10-year debt backed by a hyperscaler's pristine… credit rating," were choosing the hyperscaler "at the margin." His words: "That just doesn't happen." He calls it AI capex "beginning to compete with the dollar."

Why that breaks things faster. Here is the chain a PM should write down. The 10-year "has gone up like 100 basis points over the last… six months or even last three months." That "directly translates into the funding costs for new data centers", and it doesn't stop there, because in a stressed market the spread widens too, so financing costs can jump "150" or "200 basis points," not 100. Every rate rise, he says, means "it just got a lot more expensive and difficult for data centers to justify the future cash flows." His conclusion: "we're within six to 12 months on the outside of that aspect of it all breaking, because it's very clear that they're not going to earn 11 and 12%, which are the kinds of numbers we're seeing at the most stressed end of the market… There's no economics that does that." He thinks the actual trigger comes "somewhere unexpected", his favorite candidate is "a botched 10-year treasury auction" (insufficient demand at the offered price), which would blow spreads out and "explode your ability to raise debt at any economic price."

The "earnings before bad things" jab. On the bull claim that inference (running AI models, as opposed to training them) is hugely profitable, Kedrosky is scathing: it only looks that way if you strip out the costs. He calls it "earnings before bad things", "EBBT, not EBIT." His challenge: "Show me the hyperscalers that have stopped spending on frontier models, and I'll buy the argument that the economics now make sense purely in terms of commodity inference." And if the business really does collapse into low-cost "industrial inference," then "it's just who can throw the most energy at this", and there, "China's… adding three times as much capacity per year as the United States" and "is already ahead." His read on who wins a pure energy-and-tokens race: "it ain't going to be OpenAI and Anthropic."

A fresh, unverified tell. Kedrosky passed along secondhand chatter that Anthropic's IPO has slipped from "October 1st" to "November," that there is "raging panic inside the company," and that a large vendor told him Anthropic "underperformed their… expectations in the third quarter." He tied it to customer concentration, "two or three customers making up 60 plus percent of your revenues" who had been "token maxing" and "have increasingly cut back," producing "a kind of air pocket later in the year." Treat this as rumor, clearly labeled, but the customer-concentration point rhymes with everything else in this letter.

Why it matters: this is the same analyst who last week gave us a five-year runway, now saying the fuse is under a year. The whole thing hinges on one variable, the 10-year Treasury, which makes rates the single most important thing on every hyperscaler and neocloud screen. If he's right that AI debt is now big enough to move Treasuries, then the AI trade and the bond market have become the same trade.

2. Goldman's chief credit strategist took the other side, with the issuance math in hand. Alpha Exchange, Sept 23, Amanda Lynam, Chief Credit Strategist in Global Investment Research at Goldman Sachs, with host Dean Curnutt. (This is exactly the institutional counterweight the bull case has been missing, a person whose whole job is worrying about getting paid back.)

If Kedrosky is the prosecution, Lynam is the defense, and she came with receipts. Crucially, she does not dispute the scale, she may have the biggest number in this letter.

The size, from the lender's seat. Goldman's equity colleagues put hyperscaler capex "from 2026 through 2030" at "upwards of $6 trillion", larger still if you add 2025 and push to 2031. And hyperscalers are "only 40% of total AI-related gross issuance"; the theme is "much broader." On the debt specifically: hyperscalers "issued $108 billion of global IG debt in 2025," "$229 billion" so far in 2026, and she expects "$250 billion" for the full year, then "a 60% increase in 2027 where we're penciling in $400 billion of global hyperscaler direct debt issuance", plus "another $300 billion" for data-center and chip financing on top. These are enormous numbers, and she says them without flinching.

Why she's still constructive. The key move: she expects the share of capex that is debt-financed to "peak in 2027 at 35%." Note the gap with Kedrosky's "more than 60%", they are measuring different things (his figure is the share of external financing that is debt; hers is the share of total capex funded by debt, with operating cash flow doing the rest), and reconciling the two is arguably the most important number-chase in the whole debate. She also frames the $400 billion against a market that raised its "USIG gross issuance forecast for 2026… to $2.3 trillion" and "$2.4 trillion" for 2027, so hyperscaler supply is "large" and "meaningful," but arriving "in the context of a very robust capital market cycle."

The direct rebuttal to "crowding out." Asked whether hyperscaler borrowing is squeezing out the government or ordinary borrowers, Lynam: "We actually see very little evidence of a crowding out." Her rates colleagues think the run-up in Treasury yields "may actually be more closely linked to the commodity market moves as opposed to hyperscaler supply." Any pressure from big issuance is "mechanical", investors selling Treasuries to hedge new corporate deals, "very rarely would I attribute it to a crowding out from an asset allocation perspective." And under the surface, "the rest of the broader IG market is holding in really well": non-AI spreads are widening only "a little bit," while AI-related moves are "episodic periods of widening around supply indigestion," not fundamental distress. AI is "around 12%" of the US investment-grade index, meaningful, but "not yet the largest sector" (banks are "21%, 22%").

Where she agrees with the bear. This is the tell. The one risk Lynam keeps circling is the same one Kedrosky is pounding: rates. "Is there a tipping point where the rates volatility… increases such that rates become a headwind for credit?" She notes IG now "resembles a rates product", only "15 to 20 percent of the all-in yield" is the credit spread; the rest is the risk-free rate. Her nuance: if yields rise because growth is strong, "credit can digest" it; if they rise "for more negative reasons… maybe it's fiscal concerns," that's a backdrop "credit would not digest as easily." She also flags demand is durable, foreign investors bought "$251 billion" of US corporate credit in the first half of 2026, on pace to beat 2025's record, and own "29%" of the market.

Why it matters: this is the first heavyweight, name-brand credit voice to directly rebut the bubble mechanics, and she does it not by denying the scale but by disputing the consequences. The debate has narrowed to a single, testable disagreement: does record AI issuance break the Treasury market, or does the Treasury market absorb it? Both the top bear and the top bull now agree the answer shows up in one place, rates volatility. Watch that, and you're watching the whole trade.

3. The depreciation bear case got its most rigorous rebuttal, and a new market to price it. The Most Interesting Thing in AI, Sept 23, Azeem Azhar (analyst/researcher) with Nicholas Thompson; and The Business Brew, Sept 23, Steve Hou of Silicon Data.

The whole bear case has a hidden load-bearing assumption: that AI chips wear out (or go obsolete) fast, say, four years, which would make the hyperscalers' reported profits far too rosy. Azhar built a model of the entire AI economy and found the answer turns almost entirely on this one number. At a six-year chip life, "the math probably works." At four years, "it doesn't look great." So which is it?

Azhar sided with the data, not the theory. The theory (and Nvidia's own Jensen Huang, who "himself has said… we make our products obsolete") argues for a short life. But the field evidence keeps coming in long: Amazon "seeing chips in use six, six and a half years after we have bought them"; Google's infrastructure chief citing AI chips "eight or nine years old that were fully in use"; and neocloud CoreWeave with six-year-old chips "that were 100% used, that they had just contracted for another three years." He frames it as "empiricist" versus "rationalist", does it matter whether it works in theory, or in practice?, and lands on practice: "these chips are churning out revenues."

The pricing backs it up, and points the other way from a glut. GPU rental prices "held up reasonably well," and here's the counterintuitive part: "long-term contracts are often at a higher price than the spot price, which is the reverse of what you would think." That's a sign of scarcity, not oversupply, buyers paying up to lock in surge capacity (Azhar cited the expensive compute-rental deals struck with Anthropic and Google). His caution: this "muddies the water," because "we know what happens the moment capacity turns… if there's too much capacity, prices will collapse."

The independent corroboration. Over on The Business Brew, Steve Hou of Silicon Data, the firm that actually tracks GPU rental prices, made the same point from the data vendor's seat. "Over 50% of the cost of a data center today is chips and memory," so how you value that used hardware decides whether the loans against it are sound. Early on, lenders "applied a very punitive… schedule"; now, with "a few years of observing… rental income data," the "realized residual value implied by the currently observable rental income is quite a bit higher, substantially higher, than what's implied by those conventions." His honest caveat: "we haven't gone through a full business cycle… What happens if demand normalizes and supply catches up?" The genuinely new, actionable nugget: Silicon Data and the CME have an exclusive partnership to launch two cash-settled futures contracts, an H100 (Hopper) and a B200 (Blackwell) neocloud rental-price index, "the first… institutional grade financial instruments with which people that have AI exposure… can hedge that risk." Compute is becoming a traded commodity.

The honest other half. Azhar is not a cheerleader, he calls it "a plane that is being pushed at full throttle at a very high altitude." His demand check is sobering: outside China, real AI-services spend was about "$126 billion in the 12 months to July 2026" (a deduplicated, value-add figure), versus AI capex now "in excess of $100 billion" in 2026 alone and "much larger than the top-line revenue." And the productivity reality is thin: at a room of "160 VPs of IT," two-thirds said they had "measurable results," but only about "eight… out of 150" had results good enough to "disturb your CEO on their summer vacation." He cites a BCG survey where "70%" of CEOs said AI matters for them "to be perceived to succeed", an "Upton Sinclair moment," where people have every incentive to talk it up. His net: "finely balanced… the learning occurs more quietly than the fanfare headlines do."

Why it matters: the depreciation fight is where the bull and bear literally use different spreadsheets. Azhar and Hou just handed the bulls their best evidence, real-world chip lives running well past accounting schedules, and residuals holding up, which is exactly what you need to believe if you think the hyperscalers' profits are real. But both flagged the same fragile assumption: it all holds until demand normalizes. The CME futures matter because, for the first time, the market will publish a live price on that question.

4. Meta's capex bill got pinned down, $130–145B this year, ~$200B next, and the "why" is a decade-long platform bet. TIP848, The Investor's Podcast, Sept 24, Daniel Mahncke and Shawn O'Malley (tech/investing analysts).

The freshest episode in the batch (dated today) dug into the one hyperscaler the market actually rewarded last week. Meta's 2026 capex guidance "has been raised multiple times and is now between 130 and $145 billion," with analysts projecting 2027 "close to $200 billion." The hosts frame this as Zuckerberg's "third shot" at building a platform, after the Farmville-era desktop apps and the ~$100 billion Metaverse, and this time he "clearly sees AI spend as a defense mechanism."

The bull half is simple and immediate. AI "can turn every pixel of your screen into an ad", near-infinite new inventory in Meta's best business. "You don't need superintelligence or the next computing platform or cloud business… It's just better algos and more inventory, and that would technically be enough." And they made a point that cuts against the whole "circular financing" worry: Meta "is sort of the only company that is not partnered with OpenAI or Anthropic," so "there's no circular financing, there are no stakes in trillion dollar companies that never earned a dime of cashflow", possibly "even an advantage." Meta also owns real assets in the stack: its own custom AI chips, and PyTorch, the open-source framework "that basically the entire AI world builds on."

The bear half is the enterprise dream. The hosts are skeptical of the part of the spend justified by selling cloud compute and an enterprise business, where Meta "has no Salesforce, no support organization, no compliance track record", the very things that took Google Cloud "more than a decade and tens of billions of dollars in losses" to build. Zuckerberg's own answer on the earnings call ("the enterprise opportunity is kind of the sum of all of these different things… going to be somewhat of a new muscle that we build") struck them as too vague "to justify hundreds of billions of dollars in capital expenditures." Their blunt line: if Zuckerberg "just wanted the stock to rally… he would just cut CapEx in half and frame Meta as this advertising beast that it certainly is." Their likely path if it works at all: not the Fortune 500 but "the hundreds of millions of small businesses" already on WhatsApp and Instagram, "almost certainly something that takes you five to 10 years."

The number that stops you. Citing a Wall Street Journal deep dive, the hosts said Alphabet alone "has about a trillion dollars in off-balance sheet liabilities… through leases that haven't started yet and… purchase commitments for things like energy." That is the off-books build, quantified, and it ties straight back to Kedrosky's and Lynam's arguments about where the real leverage is hiding. Their summary of the whole cycle: "there's no way back. We've gone off the cliff", unlike the metaverse, you can't just declare a "year of efficiency."

Why it matters: this pins Meta's actual dollar commitment and separates the part of it the market will pay for (ads, now) from the part it won't yet (enterprise, someday). And the Alphabet off-balance-sheet trillion is the concrete version of the bears' central worry, the leverage that doesn't show up where you'd look for it.

5. A power operator drew the line on the demand hype: "Physics is going to win." Let's Talk Energy, Sept 23, Robert Gaudette, CEO of NRG (competitive power generator: "8 million customers, 17,000 employees, and 25 gigs of power generation").

Amid the runaway data-center-power forecasts, an operator who actually builds and dispatches generation offered a reality check that cuts both ways. It is "a demand super cycle… no question," with US demand growing "about 2.5% annually" (up from "1% on a good day" for two decades). But the wilder numbers are fantasy: on ERCOT, "there's a number out there that says 500 gigawatts. Now that's on a system that has 91 [GW peak]… That's not going to happen. Physics is going to win." His realistic figure: ERCOT peak load "increase by about 40 gigawatts" by 2030, huge, but a fifth of the hype.

The trade is behind-the-meter gas. NRG's "Bring Your Own Power" model requires a large data-center load to build "1.2 gigawatts of supply for every one gigawatt of load" in combined-cycle gas, so the site strengthens the grid instead of straining it. The first project is "1.2 gigawatts"; by 2029 it's expected to throw off "$375 million in free cash flow… about 10% of [NRG's] total free cash flow," with an "achievable pipeline" of "5.4 gigawatts." He signs only "investment grade counterparties with contracts," on assets with "a 40 year life." And the binding constraints are physical and real: "there is a turbine constraint… high load switch gear problems… definitely labor constraints," with the engineering firms that build these plants "completely booked."

Why it matters: this is the sober version of the power trade. It confirms the super-cycle (bullish for gas generation, turbines, and grid equipment) while puncturing the 500-GW dreamscape (a caution on anyone underwriting infinite demand). The actionable read is behind-the-meter gas plus the equipment bottleneck, turbines, switchgear, and the electricians to install them.

The debate

Steel-manning both sides of the $700B-plus 2026 hyperscaler capex thesis, with the useful twist that this cycle gave us the sharpest bear and the most credible bull, arguing about the same facts.

Bull, the scale is real, the debt market is deep and eager, and the chips last longer than the bears assume. Goldman's Lynam doesn't deny "$6 trillion" of capex or "$400 billion" of 2027 hyperscaler issuance; she argues the market absorbs it, sees "very little evidence of a crowding out," pins the debt-financed share of capex at just "35%," and reports foreign demand ("$251 billion" in H1) running at a record (Alpha Exchange, Sept 23). Azhar and Hou supply the other pillar: chips are lasting "six, six and a half," even "eight or nine" years, residual values are "substantially higher" than conservative schedules assumed, and long-term GPU contracts price above spot, scarcity, not glut (The Most Interesting Thing in AI, Sept 23; The Business Brew, Sept 23). And Meta shows a hyperscaler with an immediate, proven payoff (ads) sitting on top of the spend (TIP848, Sept 24).

Bear, the debt is now big enough to break the bond market, and the fuse is under a year. Kedrosky's new frame is that AI debt ("in excess of a trillion," "more than 60%… debt financed") has grown into "the largest piece" of both the IG and high-yield markets and started to "bleed into the treasury market," pushing the 10-year "over 5%." Rising rates lift hurdle rates to "11 and 12%" that "there's no economics" to meet, and the break comes "within six to 12 months", plausibly via "a botched 10-year treasury auction." Profitability claims are "earnings before bad things," and if it all collapses to commodity inference, China's 3x capacity advantage wins (Better Offline, Sept 23). The off-books leverage is real and large, Alphabet's "~$1 trillion" of off-balance-sheet liabilities (TIP848, Sept 24).

The synthesis. For once, the two best voices agree on the fact (record borrowing) and even on the tell (rates volatility), they just disagree on the outcome. Kedrosky says the AI-debt pile is now so big it will crack the Treasury market from the inside; Lynam says the Treasury market is deep enough to swallow it and the yield move is mostly about commodities and mechanical hedging. Underneath, the depreciation data has quietly shifted toward the bulls, if chips really last six-plus years and hold their value, the reported profits are more real than the bears' four-year math allows. The tie-breaker is a single number you can watch every day: the 10-year Treasury and the size of the auctions. If an auction stumbles or rate vol spikes for "fiscal" (not "growth") reasons, Lynam's own red line, the bear wins on his own clock. If the book keeps clearing at these yields, the bull's "supply indigestion" framing holds.

Sell signals to watch: - A weak or "technically delayed" 10-year Treasury auction. Kedrosky's named trigger; both he and Lynam agree rates are the fault line. A tail or a delay is the first hard evidence his six-to-12-month clock is real (Better Offline, Sept 23). - Rate vol rising for "fiscal," not "growth," reasons. Lynam's explicit line in the sand, the backdrop credit "would not digest as easily." Watch MOVE-style rate volatility and the driver behind any yield back-up (Alpha Exchange, Sept 23). - AI credit spreads widening beyond "episodic." Lynam's tell that supply indigestion has become something worse; watch hyperscaler and especially neocloud (CoreWeave) spreads decouple from the broad IG index (Alpha Exchange, Sept 23). - GPU rental prices rolling over. The moment "capacity turns," Azhar says, "prices will collapse." The new CME H100/B200 rental futures will make this visible in real time, the cleanest live read on the depreciation debate (The Most Interesting Thing in AI, Sept 23; The Business Brew, Sept 23). - Neocloud IPO slippage / frontier-lab air pockets. Kedrosky's (labeled) rumor of Anthropic's IPO delay and a soft Q3 on customer concentration, if any of it surfaces in the data, it validates the "two or three customers = 60%+ of revenue" fragility (Better Offline, Sept 23).

Stocks in play

NVDA. Bull: the depreciation data is quietly bullish, if chips run "six, six and a half" or even "eight or nine" years and hold their value, the installed base keeps paying, and Nvidia's own Jensen Huang concedes "we make our products obsolete," which only accelerates the refresh cycle (The Most Interesting Thing in AI, Sept 23). Bear: Kedrosky's framing makes Nvidia the epicenter of a debt-driven trade, GPUs so central they were "19%" of global goods trade, and warns that in a pure "industrial inference" race, "China's… adding three times as much capacity per year" (Better Offline, Sept 23). Next catalyst: Micron's Sept 30 print as the demand read-through; any China export-policy resolution.

AVGO. Bull: the custom-ASIC arms dealer, Meta's reliance on "their own custom AI chips" is the structural tailwind for merchant-silicon design partners (TIP848, Sept 24). Bear: sits inside the same debt-and-hurdle-rate math as the whole complex; no company-specific rebuttal this cycle. Next catalyst: the next hyperscaler custom-ASIC order update. Largely quiet on the podcasts this cycle.

AMD. Bull: still the credible number-two accelerator, leveraged to the same "chips are lasting longer / demand is real" data. Bear: as the higher-beta AI-hardware name, the most exposed to Kedrosky's rate/refinancing air pocket. Next catalyst: a company MI450X/Helios deployment update and a named frontier anchor customer. Quiet on the podcasts this cycle.

MSFT. Bull: the balance sheet to carry debt through a repricing, and Azhar's demand work notes Microsoft is "making less" from AI than its total revenue only because disclosure is conservative, the AI revenue is real and growing (The Most Interesting Thing in AI, Sept 23). Bear: squarely inside the "60% debt-financed / hurdle rates rising" critique, and the GPU-backed-lease structure is exactly what Kedrosky says reprices (Better Offline, Sept 23). Next catalyst: Azure growth against the capex line. No company-specific news this cycle.

GOOGL. Bull: the depreciation debate's best exhibit, Google's infrastructure chief citing AI chips "eight or nine years old that were fully in use", and "Google Cloud is currently exploding due to Gemini" (The Most Interesting Thing in AI, Sept 23; TIP848, Sept 24). Bear: the poster child for hidden leverage, "about a trillion dollars in off-balance sheet liabilities" through unstarted leases and energy purchase commitments (TIP848, Sept 24). Next catalyst: free-cash-flow trajectory against the rising capex line; disclosure on those off-book commitments.

AMZN. Bull: the depreciation anchor, Amazon's own "chips in use six, six and a half years after we have bought them," well past its ~5-year accounting schedule, and AWS is the offtake engine (The Most Interesting Thing in AI, Sept 23). Bear: carries the heaviest absolute capex to justify, inside the same rate math. Next catalyst: proof AWS demand is broad beyond the frontier labs.

META. Bull: the clearest capex-with-a-payoff story, $130–145B this year toward ads that "turn every pixel… into an ad," no circular-financing entanglements (uniquely "not partnered with OpenAI or Anthropic"), and it owns real stack assets (custom chips, PyTorch) (TIP848, Sept 24). Bear: the ~$200B 2027 number rests partly on an unproven enterprise/cloud dream where Meta has "no Salesforce, no support organization, no compliance track record," and Zuckerberg's justification struck the analysts as too vague, "cut CapEx in half" and the stock arguably rallies (TIP848, Sept 24). Next catalyst: whether the enterprise pitch gets concrete; capex-to-ad-revenue conversion.

Also in play: CoreWeave, cited on both sides: Azhar's evidence that its six-year chips are "100% used" and re-contracted for three more years (bull for residuals), and Kedrosky's archetype of a debt-laden neocloud most exposed to a rate shock (bear) (The Most Interesting Thing in AI, Sept 23; Better Offline, Sept 23). NRG, "Bring Your Own Power" behind-the-meter gas is expected to reach "$375 million in free cash flow" (~10% of FCF) by 2029 on a "5.4 gigawatt" pipeline (Let's Talk Energy, Sept 23). CME / Silicon Data, launching the first H100 and B200 GPU-rental futures, making compute a traded, hedgeable commodity (The Business Brew, Sept 23). MU (Micron), reports September 30; no fresh podcast color this cycle, but it remains the next hard catalyst on the memory question. Anthropic, Kedrosky's (labeled, unverified) rumor of an IPO slip to November and a soft Q3 on customer concentration (Better Offline, Sept 23).

Read-throughs

  • Rates / credit, this is the entire cycle, and the trade is a spread, not a stock. The single most actionable thing this cycle is to watch the 10-year Treasury and its auctions. Kedrosky says AI debt is now big enough to break them "within six to 12 months"; Lynam says the market absorbs it, but both name rate volatility as the fault line. Watch AI-related IG spreads (especially neocloud) for a move beyond "episodic," and watch whether any yield back-up is "growth" (digestible) or "fiscal" (not). If you own the AI complex, this is your hedge leg (Better Offline, Sept 23; Alpha Exchange, Sept 23).
  • GPU residuals / depreciation, now a tradeable signal. The bull-bear fight over chip life just moved from spreadsheets to a live market: the CME's coming H100/B200 rental-price futures will publish, in real time, whether GPU rents are holding (bull) or rolling over (bear). Until then, the field data (Amazon 6.5 yrs, Google 8–9 yrs, CoreWeave re-contracting six-year chips) and Silicon Data's "substantially higher" residuals favor the bulls, with the standing caveat that it all holds "until demand normalizes" (The Most Interesting Thing in AI, Sept 23; The Business Brew, Sept 23).
  • Power / thermal, behind-the-meter gas, and the equipment bottleneck. NRG's read caps the runaway demand story ("40 gigawatts," not 500, on ERCOT) but confirms the super-cycle and points the trade at behind-the-meter combined-cycle gas (1.2 GW built per 1 GW of load). The gating constraints, "turbine constraint… high load switch gear… labor" with EPCs "completely booked", are the actionable read: gas-turbine supply and electrical equipment (VRT, ETN) sit on the binding bottleneck. Nuclear (Vistra, Constellation, Talen) was quiet this cycle; NRG is talking to SMR/geothermal players only for offtake, not builds (Let's Talk Energy, Sept 23).
  • Grid, globally, the queue is the story. From Climate Week NYC, the UK grid operator (NESO) pegged its connection queue at "125 gigawatts of new demand… a large part data center driven" against a "45 gigawatt" peak, the same load-vs-capacity mismatch as ERCOT, on a smaller grid. A useful reminder the constraint is worldwide. (One wry aside from the same panel: the IEA's Fatih Birol reportedly still ranks air conditioning, not AI, as "the single biggest contributor to power demand growth", worth remembering before every megawatt gets attributed to GPUs) (Energy Gang, Sept 23).
  • Memory / networking / optics, quiet, with the catalyst on the calendar. No fresh operator color on Micron/SK Hynix HBM, Marvell/Astera/Credo networking silicon, or Coherent/Lumentum/Fabrinet optics this cycle. Micron's Sept 30 print is the next hard read on whether the memory tightness is holding.

What changed vs last issue

Last issue (Wednesday, Sept 23, "The bill comes due in 2029. Money floods in anyway.") covered the Monday, Sept 22 podcasts: Kedrosky's 2029 maturity-wall thesis (from Between Two COO's), Alibaba's "$53 billion" plan and Nvidia-challenger chip, SoftBank's OpenAI bond drawing ">$20 billion" of demand, and Meta's Muse hitting "number one on the App Store." This edition covers Sept 23–24, and the financing debate went from one-sided to a genuine heavyweight fight.

  • Kedrosky moved his own date up, a lot. Last issue he gave a "2029… five-year maturity wall." This cycle, on Better Offline, he says the break is "within six to 12 months," via a new mechanism (AI debt bleeding into the Treasury market, the 10-year "over 5%," unmeetable hurdle rates). Same analyst, far shorter fuse (Better Offline, Sept 23).
  • The bull case finally got an institutional voice. New this cycle: Goldman's Amanda Lynam directly rebutting the crowding-out thesis with issuance data, "$6 trillion" capex 2026–2030, "$400 billion" 2027 hyperscaler issuance, but "35%" debt-financed and "very little evidence of a crowding out." Last week the bull side was a stock chart (Micron) and a hopeful CEO (Meta's Muse); this week it's a chief credit strategist (Alpha Exchange, Sept 23).
  • The depreciation debate got real evidence, favoring the bulls. New this cycle: Azhar's six-vs-four-year model and the field data (Amazon 6.5 yrs, Google 8–9 yrs, CoreWeave), plus Silicon Data's "substantially higher" residuals and the CME's coming GPU-rental futures. This is the first time the "chips wear out too fast" bear case has been met with hard, contradicting numbers (The Most Interesting Thing in AI, Sept 23; The Business Brew, Sept 23).
  • Meta's capex got pinned, and a new leverage number surfaced. New this cycle: "$130 to $145 billion" 2026 / "$200 billion" 2027 for Meta, plus Alphabet's "$1 trillion" of off-balance-sheet liabilities (via WSJ), the concrete version of the bears' hidden-leverage worry (TIP848, Sept 24).
  • The aggregate capex anchor firmed up. Goldman's institutional "$6 trillion, 2026–2030" roughly corroborates the weekend's Oracle-CEO "$5 trillion, 2025–2030", two independent, similar-magnitude anchors now.
  • Power shifted from "cuts your bill" to "physics wins." Last week's grid story was Edison's "$1.4 trillion" and the argument that data-center load can lower rates; this week NRG reframes it as behind-the-meter gas and a hard cap on the hype ("40 GW," not 500) (Let's Talk Energy, Sept 23).
  • Still open from last issue: the SoftBank/OpenAI bond was set to price this Thursday (today), no Sept 23–24 podcast carried the pricing outcome, so it remains the near-term credit tell to confirm. No fresh podcast follow-up on Alibaba's chip or the Trump–Xi meeting.
  • Quiet this cycle: AVGO, AMD (no company-specific color); memory/optics operator commentary (Micron's Sept 30 print is the catalyst); dedicated nuclear PPAs (Vistra, Constellation, Talen).

Next catalysts to watch: Micron earnings (Wed, Sept 30) as the memory verdict; the SoftBank/OpenAI bond pricing outcome (was due today) as the live credit tell; any 10-year Treasury auction stress, Kedrosky's named trigger; and the CME/Silicon Data GPU-rental futures launch as the first real-time price on the depreciation debate.