Newsletter · · Ashutosh Agarwal

Oracle's Catz Puts Hyperscaler Capex at $5 Trillion as Off-Balance-Sheet Debt Grows - The AI Capex Tracker - Week of September 21, 2026

The AI Capex Tracker for the week of September 21, 2026, synthesizing investor podcasts published September 18 and 19. This issue covers Oracle CEO Safra Catz's $5 trillion 2025-2030 capex figure for the four biggest hyperscalers, Robin Wigglesworth and PGIM's Greg Peters invoking the railroad boom as hyperscaler debt and off-balance-sheet leases pile up, Tom Lee and Dan Ives making the earnings-not-bubble case, Dan Niles naming fresh cracks in AI demand, and Microsoft stretching server depreciation from 15 to 25 years.

The AI Capex Tracker

Week of September 21, 2026: Oracle's Catz Puts Hyperscaler Capex at $5 Trillion as Off-Balance-Sheet Debt Grows

Issue: Monday, September 21, 2026

(Covers podcasts published over the weekend and Friday evening, September 18–19. The weekend's through-line was a number and a warning about how it gets paid for. Oracle's CEO, quoted by adman Sir Martin Sorrell, put the four biggest cloud builders' spending at $5 trillion between 2025 and 2030, with this year's figure already revised up from about $630 billion to $830 billion, and next year pegged near $1.1 trillion. At the same time, two separate voices, a financial historian and the co-head of a $1.2 trillion credit fund, independently reached for the same analogy: the 1800s railroad boom, when the building was financed with debt, and the debt is what made the bust so much worse. Against that, the trade's loudest bulls, Tom Lee and Dan Ives, made their strongest case yet that this is an earnings story, not a bubble.)

TL;DR

  • An operator just put the biggest number yet on the build, and it's an escalation. Oracle CEO Safra Catz, relayed by Sir Martin Sorrell, said Microsoft, Amazon, Alphabet and Meta will together spend $5 trillion on capex from 2025 to 2030; this year's forecast for the four has already climbed "from about $630 billion… up to $830 billion," and next year is pegged at "$1.1 trillion. Just those four" (The Geopolitics of Business, Sept 18).
  • Two independent voices reached for the 1800s railroad boom, because this build now runs on debt, and a lot of it is hidden. Hyperscalers are on track to sell roughly "$500 billion of bonds this year," nearly double last year, and there's "another $1.5 trillion worth of lease commitments… of which only $500 billion are actually on their balance sheets," with the rest tucked in footnotes and joint ventures (Simply Put, Sept 18). PGIM Credit's co-CIO, running $1.2 trillion, calls it "the early stages of this credit crowding out… a quantum of debt in ways that we haven't seen… since the buildout of the US railroad system" (Credit Exchange, Sept 18).
  • The bulls hit back hard: this is earnings, not mania. Dan Ives says chip demand-to-supply is "13 to 1" and won't reach balance "till early 2029," with "a five to six dollar multiplier" across tech for every capex dollar; Tom Lee's line was the mood in one sentence: "you're never at a top when people want [the] P/E to be low and earnings are going up" (The Compound and Friends, Sept 18).

What's new

The story since Thursday's Fed hike has shifted from "the money just got more expensive" to a sharper question: how big is this build really, and how is it being paid for? I've ranked by what hits a hedge fund book hardest: the new, bigger capex number first (it resets the whole trajectory this letter tracks); then the debt-and-hidden-leases story that two serious people independently flagged; then the bulls' rebuttal; then a respected trader's "speed bump, not top" nuance and the cracks he's watching; then the accounting sleight-of-hand that's flattering this quarter's earnings.

A plain-English key, since these terms recur. Capex is money spent on long-lived assets: data centers, chips, power plants. Hyperscaler is a giant cloud owner: Microsoft, Google, Amazon, Meta (Oracle now often lumped in). Free cash flow is the cash a company has left after paying for its capex; when capex explodes, free cash flow can go negative even while profits look great. Depreciation is how a company spreads the cost of a long-lived asset across the years it expects to use it; stretch the years, and each year's reported expense shrinks. Off-balance-sheet means a real financial obligation that doesn't show up in the headline debt figure because of how the deal is structured (a lease, a joint venture). Open-weight model is an AI model whose guts are published so anyone can run it, the free, commoditizing competitor to the paid frontier labs. Basis points (bps): 100bps = 1 percentage point.

1. Oracle's Safra Catz put a $5 trillion number on the build, and this year's figure already jumped $200 billion. The Geopolitics of Business, Sept 18: Sir Martin Sorrell, founder of S4 Capital and WPP, relaying figures he attributes to Oracle CEO Safra Catz, in conversation about AI and global advertising.

This is the item that resets the trajectory. Sorrell, describing a breakfast with a senior Alphabet engineer and citing Catz's framing, laid out the arithmetic bluntly: "Microsoft, Amazon, Alphabet and Meta, those four… are going to spend $5 trillion between 2025 and 2030." The near-term revision is the part that matters for numbers this quarter: "going into Q2, the forecast for this year for those four was about $630 billion. And it's been up to $830 billion for this year after Q2. And for next year, it's forecast at $1.1 trillion. Just those four." His per-company sketch for this year: "Google is spending $200 billion… Amazon, $240 billion. Meta, $140 [billion]. And Microsoft, around $200, $250." And the tell on why the bulls aren't blinking: the engineer's "basic belief is that demand, that pipeline is there. So we are at the beginning" of the build, not the end.

Why it matters: the aggregate capex figure this letter has tracked keeps climbing, and it now has a named-operator source rather than a strategist's estimate. Last issue's anchor was "$700 billion in 2026 heading to $1 trillion in 2027," with Gerstner floating "$1.5 trillion a year." Catz's numbers ($830 billion this year for four names, $1.1 trillion next) say the run-rate is arriving faster than the old framing implied. That is bullish for the arms dealers who get paid on the spend (Nvidia, memory, power kit) and it raises the stakes on the only question that matters: whether the revenue ever shows up to justify it.

2. The build now runs on debt, and a lot of it is hidden. Two serious people reached for the same railroad analogy. Simply Put, Sept 18: Robin Wigglesworth, financial historian and FT journalist, on how the AI build echoes the 1800s railroad boom. Credit Exchange, Sept 18: Greg Peters, co-CIO of PGIM Credit, which runs $1.2 trillion, interviewed by Lisa Lee.

Here is the reframe that should worry a bull. Wigglesworth walked through how the funding has changed: "a few hundred billion dollars worth of bonds sold last year… now that bond boom is getting bigger. I think we're approaching $500 billion of bonds sold this year by just the hyperscalers. So that's almost twice what it was last year." And the part that doesn't show up in the headline debt figures: "there's probably another $1.5 trillion worth of lease commitments by the hyperscalers, of which only $500 billion are actually on their balance sheets… The other stuff you have to find in footnotes. Debt by another name, essentially."

His marquee example is one to file away. Meta is building a data center in Louisiana called Hyperion, "going to require the equivalent, the energy of five Hoover dams." Rather than borrow directly and risk its credit rating, Meta "set up a joint venture with Blue Owl, where they only take 20% of the equity," then "made an irrevocable guarantee to lease that data center for 20 years." The JV used that guarantee to "sell a $27 billion IG bond… the single biggest corporate debt security in the world right now… issued by this kind of obscure JV. And that debt is nowhere on Meta's balance sheet."

"Equity market bubbles are far less destructive than we think. But debt-fueled bubbles, whether it's for capex or anything else, are far more debilitating. The crash is so much nastier. And the hangover is much, much longer." (Robin Wigglesworth, Simply Put, Sept 18)

His central point is that this is the crucial difference from the dot-com bust, which was financed mostly with equity and so produced only "a bit of a nothing burger" of a recession. He's careful, "I don't think this is remotely the next big financial crisis," but he notes the tell that credit investors are already nervous: "we actually have credit default swaps trading on some of these companies now. That would have been preposterous… a year ago." In plain terms, a credit default swap is insurance against a company failing to pay its debt; the fact that traders now bother to buy and sell it on hyperscalers is itself a signal.

The striking thing is that Greg Peters, a credit manager with $1.2 trillion, not a journalist, independently used the same historical touchstone the very same day. Asked why the 10-year Treasury yield is near 5%, its highest since 2007, he named strong nominal growth ("six and a half percent"), sticky inflation "well away from the 2% target," and the deficit, then added: "you do have… the early stages of this credit crowding out. So this hyperscaler AI infrastructure, digital infrastructure buildout is a quantum of debt in ways that we haven't seen in… a very, very long time since the buildout of the US railroad system." He also flagged that his shop made an out-of-consensus call this summer: "we reversed our call from rate cuts to rate hikes… we were somewhat ridiculed for calling for three rate hikes this year."

Why it matters: this is the bear case maturing from "the multiple is stretched" into "the funding structure is fragile and partly invisible." When both a financial historian and a $1.2-trillion credit investor reach for the railroad boom on the same day, the analogy has left the fringe. The actionable read is Wigglesworth's own split, which echoes Bassman last week: the high-quality hyperscaler bonds probably clear, but the risk sits in the equity multiple and in the weaker players, and the "$1.5 trillion of lease commitments, only $500 billion on the books" gap is now something to hunt for in the footnotes before you underwrite any of these names.

3. The bulls' rebuttal: 13-to-1 demand, a five-to-six-dollar multiplier, and earnings the Street is still lowballing. The Compound and Friends, Sept 18: Tom Lee (Fundstrat) and Dan Ives (Wedbush), the trade's two most visible optimists.

If the debt story is the bear's best new material, this is the bull's. Ives's core claim is that the shortage is nowhere near over: "demand to supply right now for chips is… 13 to 1… there are 13 orders for every one chip that can be produced," and "you're not going to have true equilibrium probably till early 2029 at this pace." He argues the spend cascades far beyond the chipmakers: "for every dollar spent on capex there's a five to six dollar multiplier across the rest of tech… software… infrastructure… cybersecurity." His punchline for anyone worried about valuations: "I think investors are underestimating earnings by probably 25-30% in the next few years," and he framed the build as barely begun: "you have 12 to 1500 data centers that are going to be built in the next five, six years."

Tom Lee's contribution was the cleaner logical point. On why valuations are contracting even as earnings soar (this year "earnings are up 25% and the market's only up 10%," so "the market got cheaper"), he offered the line that captures the whole bull mood: "you're never at a top when people want [the] P/E to be low and earnings are going up. Unless it's a true deep cyclical." He named the two things that would actually end the bull market: a genuine bubble ("if one day everyone says there's only one AI model we want to use, and then capex goes to zero") and the Fed ("80% of the time [it's] the reason a bull market ends"). Notably, he waved off the debt-bubble worry with a nuance the bears would contest: "a debt bubble takes place when there's… return-free risk… the cost of money is so low that bad projects are getting funded. What we're seeing today is data center cost of money is going up. The market is already putting a brake on everything. So we're not able to get to a debt bubble."

Both men spent real time on the weekend's other drama, Anthropic's Dario Amodei calling for a pause on AI development, and the pushback from Jensen Huang, Satya Nadella and (via a phone-in) the White House. Their read was cynical and worth hearing: Ives sees regulatory capture ("when you get to the penthouse, then you stop the elevator so others can't come up"), while Lee framed it as marketing tied to the spend: "it's 800 billion a year is the capex, and it's going to go to 1.1… They make public statements to influence politicians and voters, not just shareholders." Both insisted the real risk is political: data-center permitting fights and moratoriums that, in their telling, hand the lead to China.

Why it matters: this is the strongest, most specific bull artifact of the weekend, and it squarely answers the debt bears, not by denying the leverage, but by arguing the demand and earnings power are so far ahead of supply that the projects are good, not "return-free." The two testable claims a PM can track: whether the "13-to-1" backlog holds (watch lead times out of Taiwan and Korea) and whether the coming quarters keep beating estimates by the margin Ives claims. Lee's "cost of money is going up, so we can't get a debt bubble" is the line the bears will most want to test, because rising rates plus $500 billion of annual issuance is exactly the squeeze they're pointing at.

4. Dan Niles is still in the "speed bump, not top" camp, but he's naming fresh cracks. RiskReversal Pod, Sept 18: Dan Niles, veteran tech portfolio manager (Niles Investment Management).

Niles called the summer's 29% semiconductor drawdown (June 22 to July 29) a "speed bump" back on June 20, and he's still there: "I'm in the speed bump category… we're just so early in the agentic AI trade." But the value of this interview is the honest list of what's changed for the worse. Three things put him on alert. First, politics: "the state of Texas… started to put a moratorium on data center expansion… that's really unusual because it's a deeply red state and they've been very supportive." Second, the founders' own "slow down" talk (Altman, Amodei, Musk), "not something you expected to see." Third, and most concrete, the economics of the models themselves: the average cost per token charged "has gone down 50% since the end of May," and while total token volume "quadrupled," usage is migrating to cheap open-weight (mostly Chinese) models, and "supposedly OpenAI's profitability actually got worse" after it cut prices to take share from Anthropic.

His single sharpest data point is a leading indicator worth adding to a dashboard: "the top 1% of users… actually saw a 10% reduction in spend from July to August, if you look at the ramp index," the first monthly decline since ChatGPT launched, even as the other 99% keep spending more. On the macro overlay, he's blunt: with the 10-year near 5%, "Google went cash flow negative for the first time since they went public," inflation has run above target "for 65 months," so "you should expect a series of rate hikes," and the average S&P drawdown from July 31 to November 9 in midterm years "is about 10%… double what it is in non-midterm years." He also flagged the Japanese carry-trade unwind, borrowing cheaply in yen to buy U.S. assets, as a bigger tail risk now that Japanese yields are the highest since the 1990s.

Niles's most useful insight cuts through the "slow down AI" theater: "two things can be true at once. You can be concerned about what this technology can do. And at the same time… want to have a government-sanctioned duopoly." His three-part motive read on the pause calls (genuine safety worry, a bid to choke off open-source rivals, and a desire to show better cash flow before OpenAI and Anthropic go public) is the most level-headed take of the weekend.

Why it matters: Niles is neither a permabull nor a permabear, which makes his crack-list the most credible risk map on offer. The cost-per-token collapse and the 1%-spend dip are exactly the "capex-to-revenue conversion" signals this letter watches: early, soft, but pointing the wrong way. His bottom line is a useful frame: Nvidia at "a mid-teens multiple for calendar '27" is "ridiculously cheap" if the estimates hold, but "a lot of companies [are] over-earning… memory is a great example," so the risk is in the estimates, not the multiple.

5. The depreciation debate got concrete: Microsoft just stretched its server life from 15 years to 25. Deep Values, Sept 18: an analytical breakdown of the mid-year hyperscaler and Chinese-tech numbers.

This letter has flagged the depreciation debate as abstract for weeks; here it is with numbers. Microsoft "changed their accounting rules regarding the physical lifespan of their servers… extended the depreciation life of their data center equipment from 15 years to 25 years." The hosts' plain-English analogy is the right one: "it feels exactly like refinancing your 15-year mortgage into a 25-year mortgage. Your monthly payment drops significantly, which makes it look like you have way more disposable cash flow today. But the actual house still costs the exact same amount." The effect is real and mechanical: a smaller depreciation charge each quarter lifts reported earnings per share now, "critics are aggressively arguing that this lowers the quality of their earnings," while management counters that modern data centers genuinely last longer. Either way, the timing is "convenient… right at the exact moment your capital expenditures hit all-time highs."

The same segment is a useful reminder of just how much cash the build is eating even at these companies. Amazon's free cash flow swung from "positive $18.2 billion" a year ago to "negative $7.6 billion" (a ~$26 billion swing) after "$54.2 billion" of capex in a single quarter and a plan to spend "$220 billion for the full year," even as AWS printed a record 39.4% operating margin. Microsoft's own free cash flow fell 23% to about $19.2 billion despite Copilot reaching "30 million paid seats." And the piece underlined why Nvidia's 75% gross margin is the tax everyone is trying to escape: Microsoft's push into its own "Maya 200" chips is a bid to "drop the cost per token… by 30%" and keep that margin for itself.

Why it matters: the depreciation change is a quiet lever on reported profits across the group, and now that Microsoft has pulled it, watch whether the others follow; it flatters EPS today and pushes the cost into the back half of the decade, exactly when the bears say the revenue has to materialize. The Amazon free-cash-flow swing is the cleanest illustration of the whole trade's central tension: the operating business (AWS at 39% margins) is a cash machine, but the capex is so large that group free cash flow goes negative, which is precisely why the debt story in item 2 exists.

The debate

Steel-manning both sides of the hyperscaler capex thesis. The number itself moved this weekend: Catz's "$5 trillion, 2025-2030" and "$830 billion this year… $1.1 trillion" next year is a step up from the "$700B in 2026 → $1T in 2027" framing that anchored this letter, and it now comes from an operator, not a strategist (worth flagging: the base keeps drifting up, and the figures differ by speaker). The two camps didn't argue so much as divide the evidence: the bulls own the demand curve and the earnings; the bears own the funding structure and the soft revenue signals.

Bull, this is an earnings-driven super-cycle with demand years ahead of supply, not a debt bubble. Ives's "13-to-1" backlog and "no true equilibrium till early 2029," the "five to six dollar multiplier" across tech, and his claim the Street is lowballing earnings "by 25-30%" are the strongest bull artifacts of the weekend (The Compound and Friends, Sept 18). Lee's logic that "you're never at a top when people want [the] P/E to be low and earnings are going up" is backed by the arithmetic: earnings +25%, market +10%, so multiples contracted. Crucially, Lee meets the debt bears head-on: a debt bubble needs "return-free risk" from cheap money funding bad projects, but "data center cost of money is going up… the market is already putting a brake on everything." Even the credit skeptics concede the top-tier bonds are probably "money good," and Deep Values shows the operating businesses underneath (AWS at a 39.4% margin) are genuine cash machines (Deep Values, Sept 18).

Bear, the money got more expensive, the build is increasingly debt-financed and partly hidden, and the revenue signals are softening. The Fed is hiking into this, the 10-year sits near 5%, and hyperscaler borrowing is itself crowding out other capital; Peters calls it "the early stages of… credit crowding out," a "quantum of debt… not seen since the… railroad system" (Credit Exchange, Sept 18). Wigglesworth's structural point is the sharpest: ~"$500 billion of bonds this year," "$1.5 trillion of lease commitments… only $500 billion… on balance sheets," a $27 billion off-balance-sheet Meta JV bond, and now credit default swaps trading on the hyperscalers, and "debt-fueled bubbles… are far more debilitating" than the equity-financed dot-com bust (Simply Put, Sept 18). The revenue side is wobbling at the margin: cost per token down 50% since May, share bleeding to cheap open-weight models, "OpenAI's profitability actually got worse," and the top 1% of users cutting spend 10% month-on-month (RiskReversal Pod, Sept 18). And the pricing-power tail: Lakshman argues the frontier labs have "no moat" and LLMs will "become commodity" like storage servers did in the 1990s, leaving "astronomically high" infrastructure costs against collapsing prices (The Neon Show, Sept 18). Microsoft's 15-to-25-year depreciation stretch, on this read, is a way to make the near-term math look better than it is (Deep Values, Sept 18).

The synthesis. Both camps now agree the suppliers get paid and the cash-rich hyperscalers can service their high-grade debt; they diverge on the equity multiple, on whether the labs' revenue converts fast enough, and on which players survive. What's genuinely new this weekend is the funding structure moving to center stage: the bears' argument is no longer "the multiple is too high," it's "the leverage is rising, some of it is invisible, and the money just got dearer." The tie-breakers are unchanged and sharper: whether the "13-to-1" backlog and the earnings beats hold, whether cost-per-token and top-1% spend keep sliding, where the 10-year settles, and whether that off-balance-sheet lease pile stays quiet.

Sell signals to watch:

  • Cost per token and the ramp 1%-spend index. Cost per token is down 50% since May and the top 1% of users cut spend 10% July-to-August, the first such dip since ChatGPT launched. If both keep sliding, the capex-to-revenue conversion is breaking (RiskReversal Pod, Sept 18).
  • Widening credit-default-swap spreads on hyperscalers. They barely existed a year ago; a widening would say the credit market has stopped believing the bonds are "money good" (Simply Put, Sept 18).
  • The 10-year Treasury climbing further from 5%. With the Fed hiking and hyperscalers crowding the bond market, this is the fastest way the equity multiple unwinds (Credit Exchange, Sept 18; RiskReversal Pod, Sept 18).
  • More off-balance-sheet JV bonds like Meta's Hyperion deal. The bigger the gap between reported debt and true lease commitments, the harder the eventual reckoning (Simply Put, Sept 18).
  • More hyperscalers stretching depreciation lives. If others follow Microsoft's 15-to-25-year change, it signals the group is reaching for accounting levers to protect EPS as capex peaks (Deep Values, Sept 18).

Stocks in play

NVDA. Bull: still the name even the bears exempt: Q2 revenue "$96.22 billion" (+106%), net profit "$59.68 billion," data-center revenue "$89 billion" (+117%), a "75%" gross margin that signals "an absolute total lack of pricing pressure," and a "mid-teens multiple for calendar '27" Niles calls "ridiculously cheap"; Ives's "13-to-1" backlog says the demand runs to 2029 (Deep Values, Sept 18; The Compound and Friends, Sept 18; RiskReversal Pod, Sept 18). Bear: Niles's "a lot of companies [are] over-earning" and a higher discount rate can compress even a cheap multiple; that 75% margin is a target every customer is trying to design around. Next catalyst: the labs' revenue conversion; custom-silicon share loss (Maya 200, others). One analyst pair-trade floated: long Nvidia/Micron, short Intel/Dell (Futurum Equities, Sept 18).

AVGO. Bull: the "safe money" arms-dealer whose custom-ASIC franchise gets paid regardless of which model wins; grouped with the "core inputs to the build [that] are super underappreciated" (Futurum Equities, Sept 18). Bear: it sits inside the same capex-depreciation math a slowdown would scrutinize. Next catalyst: next hyperscaler custom-ASIC order update.

MSFT. Bull: a core "building it to rent it" landlord with the balance sheet to carry the debt; Copilot hit "30 million paid seats," and its "Maya 200" in-house chips aim to cut cost per token "by 30%" and reclaim the Nvidia margin (Deep Values, Sept 18). Bear: it just stretched server depreciation from 15 to 25 years, flattering EPS as capex peaks, and free cash flow still fell 23% to ~$19.2 billion; Catz pegs its spend near "$200, $250 billion" this year (Deep Values, Sept 18; The Geopolitics of Business, Sept 18). Next catalyst: whether peers copy the depreciation change; Azure growth against the capex line.

GOOGL. Bull: a cash-machine landlord that "owns the stack", Futurum's preferred way to play, able to "fund the compute growth… that other companies cannot afford," with the deepest balance sheet and the ad engine to absorb a write-down (Futurum Equities, Sept 18; The Compound and Friends, Sept 18). Bear: it "went cash flow negative for the first time since they went public" and is now raising debt and equity, competing with a 6%-of-GDP federal deficit for capital; Catz puts its spend at "$200 billion" this year (RiskReversal Pod, Sept 18; The Geopolitics of Business, Sept 18). Next catalyst: free-cash-flow trajectory against a rising capex line.

AMZN. Bull: AWS is the offtake engine and a genuine cash machine: a record "39.4%" operating margin, "$25 billion" of annualized AI/custom-chip revenue, and a "$220 billion" full-year spend that "prices out the competition" (Deep Values, Sept 18). Bear: the heaviest absolute capex to justify, and it just drove group free cash flow from "+$18.2 billion" to "-$7.6 billion" year on year; Catz pegs its spend at "$240 billion," the largest of the four (Deep Values, Sept 18; The Geopolitics of Business, Sept 18). Next catalyst: proof AWS demand is broad beyond the frontier labs.

META. Bull: an ad engine that insulates holders from a factory write-down, "owns the stack," and just shipped its Muse open-weight consumer model to widen the funnel; Catz puts its spend at "$140 billion," the lightest of the four (Futurum Equities, Sept 18; The Compound and Friends, Sept 18). Bear: the poster child for the off-balance-sheet build: its Hyperion JV with Blue Owl issued a "$27 billion" bond that sits "nowhere on Meta's balance sheet," the template the whole industry is copying (Simply Put, Sept 18). Next catalyst: whether ad growth keeps outrunning the capex line; how much more of the build moves off the books.

Also in play: MU / memory (the split persists: Micron trades near a "six times" forward P/E while Dell trades at "20-21 times," which Futurum flags as an impossible gap where "one of those two facts [has] to be wrong"; the bear counter from Hedgeye is that "memory is a commodity business [where] the Chinese are expanding capacity," so pricing has likely "passed the peak"; Futurum Equities, Sept 18; Hedgeye Podcasts, Sept 19). VRT (Vertiv) / power kit (named as a preferred "AI power thematic" add alongside memory names; Futurum Equities, Sept 18). CoreWeave (CRWV) / Nebius (neocloud financing strain again: a "billion dollar" CoreWeave stock offering knocked Nebius back after a "20%" rate raise; the model still depends on continuous capital-raising; Futurum Equities, Sept 18). OpenAI / Anthropic (private) (OpenAI is the "price leader… making about half as much money on a dollar of revenue," a "15% gross margin versus 30%" for peers, a "commodity margin" the market may not reward; Futurum Equities, Sept 18).

Read-throughs

  • Rates / credit, still the biggest lever on the whole book, and the story got more structural. The new development isn't the hike (that was last week) but the funding picture behind it: ~$500 billion of hyperscaler bonds this year, $1.5 trillion of lease commitments with only a third on balance sheets, credit default swaps now trading on these names, and a $1.2-trillion credit manager calling it "credit crowding out" of a scale unseen "since the… railroad system" (Simply Put, Sept 18; Credit Exchange, Sept 18). The actionable posture is unchanged from Bassman last week: own the credit and the suppliers, respect the risk to the equity multiple, but add a new to-do: read the lease footnotes before underwriting any hyperscaler's "clean" balance sheet.
  • Memory / networking, the crowded trade split into a valuation argument. The bull frame is the six-times-versus-twenty-times gap between Micron and Dell that "shouldn't exist"; the bear frame is that memory is a commodity where "the Chinese are expanding capacity" and pricing has "passed the peak of the tightest" (Futurum Equities, Sept 18; Hedgeye Podcasts, Sept 19). The single-digit multiple is the market's verdict on durability; the cleanest expression is to size memory small and pair it with the picks-and-shovels names that get paid on volume, not price.
  • Power / grid, the constraint is now political and physical, not just about generation. Texas ("a deeply red state" and the country's second-biggest data-center hub) put a "moratorium on… connecting some of those data centers to the grid," and a data-center distributor told Hedgeye that "we cannot keep accelerating at this pace… you still got to get the inspector to come out and sign off," so 30% growth rates run into "town meeting" reality (RiskReversal Pod, Sept 18; Hedgeye Podcasts, Sept 19). Bullish for the constrained-equipment complex (Vertiv and peers); the rising risk is permitting friction and a "socialist Democrat" midterm wave Niles flags as "really anti data centers."
  • The labs' pricing power, a new bear vector worth a watch. Lakshman's "no moat / LLMs become commodity like storage did" argument, plus the on-the-ground fact that cost per token is down 50% and OpenAI's margins are slipping as it cuts price to fight open-weight rivals, is the softest but potentially most important signal: if the frontier models commoditize, the offtake revenue that justifies the whole $5 trillion build gets much harder to earn (The Neon Show, Sept 18; RiskReversal Pod, Sept 18).

What changed vs last issue

Last issue (Friday, Sept 18, "The Fed hiked into the boom. Gerstner won't YOLO.") covered Thursday's Sept 17 podcasts: the Fed's hawkish surprise, Harley Bassman tying rising rates to hyperscaler borrowing ("30%, 35% competition for money"), Brad Gerstner de-risking to "medium" while insisting "no bubble," and the memory bulls buying the selloff. This edition covers Friday-evening and weekend podcasts (Sept 18-19), and the story moved from "the money got more expensive" to "how big is this really, and how is it financed?"

  • The capex number jumped again, and got an operator's name on it. Last issue floated Gerstner's "$1.5 trillion a year." New this cycle: Oracle's Safra Catz (via Sorrell) put the four biggest builders at "$5 trillion, 2025-2030," with this year already revised "from about $630 billion… to $830 billion" and next year at "$1.1 trillion. Just those four" (The Geopolitics of Business, Sept 18).
  • The debt story went from "money good" to "structurally fragile and partly hidden." Last issue, Bassman called the hyperscaler bonds solid. New this cycle: ~$500 billion of bonds this year (2x last year), $1.5 trillion of lease commitments with only $500 billion on balance sheets, a $27 billion off-balance-sheet Meta JV bond, and CDS now trading on these names, with two independent voices invoking the 1800s railroad boom and warning "debt-fueled bubbles… are far more debilitating" (Simply Put, Sept 18; Credit Exchange, Sept 18).
  • The bull rebuttal crystallized into hard numbers. Last issue's bull (Gerstner) was hedged. New this cycle: Tom Lee and Dan Ives: "13-to-1" demand, equilibrium "till early 2029," a "five to six dollar multiplier," earnings underestimated "25-30%," and Lee meeting the debt bears head-on ("data center cost of money is going up… we're not able to get to a debt bubble") (The Compound and Friends, Sept 18).
  • The depreciation debate got a hard datapoint. Abstract last issue; concrete now: Microsoft stretched server life "from 15 years to 25 years," the "mortgage refinance" that lifts EPS as capex peaks (Deep Values, Sept 18).
  • New soft signals on revenue conversion. Cost per token down 50% since May, OpenAI's profitability "actually got worse," and the top 1% of users cut spend 10% July-to-August, the first monthly dip since ChatGPT launched (RiskReversal Pod, Sept 18).
  • Still open: whether the "13-to-1" backlog and the earnings beats hold; whether cost-per-token and top-1% spend keep sliding; where the 10-year settles; whether CDS spreads on hyperscalers widen; and whether more names copy Meta's off-balance-sheet JV structure and Microsoft's depreciation stretch.

Next catalysts to watch: the path of the 10-year Treasury and any Fed speak reinforcing the December-hike signal; the next ramp-index and token-price reads as the tell on capex-to-revenue conversion; hyperscaler CDS spreads as the credit market's verdict; and the next round of hyperscaler capex guidance measured against Catz's escalated "$830 billion this year → $1.1 trillion next" path.