Newsletter · · Ashutosh Agarwal

Eisman Joins the AI Bears as Oracle Earnings Become the Verdict - The AI Capex Tracker - Week of September 10, 2026

The AI Capex Tracker for the week of September 10, 2026: Steve Eisman put a marquee name on the bear case by arguing an AI bust would cause a recession, three independent hosts described the same debt financed tower resting on OpenAI and Anthropic, a new thesis emerged that AI spending could push interest rates higher rather than lower, and Oracle earnings landed as the cleanest test of the whole structure.

The AI Capex Tracker

Week of September 10, 2026: Eisman Joins the AI Bears as Oracle Earnings Become the Verdict

Issue: Thursday, September 10, 2026

(Covers podcasts published Tuesday, September 9, the day Apple showed its foldable iPhone and the day before Oracle reports. Eleven episodes cleared the filter, all dated Sept 9. Nine of the eleven are bear or macro shows, and that is the story of the day: the skeptics found a chorus. A "Big Short" investor now says an AI bust means recession; three separate hosts, working independently, drew the same picture of a trillion-dollar tower resting on two money-losing startups; and a new, awkward idea took shape, that all this spending could push interest rates up, not down, right as the Fed debates whether to hike. The power grid, memory and optics names went silent this cycle. The one thing that can settle any of it lands tonight: Oracle's earnings.)

TL;DR

  • The whole AI trade rests on two companies that lose money, and now a "Big Short" name says its collapse would cause a recession. Steve Eisman argues the Big Tech giants "used to run businesses that threw off so much cash they didn't know what to do with it," but after hundreds of billions in AI spending, "their cash flow has evaporated." The pressure point: roughly 70% of hyperscaler AI revenue comes from OpenAI and Anthropic, and Oracle's ~$600 billion order backlog is half OpenAI alone (Rebel Capitalist News, Sept 9). A separate host put hard numbers on it: OpenAI and Anthropic have signed $1.3 trillion of compute commitments, OpenAI alone is on the hook to spend ~$750 billion through 2030, while losing tens of billions a year (Better Offline, Sept 9).
  • A genuinely new worry: AI could push interest rates higher, not lower, right as the Fed weighs a hike. Economics writer Matt Klein argues the 1990s tech boom actually pushed real interest rates up, and that today's ~7% nominal growth makes locking in a 5% 30-year yield look unattractive, so yields may keep climbing (Top Traders Unplugged, Sept 9). The backdrop is hot: the market prices a ~60% chance of a 25bp Fed hike next week, one show is loudly calling for 50bp, Friday's CPI is expected at +3.4% year over year (target: 2%), and oil and diesel are surging (Broken Pie Chart, Sept 9; Wall Street Unplugged, Sept 9).
  • The verdict lands tonight: Oracle reports after the close, with options pricing an ~11% move, and it's the cleanest test of the bear thesis we'll get. Oracle's stock has fallen even as earnings estimates rose, re-rating its forward price-to-earnings multiple down to ~18.4x from ~30x in May (Broken Pie Chart, Sept 9). The only fresh bull note came from the sidelines: Apple's tiny $13 billion AI budget may out-earn the hyperscalers' combined $700 billion, one analyst argued (Schwab Network, Sept 9).

What's new

A note on the window and the ranking. This issue covers Tuesday, Sept 9, the day Apple unveiled its first foldable iPhone, and the day before Oracle's earnings. So the podcast feed skews to macro and opinion shows rather than fresh operator interviews, and it skews bearish: nine of the eleven episodes are skeptics or strategists. I've ranked by what actually moves money for a book, and flagged who is a specialist versus a commentator. The biggest development is that the bear case gained a marquee name and a unified structure. The most useful new idea is that AI spending could lift rates. And the loudest near-term event is Oracle tonight.

Quick jargon key, since several of these terms recur below. Capex is capital spending, money laid out on long-lived assets like data centers and chips. Hyperscaler means the giant cloud owners: Microsoft, Google, Amazon, Meta. Free cash flow is the cash a business generates after paying for its capex, the money truly left over. Take-or-pay is a contract where you must pay for capacity whether or not you end up using it. Circular financing is when a supplier funds its own customers so they can buy its products.

1. The biggest development: a "Big Short" investor says the AI bust would cause a recession, and the trade hangs on two money-losers. Rebel Capitalist News, Sept 9, host George Gammon walking through commentary from Steve Eisman (the investor made famous by The Big Short).

Eisman's core claim is blunt: the AI bubble, if it bursts, "will lead to a recession in the United States." His reasoning is about cash, not hype. Big Tech companies, he says, "used to run businesses that threw off so much cash, they didn't know what to do with it. But because of the hundreds of billions they are each spending on AI CapEx, their cash flow has evaporated. And in some cases, they are raising both debt and equity to fund that CapEx." Gammon adds the sharper version: in a few cases the cash flow isn't just lower, it's outright negative.

Where's the weak point? Eisman locates it precisely: "It resides with Anthropic and OpenAI because they are so central to the entire AI food chain… something like 70% of hyperscaler AI revenue comes from Anthropic and OpenAI." Gammon's mental image is an inverted pyramid balanced on a single Jenga block, and that block is two companies that don't yet make money. He layers on the specifics: Oracle carries a backlog of about $600 billion, and roughly 50% of it is OpenAI alone. OpenAI also accounts for over 20% of CoreWeave's contracted order book. So if that future demand gets marked down, say from $600 billion to $450 billion, "what happens to their share price? Because they've still borrowed all the money to provide that, and they've got to pay that back, but now there's no revenue to pay it back."

The freshest, most concrete detail was a run of OpenAI departures Gammon reads as a warning light: chief revenue officer Denise Dresser resigned in August after only eight months; COO Brad Lightcap left in August after eight years; the head of robotics and consumer hardware left in March over governance concerns tied to a Pentagon agreement; and the CMO and a science-research lead both departed in April. His point: senior people with stock options usually stay until an IPO cashes them out, so a cluster of exits right before a rumored listing is a tell. He also flagged the financials the Wall Street Journal reported: OpenAI's June-quarter revenue hit $6.7 billion, up only ~18% from the March quarter, while costs reached $12.3 billion, up $3 billion, meaning that in three months revenue grew $1 billion but costs grew three times that. Anthropic, by contrast, did $11 billion+ in the quarter, up over 100%, but is still deeply unprofitable. Why it matters: this is the bear case moving from "expensive stocks" to "the plumbing itself is fragile," and it names the two pipes that carry the water.

2. Three hosts, one picture: a trillion-dollar tower financed against IOUs from companies that don't earn. Better Offline, Sept 9, Ed Zitron (tech critic); corroborated independently by Commodity Culture, Sept 9, Jay Martin (investor).

What made Tuesday different from a single grumpy pundit is that separate people, on separate shows, described the identical machine. Zitron's version, sparked by a Financial Times report that bankers are seeking investment-grade credit ratings for OpenAI and Anthropic ahead of their IPOs: the two labs have over $1.3 trillion in compute commitments, OpenAI alone "is expected to spend $750 billion through 2030," and Anthropic has "made $517 billion in compute commitments." To put that in scale, he noted it "dwarfs the operating expenses of Meta at around $141 billion a year, and Microsoft at about $175 billion a year." Worse, he says, 80% of the two labs' enterprise revenue comes from the top 1% of their customers, mostly other unprofitable AI startups subsidizing their own users' token bills with venture money. And the commitments are take-or-pay: whether or not the demand shows up, "Anthropic and OpenAI will have to pay Amazon, Google, Microsoft, Cerebras, CoreWeave" once the capacity comes online, "if they can afford it." His warning: a rating just above junk means "so much debt with onerously high interest," and unless the labs "magically find ways to become profitable," he doesn't "see how they service the debt they're gonna need." (For a sense of scale on how long money-losing infrastructure can take to turn: he noted Amazon Web Services burned roughly $29.7 billion of total capex between 2003 and turning profitable in 2015.)

Jay Martin arrived at the same place from the commodity world, and drew the cleanest analogy of the day, to 2008. Future compute orders owed "mainly from OpenAI and Anthropic to mainly Oracle and Microsoft and also Amazon and Google is now over $2 trillion." The tech giants building the data centers spent "over $750 billion last year… more than all the income of those tech companies," so "every new shovel in the ground is debt-financed", borrowed against those $2 trillion of IOUs. His housing-crash parallel: in 2006 the "2 and 28" mortgage worked only as long as home prices rose 15% a year, letting buyers refinance; when appreciation slowed to 8%, the refinancing stopped and defaults began. "The AI companies can keep going back to the market for the capital they need… as long as their valuations keep climbing." He was careful to separate the technology from the finance, "I'm super bullish on AI… it wasn't the housing market that crashed in 2008, it was the financial structure built on top of it", and noted the systemic stakes: "40% of the S&P 500 is 10 companies, nine of which are incredibly dependent on the AI train," and AI is "the biggest driver of U.S. GDP, and it's not even close. Number two is gambling." Why it matters: when independent voices converge on the same failure mechanism, valuations must rise forever or the debt tower cracks, that's the shape of a consensus short forming.

3. The new idea worth chewing on: AI capex might push rates up, not down. Top Traders Unplugged, Sept 9, Matt Klein (economics writer, The Overshoot), the most analytically careful voice of the day.

Most people assume a tech boom is disinflationary and rate-friendly. Klein pushed hard the other way. Going back to the 1990s, the last "genuine, unprecedented real investment boom", he argued real interest rates actually rose "from levels that are already elevated compared to where we are now," because a wave of attractive new investment opportunities lifts the demand for capital and therefore its price. And the disinflation people remember from that era, he says, mostly came from outside forces (emerging-market crises, a strong dollar, falling commodity prices), not from the tech itself. His punchline for today: "even if you were to naively say it's the same thing, the implication that interest rates would go down, I think, is not correct… it's very plausible that if that were to happen, interest rates would go up." His simple test: with nominal growth running "like 7% in dollar terms," "why would you want to lock in a fixed nominal return of 5% for 30 years?"

Klein also delivered the single most important nuance against the doom case: most of the incremental data-center spend flows abroad, to Taiwan and Korea, rather than into U.S. production. So "the idea that the U.S. economy is entirely dependent on AI growth, I think, is a misreading", if the build slows, the direct hit to U.S. output is smaller than the bears assume (though he conceded the equity-market damage would still be large). Why it matters: it reframes the biggest risk to the AI trade as rates, not demand, and rates are exactly what's in play with the Fed next week.

4. The macro pincer: a Fed that may hike into hot inflation, and a bond market nobody wants to fund. Wall Street Unplugged, Sept 9, Frank Curzio (newsletter publisher); The Peter Schiff Show, Sept 9, Peter Schiff.

Curzio's title said it: "The Fed needs to hike rates by 50 basis points." His case is that inflation is re-accelerating, Friday's CPI is expected at +3.4% year over year (core +2.5%) against a 2% target, with WTI oil up ~25% in a month, airline fares up 20–25%, and diesel closing on $6 a gallon, up ~60% from a year ago and 15% in the past month alone. Because diesel moves nearly every physical good in the country, he sees higher prices coming "across the board." He also gave the memory-cost read that matters for hardware: prices on "any technology device that requires memory… are going up at least 30% across the board in 2027, likely another 30% in 2028." His bottom line for a book: if the Fed does nothing next week, "you just really need to sell almost everything and go into cash," because nothing stops the 10-year yield from climbing.

Schiff supplied the mechanism underneath. The hyperscalers, "Meta and Google and Amazon… had all this cash, and they were parking it in U.S. Treasuries." Now "they blew through their cash… and now they're borrowing hundreds of billions of dollars from the same savings pool that the U.S. government is dipping into." And because these companies "believe they can generate enormous returns," they "don't care if they have to pay 9%, 10%, 11%." That competition for capital, he argues, forces yields up, with the government now paying over $1.2 trillion a year in interest on the debt and heading toward $2 trillion. Both men flagged the same two data points due imminently: PPI today and CPI Friday, either of which "could easily come out much hotter than expected," and if so, "serious carnage in the bond market and then in the stock market." Why it matters: this is the transmission belt from "AI borrows a lot" to "everyone's discount rate rises", the risk Klein described, told from the credit side.

5. Apple crashed the narrative, with a $13 billion budget against the giants' $700 billion. Schwab Network, Sept 9, Dave Nicholson (Chief Research Officer, Futurum).

On the day Apple showed a foldable iPhone, Nicholson argued the hardware was a sideshow and the real story was the iOS 27 software update and Apple's "personal context" idea, an assistant that already knows your calendar, payment methods and habits. His investment framing is the one to keep, because it's the closest thing to a bull argument in the whole feed and it cuts against the spenders: Apple's AI capex is about $13 billion, versus roughly $700 billion combined for Amazon, Microsoft, Alphabet and Meta. "I would say that the investment of $13 billion is going to have much higher returns than the $700 billion will across a similar time horizon," he said, validated, he thinks, in two to three years. Apple spends only about $1 billion a year renting a custom Gemini model from Google, on the logic that "good enough always will beat perfect." Why it matters: it's a live challenge to the "more capex equals more value" reflex, the suggestion that the cheapest AI budget in Big Tech might earn the best return, precisely because Apple already owns the users. Caveat for a book: the features start English-only and won't launch in the EU or China, which delays the payoff.

The debate

Steel-manning both sides of the hyperscaler capex thesis, still running near $700 billion in 2026 and heading toward $1 trillion-plus in 2027–2028. This cycle didn't move the levels; it moved the mood. The bears gained a marquee name (Eisman), a shared blueprint (the two-money-loser tower), and a fresh macro weapon (rates might rise). The bulls barely showed up, and when they did, it was from the sidelines.

Bull, the returns are real, the financing is normal, and the U.S. isn't as exposed as the doom case claims. The strongest bull voice came from a portfolio manager on Money Life, who argued the hyperscalers' return on invested capital is "extraordinary" with equity returns "in the high 20s," so the level of rates "kind of doesn't matter, they'll issue bonds where the market clears, and it's just not material to their operations." He dismissed the short thesis directly: "we view circular finance, as it's called by the short sellers, as actually… healthy vendor finance, which you've had for decades in the U.S.," and pointed to Nvidia's 1990s investment in Marvell (bought around $90, now ~$230) as proof these are "major companies with great business models," not money thrown at frauds (Money Life with Chuck Jaffe, Sept 9). Matt Klein added the macro cushion: because most data-center spend flows to Taiwan and Korea, a slowdown would hit U.S. production less than the bears fear, even if equities suffer (Top Traders Unplugged, Sept 9). And Apple's $13B-vs-$700B pitch says at least one giant thinks the smart move is to spend less (Schwab Network, Sept 9).

Bear, the cash is gone, the tower is debt-financed against two money-losers, and rates are turning against it. Eisman: the giants' cash flow "has evaporated," and the food chain hangs on OpenAI and Anthropic, which supply ~70% of hyperscaler AI revenue and don't earn a profit (Rebel Capitalist News, Sept 9). Zitron and Martin: $1.3–$2 trillion of take-or-pay compute commitments, financed with debt borrowed against IOUs from companies losing tens of billions a year, a structure that only holds "as long as their valuations keep climbing," exactly like 2008 housing (Better Offline, Sept 9; Commodity Culture, Sept 9). Dillian: hundreds of billions in capex but GDP of only 1.5% means "the rest of the economy is actually in a mild recession," and history says "every shortage is followed by a glut", the same overbuild that bankrupted the fiber-optic firms after 2000 (Full Signal, Sept 9). And the macro just turned hostile: rates may go up (Klein), the Fed may hike into +3.4% CPI, and hyperscaler borrowing is now competing with the U.S. Treasury for scarce savings (Wall Street Unplugged, Sept 9; The Peter Schiff Show, Sept 9).

The synthesis. Notice that both sides now agree on the facts and fight over the label, exactly as they did last issue on free cash flow, only the argument has widened. Everyone agrees AI capex is ~$700 billion, that it's carrying U.S. growth (bulls call it strength, bears call it dangerous concentration), and that OpenAI and Anthropic sit at the center. The bull says the returns justify the borrowing and the financing is ordinary vendor credit; the bear says it's a valuation-dependent debt machine feeding two loss-makers, into a rising-rate wind. The tell that would break the tie isn't a podcast, it's a print. Oracle, tonight.

Sell signals to watch:

  • Oracle's reaction tonight. With ~50% of its $600B backlog tied to OpenAI, Oracle is the purest read on whether the market still trusts the two-money-loser tower. Options are pricing an ~11% move on the ~$162.50 strike (implied volatility ~132%); a sharp sell-off on an in-line or soft number would say the backlog is being discounted (Broken Pie Chart, Sept 9).
  • A hot CPI on Friday that markets actually respect. +3.4% headline is already the expectation; a hotter core would raise hike odds and lift the discount rate on every long-duration AI bet (Wall Street Unplugged, Sept 9).
  • The 10-year yield breaking higher despite Treasury buybacks. If the government's new long-end buying can't cap yields, that's the "rates go up" thesis confirming, and the direct hit to AI multiples (Full Signal, Sept 9; The Peter Schiff Show, Sept 9).
  • Any sign OpenAI's or Anthropic's fundraising slows. The tower needs continuous capital raises at rising valuations; a down round, a delayed IPO, or more senior departures would crack the base (Rebel Capitalist News, Sept 9).
  • Momentum rolling over first. As one host put it, the momentum factor, packed with Micron, Nvidia, Broadcom, Google, AMD, is "probably the first domino to fall" if a downturn starts (Broken Pie Chart, Sept 9).

Stocks in play

NVDA. Bull: the referee everyone still routes through; a PM cited Nvidia's own investing track record, its 1990s Marvell stake went from ~$90 to ~$230, as proof its circular-financing bets are in "major companies with great business models," not frauds (Money Life, Sept 9). Bear: it sits at the center of the "circular financing" critique, Eisman flagged Nvidia's investments in Perplexity and Poolside as examples of Nvidia "either lending money or making equity investments" that customers then spend back on Nvidia chips, a loop that only looks healthy while the boom lasts (Rebel Capitalist News, Sept 9). Next catalyst: Oracle's print tonight as a demand read-through, then any Rubin roadmap detail.

AVGO. Bull: a bull PM called Broadcom cheap outright, "trading at 15 times FY28 [earnings]", and lumped it in with the high-return AI names he'd own (Money Life, Sept 9). Bear: it's named as one of the two chip suppliers (with Nvidia) whose revenue depends on the same debt-funded OpenAI/Anthropic spending, "TPUs from Broadcom, who can only continue to make hundreds of billions… as long as they can provide justifications for lenders to keep issuing debt" (Better Offline, Sept 9). Next catalyst: any Amazon custom-silicon program (an open door flagged last issue), and Marvell's Analyst Day, Oct 6, as a custom-ASIC read.

AMD. Quiet on the podcasts this cycle, no fresh MI450X/Helios catalyst in the Sept 9 batch. The only mention: AMD sits among the momentum-factor leaders one host warned would fall first in a downturn (Broken Pie Chart, Sept 9). Next catalyst: its second-half data-center step-up and any MI450X deployment color.

MSFT. Bull: still judged the most durable spender; Zitron's own numbers underline the gap in scale between Microsoft's ~$175 billion of annual operating expense and the labs' commitments, i.e. Microsoft can fund its build from a real business (Better Offline, Sept 9). Bear: it's one of the take-or-pay counterparties owed by OpenAI, so a stumble at the labs flows straight back to its backlog; and Schiff lists it among the hyperscalers now borrowing rather than lending to fund the build (Better Offline, Sept 9; The Peter Schiff Show, Sept 9). Next catalyst: Oracle's read-through tonight; next capex commentary.

GOOGL. Bull: it's quietly collecting AI rents from a rival, Apple pays it ~$1 billion a year to rent a custom Gemini model, a high-margin annuity that doesn't require Google to win the consumer race (Schwab Network, Sept 9). Bear: it carries last issue's baggage, the first-ever negative free cash flow, and this cycle it's named among the giants "borrowing hundreds of billions" and among the take-or-pay providers exposed if the labs can't pay (The Peter Schiff Show, Sept 9; Better Offline, Sept 9). Next catalyst: whether next-quarter free cash flow stabilizes.

AMZN. Bull: the reference point for "money-losing infrastructure eventually pays", Zitron noted AWS burned ~$29.7 billion of capex from 2003 before turning profitable in 2015, the exact template bulls invoke (Better Offline, Sept 9). Bear: it's both a take-or-pay provider owed by the labs and, per Schiff, one of the hyperscalers that "blew through their cash" and is now borrowing to build (Better Offline, Sept 9; The Peter Schiff Show, Sept 9). Next catalyst: proof AWS demand is broad enough to justify the spend.

META. Bull: the market clearly still likes it, Meta rose ~6.5% on Sept 9, the standout gainer among the mega-caps that day (Know Your Risk, Sept 9). Bear: its ~$141 billion annual operating expense is dwarfed by the compute commitments piling up across the ecosystem it's part of, and Schiff names it among the giants now borrowing to fund the build (Better Offline, Sept 9; The Peter Schiff Show, Sept 9). Next catalyst: whether ad growth keeps outrunning the capex line.

Also in play: ORCL (tonight's print is the whole ballgame, $600B backlog, ~50% OpenAI, forward P/E re-rated to ~18.4x from ~30x in May, options implying an ~11% move). AAPL (foldable iPhone at ~$2,000+, iOS 27 "personal context," a $13B AI budget the bull case says will out-earn the giants' $700B). BE (Bloom Energy confirmed as an addition to the S&P index in the Sept 21 rebalance, the behind-the-meter power name). MU (memory prices called to rise ~30% in 2027 and again in 2028, a cost headwind for device makers but a pricing tailwind for the memory sellers). (Broken Pie Chart, Sept 9; Schwab Network, Sept 9; Wall Street Unplugged, Sept 9)

Read-throughs

  • Financing / market structure, this is where the action moved, and it points one way: up on rates. The dominant read-through this cycle isn't a stock, it's the cost of money. Schiff's mechanism is the one to hold: the hyperscalers used to be lenders to the government (parking cash in Treasuries); now they've "blew through their cash" and are borrowers competing for the same savings pool, willing to pay "9%, 10%, 11%" because they believe in the returns, which pushes yields higher for everyone (The Peter Schiff Show, Sept 9). Klein's academic version says the same: real rates rose in the 1990s boom and likely would again (Top Traders Unplugged, Sept 9). The Treasury is fighting it, Dillian noted the government began buying ~$4 billion a day of long-dated bonds on Sept 9, a yield-curve-flattening operation, but "the market is daring him to increase the size," and yields have drifted back up (Full Signal, Sept 9). Actionable: the AI trade's near-term fate runs through the 10-year, PPI today, CPI Friday and the FOMC next week, hedge duration risk, not just semiconductor risk.
  • Memory / networking, no operator episodes, but a fresh price signal. No dedicated Micron/SK Hynix/Marvell/Astera/Coherent interviews surfaced in the Sept 9 window. The one new, quotable data point: memory-heavy devices face price increases of "at least 30% across the board in 2027, likely another 30% in 2028," and Klein separately flagged memory-chip inflation already filtering into consumer prices (Apple raising MacBook/iPad prices ~20%) (Wall Street Unplugged, Sept 9; Top Traders Unplugged, Sept 9). Actionable: the HBM-tightness thesis has no new operator fuel, but the pricing-power story for memory sellers (Micron) just got a second-hand confirmation from the device-cost side, a tailwind for the seller, a margin drag for the box makers.
  • Power / thermal / utilities / nuclear, silent this cycle. Zero power, grid, ERCOT, Vertiv, Eaton, Bloom-operator, or nuclear episodes cleared the filter for Sept 9, a real gap versus last issue, which was thick with ratepayer-tariff and behind-the-meter detail. The only power-adjacent item is a market-structure one: Bloom Energy is confirmed as an S&P index addition in the Sept 21 rebalance, which mechanically adds index buying (Broken Pie Chart, Sept 9). Actionable: nothing new to trade on the power thesis today; last issue's framing (on-site generation and equipment names as the picks-and-shovels closest to the constraint) still stands until fresh operator color arrives.
  • The "is AI holding up the economy?" debate, the numbers are converging. Three shows sized AI's macro footprint and landed close together: raw AI capex is ~1.7% of GDP (Welsh), while capex spending is "3%, 3.5% of the entire economy this year" (Know Your Risk), and it accounts for "about half of the 2% GDP growth projected for 2026" (Gammon), the same order of magnitude as last issue's Deutsche Bank "single-handedly preventing recession" claim (Money Life, Sept 9; Know Your Risk, Sept 9; Rebel Capitalist News, Sept 9). Actionable: whichever way you lean, the read is that AI capex and the broad economy are now the same trade, a pullback in the build would show up in GDP, jobs and, per Klein, possibly rates.

What changed vs last issue

Last issue (Wednesday, Sept 9, "Qualcomm lands Amazon. Nvidia's grip cracks again") covered the Sept 8 batch: the Qualcomm-Amazon custom-chip deal, Anthropic's clean $25M/GW-vs-$50M/GW unit economics, Google's first negative free cash flow, and a credit picture that was quietly calming into Oracle. This edition covers Sept 9, and the tone flipped hard toward the bears, with the levels unchanged but the mood and the macro darker.

  • The bear case gained a marquee name and a unified structure. Last issue the skeptics were a scattered set of individual voices. New this cycle: Steve Eisman of "The Big Short" explicitly frames an AI bust as recession-causing, and three independent hosts (Eisman via Gammon, Zitron, Jay Martin) described the same machine, a debt-financed tower resting on OpenAI and Anthropic. Nine of eleven fresh episodes were bear/macro (Rebel Capitalist News, Sept 9).
  • The dependency thesis got hard numbers. Last issue: Anthropic's economics were the bull's rebuttal. New this cycle, the same two labs are the bear's target: $1.3 trillion of compute commitments (OpenAI ~$750B through 2030; Anthropic $517B), >$2 trillion of future compute obligations, take-or-pay terms, and a concrete OpenAI staff exodus (CRO and COO both out in August) (Better Offline, Sept 9; Rebel Capitalist News, Sept 9).
  • A brand-new macro thesis: AI may push rates HIGHER. Last issue's rate worry was about the discount rate on future AI payoffs. New this cycle, Matt Klein argues the level of rates could rise, the 1990s boom lifted real rates, and ~7% nominal growth makes 5% yields look cheap (Top Traders Unplugged, Sept 9).
  • The Fed backdrop stayed hot, and got louder. Hike odds held near ~60% for 25bp, unchanged from last issue, but now one show is calling for 50bp, Friday's CPI is pegged at +3.4%, and oil (WTI +25% in a month) and diesel (~$6/gal, +60% YoY) are stoking the inflation case (Broken Pie Chart, Sept 9; Wall Street Unplugged, Sept 9).
  • New market-structure wrinkle: the Treasury started buying its own long bonds. The government began ~$4 billion/day of long-end buybacks on Sept 9 to cap yields, but the market "is daring him to increase the size," a fresh tell to watch (Full Signal, Sept 9).
  • The credit read reversed again, from "calming" back to "competition." Last issue: hyperscaler spreads were calming, a partial re-risking. New this cycle: Schiff reframes hyperscalers as borrowers competing with the Treasury for capital, willing to pay 9–11%, a re-tightening lens, though still opinion rather than a spread print (The Peter Schiff Show, Sept 9).
  • Apple entered the frame as a bull-adjacent contrast. New this cycle: the $13B (Apple) vs $700B (the four giants) capex comparison, with the claim Apple's tiny budget out-earns the giants, a live challenge to "more capex = more value" (Schwab Network, Sept 9).
  • Quiet vs last issue: all of power/grid/nuclear (a total blackout this cycle after a rich last issue); memory/optics operators; and the Qualcomm-Amazon deal, which wasn't re-litigated. AMD quiet again.
  • Still open: whether Oracle's print tonight validates or discounts the $600B backlog; whether OpenAI/Anthropic can keep raising at rising valuations; whether the 10-year breaks higher through the Treasury's buybacks; and the unresolved core question, is this offensive, ROI-positive spending, or a valuation-dependent debt tower feeding two money-losers.

Next catalysts to watch: PPI today (Thu, Sept 10) and Oracle earnings after today's close, the reopened calendar's marquee print, options implying an ~11% move and the cleanest test of the OpenAI-backlog thesis; CPI Friday, Sept 11 (expected +3.4%); the September FOMC (~Sept 16–17) and whether a hike (25 or 50bp) materializes; the S&P index rebalance Sept 21 (Bloom Energy added); Anthropic's and OpenAI's IPO filings whenever they drop; and Marvell's Analyst Day, Oct 6.