Newsletter · 2026-09-07 · Ashutosh Agarwal

The AI Trade Now Rests on Two Money Losing Labs, OpenAI and Anthropic - The AI Capex Tracker - Week of September 7, 2026

The AI Capex Tracker for the week of September 7, 2026: over a US Labor Day weekend, three independent skeptics converged on the same crux, that the whole AI build-out is load-bearing on two unprofitable labs, plus a new SEC no-action letter that lets banks securitize data-center debt and Morgan Stanley's 1.5 trillion dollar 2027 capex call.

The AI Capex Tracker

Week of September 7, 2026: The AI Trade Now Rests on Two Money Losing Labs, OpenAI and Anthropic

Issue: Monday, September 7, 2026. US markets were closed for Labor Day, so this is the long-weekend edition, covering podcasts published Friday evening, September 4, through Sunday, September 6, the stretch since Friday's Nvidia and Broadcom recap.

TL;DR

The long weekend belonged to the bears, and three of the sharpest ones independently landed on the same point: the entire AI build-out is now load-bearing on two unprofitable startups, OpenAI and Anthropic. Ed Zitron laid out the math: 80% of OpenAI's business revenue comes from just 1% of its customers, the two labs have signed more than $1.1 trillion in "take-or-pay" compute contracts (they pay whether they use the capacity or not), and together they account for over $174 billion of expected 2027 revenue at Microsoft, Google and Amazon, all of it contingent on the labs raising fresh money. He calls those commitments "the subprime mortgages of the AI bubble." (Better Offline, Sept 4; The Real Eisman Playbook, Sept 4)
Paul Kedrosky called Nvidia the bubble's "single point of failure" and flagged a genuinely new, under-covered development: the SEC has quietly opened the door to Wall Street packaging and reselling data-center debt the way it once did mortgages. He argues more than half of hyperscaler data-center spending is now borrowed money, sitting in off-the-books vehicles, and that a recent SEC "no-action letter," cited in two bond deals in the last six weeks, lets banks syndicate this paper with far less scrutiny. His verdict: "This time is different because we have pristine balance sheets and a wildly accelerating spending ramp, which guarantees oversupply." (RiskReversal Pod, Sept 4)
The bull case did show up, and it is a good one, but even its best spokesmen now hedge. Morgan Stanley's Dan Skelly says AI infrastructure drove 60% of second-quarter earnings growth, sees 2027 capex at $1.5 trillion (above the Street's roughly $1.2T), and notes the healthiest sign versus 2000 is that hedge funds have not piled back in. But he concedes a June note warning on "double-ordering" ahead of 40 to 50 gigawatts of data-center construction, and that "we almost always boom-bust on the capex." (The Compound and Friends, Sept 4)

What's new

Markets were shut Monday for Labor Day and there were no earnings prints over the weekend. What the window did give us was a long, quiet stretch that let the smartest skeptics talk at length, and they used it to converge, from very different starting points, on one argument. The developments below are ranked by where the dollars land hardest.

1. The single biggest development of the weekend is a convergence, not a headline: three credible, independent voices all concluded the whole trade now rests on OpenAI and Anthropic, two companies that lose tens of billions of dollars a year. Better Offline, Sept 4 with Ed Zitron (tech writer and critic); The Real Eisman Playbook, Sept 4 with Steve Eisman (the investor from The Big Short); RiskReversal Pod, Sept 4 with Paul Kedrosky (venture investor, MIT fellow).

Start with Zitron, who did the plumbing. His term of the week is "concentration risk", having too many of your eggs in too few baskets. The numbers he assembled:

Per fintech firm Ramp, 80% of OpenAI's enterprise revenue comes from just 1% of its customers, a figure that hasn't improved in three years, and many of those customers are themselves venture-funded AI startups burning cash.
OpenAI and Anthropic have together signed more than $1.1 trillion in compute commitments. These are largely "take-or-pay" deals (a contract where you pay for the capacity whether you use it or not).
OpenAI's spend equals about 70% of Microsoft's AI revenue in the fiscal year that just ended in June. UBS estimates OpenAI and Anthropic will account for 48% of Google Cloud's entire revenue next year, roughly $84-100 billion in 2027, plus about $40 billion at Amazon Web Services and at least $50 billion at Microsoft Azure. Add it up: over $174 billion of expected 2027 revenue at the three biggest clouds hangs on these two labs' ability to keep raising money.

His analogy is deliberately provocative: "OpenAI and Anthropic's massive compute commitments are the subprime mortgages of the AI bubble." The trigger he watches is timing. Most of these contracts have only a small upfront fee so far, because the capacity hasn't come online. "That's going to start happening next year and get dramatically worse month after month as capacity starts powering up and they start actually having to pay for it." The reset wall, in other words, is 2027.

Eisman, who is careful and explicitly not calling a top, arrived at the same node from the accounting. He pointed to Note 7 of Nvidia's latest 10-Q, which discloses that roughly 70% of Nvidia's revenue comes from just five customers (the hyperscalers), whose own AI revenue leans heavily on the same two labs. His summary: "If Anthropic or OpenAI ever get into trouble, the whole AI ecosystem will slow to a crawl." He was blunt about the stakes for the country, not just the stock: "There is no question in my mind that the entire U.S. economy hinges on the success of AI. The amount being spent is just so large that were it to stop, the economy would go into a recession almost immediately." He also poured cold water on a headline the business press cheered, OpenAI's claim of $1 billion in advertising annualized run-rate revenue. Eisman noted OpenAI had earlier projected $2.4 billion of ad revenue for all of 2026, so $1 billion now means it "won't be even close," against a June quarter that had $6.7 billion of revenue and $12.3 billion of costs.

"OpenAI and Anthropic's massive compute commitments are the subprime mortgages of the AI bubble." Ed Zitron, Better Offline, Sept 4

Why it matters: for months the bear case has bounced between "it's too expensive" and "it's debt-financed." This weekend it sharpened to a single, testable claim, that the whole tower is balanced on two loss-making labs. That gives a book something concrete to monitor (the labs' fundraising and their 2027 payment ramp) rather than a vague unease.

2. Paul Kedrosky called Nvidia the "single point of failure" and surfaced a real, new regulatory change that makes the debt story worse. RiskReversal Pod, Sept 4 with Paul Kedrosky, in conversation with Dan Nathan.

Kedrosky's frame is that AI capital spending, now 30% to 60% of all U.S. GDP growth and the largest single piece of business investment in the country, having "crowded out pretty much all other commercial fixed investment," sits at the intersection of the four forces that have driven every great American bubble: a genuine technology story, loose credit, government policy (the "existential race with China"), and real estate (the rise of "powered shells," speculative data-center sites with power and water but no customers or chips). "Each one of those actors thinks they're acting rationally," he said. "The intersection of all their forces isn't just additive, it's multiplicative and geometric."

The genuinely new, actionable nugget: the SEC recently issued a "no-action letter", a signal that it won't object, that lets banks package and syndicate data-center debt (bundle up the loans and sell the pieces to other investors), something they can no longer do with mortgages after 2008. Kedrosky says a law firm (Latham & Watkins) argued data-center debt is legally different from mortgage debt, the SEC effectively shrugged, and the letter was "immediately cited in two recent issues" in just the last six weeks. His point: this removes friction and pushes even more capital into the build, "which guarantees oversupply."

He tied this to Nvidia specifically, hence the episode's title. Nvidia's roughly $500 billion financing pool (with David Solomon of Goldman, plus Blackstone, BlackRock, Brookfield and KKR), in which Nvidia backstops up to 25% of GPU-compute loans, plus its acquisition of Hugging Face, is Nvidia "backward-integrating" to control the whole chain, which also makes it the node everything routes through. He also connected the dots to bond yields: with roughly 60% of U.S. government debt now bought by yield-sensitive "retail" money rather than foreign central banks, a Google bond paying 6% is a "clean, higher-yield, more secure" competitor to the 10-year Treasury at the same roughly 10-year duration as a data-center lease, so all this corporate issuance is helping push government borrowing costs up. Why it matters: it reframes Nvidia not as the safest name but as the concentration point, and it names a specific plumbing change (debt securitization) that could accelerate the cycle's late innings.

3. The most balanced and most useful bull came from Morgan Stanley, and even it now carries caveats. The Compound and Friends, Sept 4 with Dan Skelly (Head of Morgan Stanley Wealth Management's market research team), Josh Brown and Michael Batnick.

Skelly's core case is strong and worth steel-manning. AI infrastructure drove 60% of second-quarter earnings growth (28% of the headline number). His team's 2027 capex estimate is $1.5 trillion, above the Street's roughly $1.2 trillion, and he says the Street has been playing catch-up for 18 months in what his desk calls "the quarterly tradition" of numbers only going higher. Josh Brown added an under-appreciated point: once you build a data center, you've "embedded guaranteed purchases of servers and chips as far as the eye can see," because GPUs are not 20-year assets, a durable tailwind for Nvidia and Dell (which just posted a blowout).

The caveats are what made it credible. Skelly flagged his own June 1 note cautioning on semiconductor momentum, arguing the real constraint is deployment, not demand: companies are ordering chips and gear ahead of 40 to 50 gigawatts of planned construction over two to three years, which raises real "double-ordering" risk ("Is it a risk today? No. But is there a risk from here? Could be"). He conceded the dot-com analog is "perfectly reasonable" (someone held up a chart of fiber-optic name Ciena, "straight up, straight down"). And crucially, from Morgan Stanley's perch as the biggest prime broker, he reported that hedge funds have not re-grossed: the first-half AI winners have been waiting, not chasing, and money has rotated toward Mag 7, software and healthcare. He calls that rotation "the healthiest thing" separating this cycle from 2000. On the circular-financing worry, he cited a Michael Cembalest (JPMorgan) figure of $320 billion of new hyperscaler and Nvidia investment-grade debt and off-books vehicles issued in 2026, but argued spreads remain benign, buyers (including insurers) are plentiful, and the market is already discriminating: Microsoft and Johnson & Johnson are "the only two AAA-rated names left," while Oracle's credit-default-swap spreads trade wide. Why it matters: this is the bull case a book should actually pressure-test against, because it agrees with the bears on the facts and disagrees only on the timing.

4. David Rosenberg made the freshest macro-bear case on Sunday, Sept 6: the knock-on effects could be worse than the dot-com bust. Thoughtful Money, Sept 6 with David Rosenberg (Rosenberg Research).

Rosenberg's central observation is about breadth in reverse. In the late 1990s, only media and telecom moved with tech. Today, he says, only health care and staples are not correlated with the AI trade: industrials, financials, consumer discretionary and utilities all are. "Everything is converging on one trade, which is generative AI... if this thing does reverse, it's going to have much broader impacts on the stock market than even what happened when we had the tech meltdown back in the early 2000s." He stacked up the froth: a CAPE ratio around 40 (a cyclically-adjusted price-to-earnings measure, the market's valuation smoothed over 10 years); a negative equity risk premium (stocks yielding less than the "safe" real bond yield, meaning investors are treating equities as riskless); 73% of household financial assets in equities per the Fed; portfolio-manager cash near a record-low 1%. His base case for 2027 is that both earnings estimates and the multiple come down: "I would have no problem formulating that as a base case." And he tied rising rates directly to the build, since corporate AI borrowing in "long-dated maturities and multiple currencies" is now competing with the government for funds, backing up real rates and quietly strangling housing. Why it matters: it's the cleanest articulation of the "this time it's more correlated, so the fall is broader" risk, and it came from a macro strategist, not a perma-bear pundit.

5. Power stayed the physical ceiling, and the read moved north to Canada. Hub Podcasts, Sept 4 with an Alberta grid and industrial-load panel; Tech Talks Daily, Sept 5 with the CEO of EdgeCore (a data-center developer, so an operator with a book to talk); Volts, Sept 4 with Saleem Chapman (Climate Cabinet).

Given Google's record Canadian bond deal last issue, the Alberta episode is a timely read-through. Alberta runs a 12.5-gigawatt grid today and has 20 gigawatts of data centers looking to connect, including an incoming 1,000-megawatt Meta site that, on its own, rivals the 600 MW consumed by the province's seven largest industrial members combined. Alberta's answer is "BYOG," bring your own generation, a rule forcing any load over 75 MW to supply its own power so existing ratepayers aren't stuck with the bill; the province is also lifting its power-price cap from $1,000 to $3,000 per megawatt-hour to push big users to throttle back at peak. The catch a book should note: a three-year bridging period means load can arrive before the promised generation does, risking "significant pricing impacts" in the gap. EdgeCore's chief, meanwhile, framed power as "a decade-long issue" and said the company has locked $1.5 billion of fully-leased financing and is over-building floor space and liquid-cooling capacity to avoid stranded assets. And on the market-design front, Climate Cabinet's Saleem Chapman argued all three U.S. grid structures, Georgia's vertically integrated utilities, PJM's competitive market, and ERCOT's free market, are failing to fairly allocate data-center demand risk, leaving ordinary ratepayers exposed. Why it matters: the power bottleneck is now a cross-border, ratepayer-politics story, and the "who pays for the grid" fight is the slow-burn risk to how fast capex converts to revenue.

The debate

Steel-manning both sides of the hyperscaler capex thesis, now roughly $800 billion in 2026, heading toward $1.3 trillion in 2027 on Nvidia's own math, with Morgan Stanley as high as $1.5 trillion. Last issue, Nvidia's blowout print settled the near-term demand question in the bulls' favor. This weekend the argument moved one layer down, to who is actually paying, and with whose money.

Bull: the demand is real, the spenders are credible, and the smart money hasn't even chased it yet. AI infrastructure drove 60% of Q2 earnings growth; scaling laws still work, so there's no technical reason to cut spending; and building the centers locks in years of chip re-buys (The Compound and Friends, Sept 4). BetaShares strategist Tom Wickenden made the clean version: this is "the biggest capex cycle in history," the funding has shifted to the bond market but "bond investors are happy to fund this," and recent cloud results showed rising revenue and margins, "a strong indication of potentially their path to profitability." His marquee data point: Anthropic, which didn't exist five years ago, is now at a $65 billion annualized revenue run-rate, "more than any Australian company earns in revenue" (Equity Mates, Sept 6). Economist Tyler Cowen went further, arguing the bears are "asking the wrong questions" and that the labs themselves "are not very heavily financed by debt" (Prof G Markets, Sept 4). And Morgan Stanley's tell is the most reassuring: hedge funds have not re-grossed, so this isn't a crowded, everyone's-all-in top.

Bear: the money is borrowed, it's hidden, and it all rests on two firms that lose billions. The weekend's bear case was unusually coherent because three people built it independently: Zitron's concentration math ($1.1 trillion of take-or-pay commitments, 80% of OpenAI revenue from 1% of customers, a 2027 "reset wall"), Eisman's Note 7 (70% of Nvidia's revenue from five customers who lean on the same two labs), and Kedrosky's plumbing (more than half of data-center spending now externally financed, off the books, and, new this weekend, now legally securitizable). Layer on Rosenberg's warning that the trade is now correlated with almost every sector, so a reversal hits far wider than 2000 (Thoughtful Money, Sept 6); Hedgeye's point that hyperscalers are "burning cash at an extremely high rate" while price competition (Meta, OpenAI cutting prices) destroys returns (Hedgeye, Sept 4); and Macro Hive's Dominique Dwor-Frecaut arguing much of the capex is imported and the boom will "run out of steam" with "limited impact on growth" (Macro Hive, Sept 4). Even the Dividend Cafe's David Dietze, from a wealth-management seat, said free cash flow is collapsing despite rising revenue and that three unknowns, customer-acquisition cost, future utilization, and whether profitable use arrives before the assets depreciate, make honest ROI analysis "impossible" (The Dividend Cafe, Sept 4).

The synthesis. The two sides now agree on almost all the facts and disagree only on timing and consequence. Both accept the demand is real; both accept the funding has gone from cash to debt; both accept 2027 is the year the take-or-pay bills come due. The bull says the spenders are credible, the buyers of the paper are plentiful, and the money hasn't even crowded in yet. The bear says credible balance sheets are exactly how capital markets "turn a prime credit into a subprime credit," and that the concentration on two loss-making labs plus off-books, now-securitizable debt is the fault line. The most important new development is not a number but a shape: the bear thesis has become specific and monitorable, and it points at two private companies most of us can't even see the financials of.

Sell signals to watch:

OpenAI or Anthropic fundraising stumbling, or their 2027 payment ramp starting late. This is now the single load-bearing variable. Watch for a down round, a delayed raise, or any sign the labs are renegotiating their take-or-pay commitments (Better Offline, Sept 4).
A data-center bond deal repricing wider, or a securitized tranche marking down. Kedrosky's SEC-letter point means more of this paper is about to be sold to investors who don't hold the risk, so the first crack shows here, not in equities (RiskReversal Pod, Sept 4).
Oracle's credit spreads widening further. Morgan Stanley already flags Oracle's CDS as the wide man among the big spenders; a further blowout would be the market pricing the "prime-to-subprime" drift in real time (The Compound and Friends, Sept 4).
Token prices falling faster than volume grows. Kedrosky's point that the top models are converging and prices are down 50-80% year-over-year means the labs are increasingly "losing money on every unit and making it up on volume," a squeeze that would hit the labs' economics before it hits the chipmakers (RiskReversal Pod, Sept 4).
Hedge funds finally re-grossing into the first-half winners. Skelly frames the fact that they haven't as the bull's safety margin; if that flips and everyone chases, the crowded-top risk returns (The Compound and Friends, Sept 4).

Stocks in play

NVDA. Bull: still the demand referee. Eisman notes revenue grew in excess of 100% last quarter, an acceleration from 85% the quarter before, and Telltales pegs Nvidia's run-rate free cash flow at roughly $186 billion, about $50 billion ahead of Apple's (The Real Eisman Playbook, Sept 4; Telltales, Sept 6). Bear: Kedrosky's whole thesis is that Nvidia is now the bubble's "single point of failure," backstopping about 25% of a $500B GPU-financing pool, backward-integrating (Hugging Face), and sitting atop a revenue base that Note 7 shows is 70% concentrated in five customers who themselves depend on two loss-making labs; its customers "are no longer able to buy its GPUs through cash flow alone" (RiskReversal Pod, Sept 4; Better Offline, Sept 4). Next catalyst: any detail on how much of the $500B financing pool actually draws down, and the labs' next fundraising rounds.

AVGO. Bull: Broadcom generates $34 billion of free cash flow and is growing about 65% over the next twelve months, twice the growth of names trading at twice its multiple (Telltales, Sept 6). Bear: Telltales, an AI-produced markets recap, dissected why it fell on a great quarter: the guide came "$200 million light on $35 billion," inside the rounding error, but "a multiple that full removes your tolerance for a guide that isn't perfect." And Zitron notes Broadcom itself now says Anthropic and OpenAI will be its top two customers, wiring it into the same concentration risk (Telltales, Sept 6; Better Offline, Sept 4). Next catalyst: whether the fiscal-2027 AI number gets formally lifted above "$100 billion."

AMD. The read-through this window is indirect: Morgan Stanley's "deployment, not demand" framing and its double-ordering caution apply to the whole silicon supply chain (The Compound and Friends, Sept 4). Next catalyst: its second-half data-center step-up.

MSFT. Bull: the cleanest credit and the deepest cloud, one of only two AAA-rated names left in the market per Morgan Stanley, and rising cloud margins point to a path to profitability (The Compound and Friends, Sept 4; Equity Mates, Sept 6). Bear: Zitron's math puts about 70% of Microsoft's AI revenue through OpenAI, with at least $50 billion of Azure spend from the two labs due in 2027, the most concentrated exposure of any hyperscaler (Better Offline, Sept 4). Next catalyst: the first quarter reported under its new "Agents and Infra" segment, which turns the Azure number into a headline print.

GOOGL. Bull: elite credit and a strong balance sheet let it fund the build cheaply, and its bonds now yield about 6%, drawing yield-hungry buyers (RiskReversal Pod, Sept 4). Bear: UBS estimates OpenAI and Anthropic will be 48% of Google Cloud's entire revenue next year (about $84-100 billion), a staggering dependence on two customers, and its own record debt issuance is part of what's nudging up long-term rates (Better Offline, Sept 4; RiskReversal Pod, Sept 4). Next catalyst: evidence cloud revenue is converting to free cash flow, and any read on the OpenAI and Anthropic concentration.

AMZN. Bull: AWS remains a core demand engine, with an estimated $40 billion of 2027 compute spend coming from the two labs, real revenue even if concentrated (Better Offline, Sept 4). Bear: that same $40 billion is the concentration risk in another guise. Next catalyst: proof AWS AI demand is broad enough to justify the GPU commitments made last quarter.

META. Bull: Eisman relayed a favorable legal read on Meta's settlement with state attorneys general over teen mental health. Law professor Ben Zipursky called it smart, since it likely lets Meta "avoid legislation and regulation from many different states" and sets a legal "line in the sand" for future claims (The Real Eisman Playbook, Sept 4). The reported size of that settlement varies widely across sources, so treat the dollar figure as unconfirmed and weight the qualitative legal win. Bear: Meta is the poster child for hiding the spend. Kedrosky singled out its $100 billion-plus of off-balance-sheet data-center vehicles, and it has reportedly issued bonds at about 8% (RiskReversal Pod, Sept 4; Deep Values, Sept 5). Next catalyst: whether ad-revenue growth keeps outrunning the capex line, and any move to bring off-books spending onto the balance sheet.

Read-throughs

Financing and private credit: this is still where the first crack shows, and the weekend made it more urgent. The new fact is regulatory. Kedrosky's SEC "no-action letter" lets banks securitize data-center debt with far less scrutiny, already cited in two deals in six weeks, restoring the "I'll be gone, you'll be gone" incentive to originate and sell that predated 2008 (RiskReversal Pod, Sept 4). Morgan Stanley quantifies the flow at $320 billion of new IG debt and off-books vehicles in 2026, notes insurers are big new buyers, and flags Oracle's CDS as the wide man to watch; a Deep Values explainer put the scale in macro terms, with U.S. labs possibly needing $5 trillion of new debt through 2029 to fund $10 trillion of total capex by 2030 (The Compound and Friends, Sept 4; Deep Values, Sept 5). Actionable: watch compute-backed spreads and any securitized tranche's mark before it reaches a headline earnings line.
Power, thermal and grid: the bottleneck went cross-border and turned into a ratepayer fight. Alberta's 12.5 GW grid versus 20 GW of data-center interest, its BYOG rule, and a three-year gap where load could arrive before generation is the template for the "who pays for the grid" problem now spreading beyond ERCOT (Hub Podcasts, Sept 4). EdgeCore's $1.5 billion of leased financing and behind-the-meter design shows the operators are still funding aggressively (Tech Talks Daily, Sept 5), while the Climate Cabinet critique says every U.S. market design is mis-allocating the demand risk (Volts, Sept 4). Actionable: this still supports anything shortening the path to reliable power (grid gear, behind-the-meter gas, flexibility software), but the political and ratepayer backlash is now a genuine second-order risk to build speed.
Memory, networking and optics: hold last issue's read. Memory remains the tightest link in the chain, and the easy margin gains there have already played out. Nothing in a holiday-shortened window with no semiconductor prints changes that call; refresh it when fresh operator color arrives.
The rotation trade: the constructive flip side of all the bearishness. Two independent voices pointed the same way. Morgan Stanley sees money already rotating from the first-half AI winners into Mag 7, software and healthcare, and Eisman is explicitly preparing to reallocate toward "safety" subsectors, naming property-and-casualty insurance as the safe corner of financials and flagging low-volatility and high-dividend ETFs (The Compound and Friends, Sept 4; The Real Eisman Playbook, Sept 4). Actionable: the trade that rhymes with this weekend's thesis isn't just "short the fragile," it's "own the uncorrelated," and health care and staples are the two sectors Rosenberg says aren't tied to the AI trade at all.

What changed vs last issue

Last issue (Friday, Sept 4, "Nvidia's $1.3 trillion call. Broadcom beat and fell.") was an earnings recap: Nvidia's blowout settled the demand debate bullishly, Broadcom's "beat and fall" showed the market's mood had hardened, and the bear case had migrated to financing and power. This long-weekend edition moved the story one layer deeper, from "is demand real?" (answered) to "who is paying, with whose money, and what if they can't?"

The bear case became specific and converged. New this weekend: three independent voices (Zitron, Eisman, Kedrosky) landed on the same crux, that the whole complex rests on OpenAI and Anthropic. Concrete new numbers: 80% of OpenAI revenue from 1% of customers; $1.1 trillion of take-or-pay compute commitments; more than $174 billion of 2027 hyperscaler revenue contingent on the two labs; Note 7's 70% five-customer concentration at Nvidia; 48% of Google Cloud from the two labs; a 2027 "reset wall."
A genuinely new regulatory delta. The SEC's "no-action letter" allowing data-center debt to be securitized, cited in two deals in six weeks, is new information, and it's a structural accelerant the bulls didn't have to reckon with last issue (RiskReversal Pod, Sept 4).
"Beat and fall" got a cleaner mechanism. Last issue we noted Broadcom beat and fell; Telltales this weekend explained it precisely, saying the market "repriced two guidance revisions, not AI," rewarding Snowflake's raise (100x FCF, about 29% growth) and punishing Broadcom's rounding-error miss (53x FCF, about 65% growth). "The tape is trading the delta and ignoring the denominator" (Telltales, Sept 6).
The capex numbers spread further apart. Last issue anchored on Jensen's roughly $800B (2026) and $1.3T (2027). New this weekend: Morgan Stanley is at $1.5T for 2027, the Street at about $1.2T, a widening $300B disagreement that is itself a signal of how uncertain the out-year is (The Compound and Friends, Sept 4).
Power coverage moved northward. Last issue flagged no named ERCOT or nuclear color; this weekend brought a detailed Alberta grid read (12.5 GW versus 20 GW of interest, BYOG, a three-year gap) and a U.S. market-design critique touching PJM and ERCOT.
A new, constructive rotation thesis. New versus last issue: both Morgan Stanley (hedge funds not re-grossing, money into healthcare and software) and Eisman (preparing to shift to P&C insurance and low-vol safety) are pointing at the same "own the uncorrelated" move, the actionable counter-trade to the bear thesis.
Macro backdrop: Rosenberg's "worse than the dot-com bust" correlation argument is new color (CAPE around 40, a negative equity risk premium, 73% of household assets in equities); the 10-year was cited around 4.80% (Hedgeye), roughly in line with last issue; AI capex remains 30 to 60% of GDP growth (Kedrosky), consistent with prior. The August jobs signal turned soft (an ADP miss was flagged), a cross-current to the Fed's hawkish turn.
Deltas on the big numbers: the 2027 capex range widened to $1.2T (Street), $1.3T (Nvidia), $1.5T (Morgan Stanley); new IG plus off-books debt $320B in 2026 (Cembalest via Morgan Stanley); OpenAI and Anthropic compute commitments $1.1T; their 2027 hyperscaler revenue footprint more than $174B; Nvidia run-rate FCF about $186B; Broadcom FCF about $34B; Alberta grid 12.5 GW versus 20 GW of data-center interest.
Still open: whether OpenAI and Anthropic keep raising fast enough to meet the 2027 payment ramp; whether securitized data-center debt finds willing buyers or reprices; whether the memory sell-out proves the bull case or taxes the chipmakers; and whether the "beat and fall" pattern is early distribution or a healthy, crowded-trade reset.

Next catalysts to watch: Oracle earnings Thursday, Sept 10 (the first big print of the reopened calendar, and the widest-CDS name among the spenders); Apple's iPhone launch Sept 9 (John Ternus's first as CEO); the September FOMC; Marvell's Analyst Day, October 6; and, the new top of the list, any sign of stress in OpenAI and Anthropic fundraising or in securitized data-center paper.