Newsletter · · Ashutosh Agarwal

Broadcom Borrows 60 Billion as the Bears Find a Fed Voice - The AI Capex Tracker - Week of August 21, 2026

AI capex and semiconductors newsletter for the week of August 21, 2026. Broadcom moved to raise more than $60 billion of debt to finance AI data centers, extending Nvidia's borrow-to-build model to a second chip giant, while former New York Fed president Bill Dudley put a valuation scorecard and a $2 trillion revenue gap on the bubble question and the private AI labs delivered the cycle's first below-whisper run-rates.

The AI Capex Tracker

Week of August 21, 2026: Broadcom Borrows 60 Billion as the Bears Find a Fed Voice

Issue: Friday, August 21, 2026

TL;DR

  • Broadcom is trying to raise more than $60 billion of debt to help finance AI data centers, the clearest sign yet that Nvidia's "borrow to build" playbook is now the whole industry's. The deal would fund compute for Anthropic and others, and could reach as much as $100 billion in total: roughly $30 billion of riskier "junior" money from Apollo and Blackstone, and a $60–70 billion "senior secured" slug that Broadcom is shopping to banks. It extends the June Broadcom, Blackstone and Apollo pact to build 20 gigawatts of compute (for scale, an earlier $35 billion deal bought just one gigawatt). (Bloomberg Businessweek, Aug 20)
  • A former Federal Reserve official put hard numbers on the bubble question, and a $2 trillion revenue hole. Bill Dudley, ex-president of the New York Fed, argues stock valuations are stretched to near dot-com extremes (the Shiller CAPE ratio at 41 versus a 44 peak in December 1999 and a long-run average near 17), and that the AI giants will struggle to generate the roughly $2 trillion of annual revenue needed to justify a ~$5 trillion investment base. His key insight: it's the change in spending that matters, so even a 2027 that's still huge but growing more slowly could crack the earnings story. (Bloomberg Talks, Aug 20; The Financial Exchange Show, Aug 20)
  • Meanwhile the chip supply chain keeps minting deals (Google's Marvell partnership only triggers its stock warrants if Google spends $120 billion on Marvell products over six years) but the private AI labs gave the first "below expectations" print of the cycle. New leaked run-rates put Anthropic near $65 billion and OpenAI near $40 billion, impressive, but under the ~$80 billion whisper for Anthropic, and that, plus a fresh reminder of ~$3 trillion in off-balance-sheet commitments, triggered a sharp one-day sell-off in AI infrastructure names. (The AI Investor Podcast, Aug 20)

What's new

A note on the window: this is the Friday issue, sweeping roughly the last 24 hours (Aug 20). Yesterday we mapped the plumbing of Nvidia's $500 billion financing plan and the moment Google's own cash ran dry. This cycle the story widened in two directions at once: the financing model spread to a second chip giant (Broadcom), and the bear case picked up its most credentialed voice yet (a former Fed president). One caveat, same as yesterday: this window was heavy on market-desk shows, macro commentators and value investors, with almost no operator commentary, so I've flagged who's an insider and who's an opinion. Ranked by where the dollars land hardest for a book.

1. Broadcom wants to borrow $60 billion-plus to finance AI data centers, the "borrow to build" model is no longer a Nvidia exclusive.

Bloomberg Businessweek, Aug 20: Ryan Gould, Bloomberg News deals reporter (a reporter relaying his own scoop, not an analyst pushing a view).

The news crossed just after the closing bell: Broadcom is in talks with lenders to raise more than $60 billion in debt for an AI-chip financing deal that would benefit Anthropic and other companies. Gould's reporting frames it as an extension of the partnership Broadcom struck with Blackstone and Apollo back in June to provide roughly 20 gigawatts of compute globally. Put the pieces together and Broadcom could line up as much as $100 billion in debt: about $30 billion from Apollo and Blackstone as the junior (first-loss) piece, and a $60–70 billion senior secured slug that Broadcom itself is backstopping and shopping out to banks.

Why it matters, in one number: the earlier Blackstone and Apollo deal was $35 billion for just one gigawatt of compute. This one targets twenty times the capacity. Gould's read is that this is "the tip of the iceberg," and he name-checked Broadcom's existing accord with Apple as another building block. He also could not avoid the word everyone is now circling:

"Do we start to talk about the circular financing thing? Yes. I don't know, perhaps. I don't know. But here we are."

The bull framing: Broadcom's CEO Hock Tan is famous for "very, very intricate financial engineering," and Broadcom has become an industry leader in this corner of chip finance, so if anyone can pull off a raise this size efficiently, it's them, and Bloomberg Intelligence's Deep Singh notes the underlying demand and backlogs still look real. The bear framing is the same one that hangs over Nvidia's plan: a chip supplier is now arranging the debt that lets its customers buy its chips, and as Gould put it, this is "two sides of the same coin" with the Nvidia deal, "when you start to innovate or disrupt too much, is it a sign that maybe something's not quite right?" For a book, this is the single most actionable new fact of the cycle: the vendor-financing wave that started with Nvidia has now reached Broadcom, and it is being funded increasingly through private credit rather than the companies' own cash.

2. A former New York Fed president called the bubble, with a valuation scorecard and a $2 trillion revenue gap.

Bloomberg Talks, Aug 20 and The Financial Exchange Show, Aug 20: Bill Dudley, former president of the Federal Reserve Bank of New York, now writing for Bloomberg Opinion (a policy heavyweight, though an economist rather than a company operator).

Dudley laid out the most rigorous bear scorecard of the cycle. On valuations: the Shiller CAPE ratio (a smoothed price-to-earnings measure) sits at 41, near the all-time high of 44 in December 1999 and more than double the 25-to-30-year average of about 17. The extra return you get for holding stocks over inflation-protected Treasuries is just 1.1%, less than half the average since 2010. And the "Buffett indicator" (total US market value versus GDP) is around 240%, versus the 100% level Warren Buffett once called risky.

His most important argument is subtle and worth internalizing: it's not the level of AI investment that matters for markets, it's the change in it.

"CapEx in [20]27 will be higher than [20]26, but the increase in CapEx in 27 will be smaller than the increase in 26."

When the growth of the boom slows, he argues, profit-growth expectations at the suppliers come down and margins get squeezed, "just like you had a really great story on the way up, expanding profit margins, higher volume; on the way down, you'll have slowing volume growth and more constrained profit margins." Layer on the demand question (whether the hyperscalers can actually generate the ~$2 trillion of annual revenue needed to justify a projected ~$5 trillion AI capital base) and you get his call that this rhymes with the railroad, telecom and dot-com build-outs: transformative technology and a boom-bust at the same time. He also flagged the financing as "getting pretty incestuous" (suppliers lending to hyperscalers, private credit filling gaps, "greater opacity"), and even referenced the Broadcom deal crossing the wire live during his interview. His timing view: with 7 to 8 firms all-in and probably only 2 to 3 long-run survivors, "getting from the 7 or 8 to the 2 or 3, it's going to be pretty painful."

3. The chip supply chain kept signing marquee deals, but the private AI labs delivered the first "below expectations" data point.

The AI Investor Podcast, Aug 20: hosts of a retail-facing AI investing show (committed bulls; treat as opinion, not operator insight).

Fresh detail on the Google and Marvell deal our last issue introduced: Google's warrants to buy Marvell stock only vest if Google spends $120 billion on Marvell's products over the next six years, and the hosts cite estimates that Google could be spending $20–25 billion a year on Marvell by 2028, more than twice Marvell's entire current revenue (roughly $11 billion this year). They also relayed Bank of America's new map of the custom-chip ("TPU") market: Broadcom's addressable market at $250–350 billion, MediaTek at $100–150 billion, and Marvell at $50–100 billion, meaning Broadcom still keeps more than 60% of the pie even as Google multi-sources. Their framing of the whole space: AI networking revenue could grow from about $10 billion in 2023 to $245 billion by the end of the decade. Notably, Broadcom shares fell about 10% in the two days around the Marvell news, which the hosts think is the market misreading a stock that is now "amongst most hated and potentially most misunderstood."

But the same episode carried the cycle's first genuine cold-water moment. New leaked run-rates for the two flagship AI labs (Anthropic near $65 billion, OpenAI near $40 billion) are enormous, but they came in below the whisper numbers (some had Anthropic closer to $80 billion). Because those two loss-making labs are on the hook for so much of the buildout, a below-whisper print rattles the demand story, and it collided with a widely-read reminder of the ~$3 trillion of off-balance-sheet AI commitments to trigger a sharp Tuesday sell-off, "most momentum companies were down 10% in a single day." The read: the demand side, so far unshakeable, just showed its first hairline crack.

4. A market strategist warned the real risk isn't the ROI, it's that the capital markets could simply "close" before the trillion is raised.

Excess Returns, Aug 20: Andy Constan, macro and rates strategist (a markets professional known for grading every Treasury auction; an analyst view, not an operator).

Constan's argument reframes the whole debate. Forget, for a moment, whether AI's long-run return on investment is good, nobody knows yet. The near-term risk is funding. For decades, US companies have quietly retired about $1 trillion of stock a year (buybacks minus new issuance), a "passive bid" that helped push prices up. He expects that to flip to roughly zero or even net new issuance next year, just as corporate borrowing explodes, "a double whammy for bonds." His central worry, from a career watching capital markets:

"After a huge burst, the capital markets close. They just close… So the question is, do they close before the trillion is through the door?"

If markets seize up at, say, $500 billion raised with another $500 billion still needed, "this CapEx circular finance machine just pencils down." On Nvidia's $500 billion plan specifically, he made a sharp technical point: Nvidia used to do these deals directly (backstopping one data center at a time); the new structure is the first indirect one, with collateral pooled into a trust, tranched and sold to the likes of BlackRock and Apollo. His read of that shift: "Nvidia is out of capacity to do direct deals… but still very, very in need of supporting its customer," a "slight negative" heading into Nvidia's own quarter next week.

5. The bear chorus got louder and more specific: Gromen's five-boom history, Burry's "shades of Enron," and Faber's "devastating."

Monetary Matters, Aug 20 (Luke Gromen, macro strategist), The Accounting Podcast, Aug 19 (accountants Blake Oliver and David Leary relaying Michael Burry), and WTFinance, Aug 19 (Marc Faber).

Gromen pointed to a Bloomberg chart of the five previous US capital-spending booms going back to the 1840s canal era (several states literally defaulted after the canal boom) and noted every one ended in a bust: "Every single one of those prior 5 booms ended in a bust, and this one will end in a bust too." His tell that the top isn't in yet: the machine keeps needing new credit, and the plumbing keeps getting easier, "the SEC's changing the rules around securitizations for anything related to AI" and Nvidia's Jensen Huang is "bragging about creating a compute security derivative." His advice is nuanced, not apocalyptic: take profits and rotate to gold and industrials, but don't short, as his host countered, "the track has been laid for NVIDIA to make $300 billion in operating profits," and Google alone has $800 billion in forward purchase commitments that will show up in Q4 supplier earnings.

The sharpest single line came via the accountants relaying Michael Burry (the investor who called the 2008 crisis) who has increased his short positions against AI names and called Nvidia's $500 billion structure a case of "shades of Enron," using "unnatural credits to prolong momentum late in the bull phase… a sign of desperation." Their accounting-desk detail worth filing: AI chips realistically last 2–3 years, but companies have stretched the accounting assumption to 5–6 years, which lowers annual depreciation and flatters profits today. And the bookend came from perennial bear Marc Faber, who warned the eventual unwind "will not be an ordinary recession… it will be devastating. It will be terrible."

The debate

Steel-manning both sides of the hyperscaler capex thesis (~$700 billion in 2026, JPMorgan's tally is ~$697 billion, heading toward ~$1 trillion-plus in 2027). Yesterday the fight was about Nvidia's financing plumbing and Google's cash crunch. This cycle it widened: a second chip giant (Broadcom) is now borrowing tens of billions to fund its customers, and the bears added a former Fed president and a valuation scorecard. The narrowed question: does the demand hold up, and the funding stay open, long enough for revenue to catch the spending?

Bull: the demand is real, the deals keep getting done, and the profit pool is enormous. Even the loudest macro bear on tape this cycle refuses to short: as Luke Gromen's co-host put it, "the track has been laid for NVIDIA to make $300 billion in operating profits… pure profit," and Google has $800 billion of forward purchase commitments that will convert to supplier revenue (Monetary Matters, Aug 20). The chip supply chain keeps signing deals that only make sense if demand is durable: Google's Marvell warrants require $120 billion of spending over six years, and AI networking alone is projected to grow from ~$10 billion to $245 billion by decade's end (The AI Investor Podcast, Aug 20). Broadcom's ability to line up $60–70 billion of senior secured debt is itself a demand signal, sophisticated lenders are willing to fund it (Bloomberg Businessweek, Aug 20). And one retail-investor framing is worth steel-manning: Nvidia and Broadcom financing their customers is less "circular fraud" than "car-maker offering loans to buy the car," normal when real outside demand exists (Buy Hold Rant, Aug 19).

Bear: valuations are near dot-com extremes, the growth rate is about to slow, and the financing is a structure that has failed before. Bill Dudley's scorecard is the cleanest bear case of the cycle: Shiller CAPE 41, a real equity risk premium of just 1.1%, a Buffett indicator at 240%, and a $2 trillion revenue gap to justify a ~$5 trillion capital base, plus the timing argument that a slower increase in 2027 capex is enough to compress supplier margins (Bloomberg Talks, Aug 20). Andy Constan's warning is about mechanics, not sentiment: capital markets "just close" after a burst, and Nvidia's shift to indirect SPV financing signals it is "out of capacity to do direct deals" (Excess Returns, Aug 20). Michael Burry has raised his shorts and called the structure "shades of Enron," resting on chips whose accounting life has been stretched from 2–3 years to 5–6 to flatter profits (The Accounting Podcast, Aug 19). Luke Gromen's history lesson: all five prior US capex booms since the 1840s ended in a bust (Monetary Matters, Aug 20). And the private labs just missed their whisper numbers (Anthropic ~$65B vs ~$80B expected; OpenAI ~$40B), the first demand wobble (The AI Investor Podcast, Aug 20).

The synthesis. Both camps still agree the technology is transformative. What's new this cycle is that the bear case moved from vibes to a scorecard (Dudley) and from Nvidia-only to industry-wide (Broadcom's debt raise), while the bull case picked up its own hard proof (the $120 billion Marvell commitment, the $800 billion Google backlog, lenders lining up for Broadcom). The debate now turns on two referees, both arriving fast: Nvidia's quarter next Wednesday (does it show demand it is not itself financing?), and the health of the funding market (does a compute-linked bond deal or a private-credit fund wobble before the money is raised?).

Sell signals to watch:

  • A funding-market stumble. Broadcom's $60–70 billion raise, Nvidia's SPV structure, and the SEC's easier AI-securitization rules all depend on lenders staying hungry. The first oversubscribed deal that suddenly isn't, or a private-credit fund (watch Blue Owl) marking AI paper down, is the crack (Bloomberg Businessweek, Aug 20; Excess Returns, Aug 20).
  • Demand missing the bar. The Anthropic/OpenAI run-rate miss was the first; a second below-whisper print, or soft cloud-backlog conversion, would confirm the "air pocket" (The AI Investor Podcast, Aug 20).
  • The growth rate rolling over. Per Dudley, watch for any hyperscaler whose 2027 capex increase is smaller than 2026's, that's the margin-compression trigger, not an outright cut (Bloomberg Talks, Aug 20).
  • Rising financing costs and yields. With the 30-year Treasury near 5.3% and issuance surging, a hyperscaler forced to pay up on a bond, or long yields spiking on fiscal fears, weighs directly on valuations (Bloomberg Talks, Aug 20).

Stocks in play

NVDA. Bull: the demand backdrop is still overwhelming and the profit pool is real, even committed bears won't short it because "the track has been laid for NVIDIA to make $300 billion in operating profits" (Monetary Matters, Aug 20). Bear: its move to indirect SPV financing signals it is "out of capacity to do direct deals" while still needing to prop up its own customers, and Michael Burry calls the structure "shades of Enron," resting on chips whose depreciation life has been stretched 2–3 to 5–6 years (Excess Returns, Aug 20; The Accounting Podcast, Aug 19). Next catalyst: Q2 FY2027 earnings, Wednesday Aug 26 (after the close), the demand referee for the entire complex.

AVGO. Bull: it is now the second chip giant to industrialize AI-scale financing, arranging $60–70 billion of senior secured debt (up to $100 billion with Apollo/Blackstone) to fund 20 gigawatts of compute for Anthropic and others, under a CEO famous for financial engineering; Bank of America still pegs its custom-chip market at $250–350 billion, over 60% of the pie (Bloomberg Businessweek, Aug 20; The AI Investor Podcast, Aug 20). Bear: it is taking on enormous debt to finance its own customers' purchases (the vendor-financing risk in sharpest form), Google is actively multi-sourcing its TPU to Marvell and MediaTek, and the shares fell ~10% around the Marvell news (The AI Investor Podcast, Aug 20). Next catalyst: terms and take-up of the $60B+ debt deal; read-through from Nvidia's Aug 26 print.

AMD. Bull: it remains a second source as hyperscalers deliberately diversify away from single suppliers, and the multi-year demand case (Lisa Su's "100x more compute," from last cycle) is intact. Bear: no fresh MI450X/Helios catalyst surfaced this cycle, and it is squeezed between Nvidia's dominance and hyperscalers' own custom silicon (now Broadcom and Marvell). Largely quiet on the tape beyond being named in the off-balance-sheet review (The Banker Next Door, Aug 20). Next catalyst: its second-half data-center step-up; Nvidia read-through.

MSFT. Bull: it is quietly the "good collateral" that makes the new financing work (a Microsoft lease is what lets junk-rated borrowers raise money), and neoclouds like IREN are delivering capacity for it, a sign its cloud demand is real and outsourced where useful; its balance sheet also looks relatively cleaner, with on-balance-sheet debt of ~$40 billion against larger peers' commitments (Buy Hold Rant, Aug 19; The Banker Next Door, Aug 20). Bear: it still carries ~$229 billion of off-balance-sheet purchase commitments and ~$329 billion of leases not yet started, and the open question is whether cloud demand converts capex to cash (The Banker Next Door, Aug 20). Next catalyst: evidence cloud demand keeps free cash flow positive as capex climbs.

GOOGL. Bull: it keeps signing the deals that define the trade, the Marvell warrants imply $120 billion of spending over six years, and its "sleeper" holdings delivered a ~$100 billion gain on SpaceX this quarter; the cloud backlog remains enormous (The AI Investor Podcast, Aug 20). Bear: it is the poster child for the off-balance-sheet build, $811 billion of purchase commitments as of June 30 (up from $332 billion just a quarter earlier), off-balance-sheet obligations up roughly 800% year over year, energy deals running out to 2054, and its recently issued bonds are trading below their issue price (The Banker Next Door, Aug 20; Excess Returns, Aug 20). Next catalyst: whether accelerating cloud revenue converts to real free cash flow rather than more debt and equity.

AMZN. Bull: AWS keeps compounding, and Amazon holds warrants in Anthropic, real equity upside in the fastest-growing AI lab. Bear: it sits inside the ~$3 trillion off-balance-sheet tally, with ~$200 billion of annual capex that has to convert to durable outside demand; no fresh Amazon-specific print this cycle beyond being named in the WSJ review (The Banker Next Door, Aug 20). Next catalyst: evidence AWS AI demand is broad and durable.

META. Bull: it retains the most ad-funded optionality of the group to absorb its capex line. Bear: its Hyperion mega-campus is the textbook case of off-balance-sheet engineering, a joint venture that is 20% Meta, 80% Blue Owl, with $27 billion of construction debt and a ~$12.3 billion initial lease commitment kept off Meta's balance sheet until rent starts; Meta's off-balance-sheet obligations are up roughly 780% year over year, and its private-credit partner Blue Owl is itself under pressure from bad loans and redemptions (The Banker Next Door, Aug 20). Next catalyst: whether ad-revenue growth keeps outrunning capex; otherwise quiet on company-specific news this cycle.

Read-throughs

  • Custom silicon and networking (Marvell, Broadcom, optics), this is where the incremental dollars are landing. Google's Marvell deal ($120 billion of spending over six years to trigger the warrants; ~$20–25 billion a year by 2028, more than twice Marvell's current revenue) plus Bank of America's TPU market map (Broadcom $250–350B, MediaTek $100–150B, Marvell $50–100B) confirm the interconnect/custom-chip leg is the fastest-growing slice, AI networking alone projected to go from ~$10 billion (2023) to $245 billion by decade's end. Broadcom's ~10% two-day drop around the news is, per the (bullish) hosts, a market misread of a "most hated and misunderstood" stock, not a fundamental crack (The AI Investor Podcast, Aug 20). Actionable: the networking/optics complex is the highest-conviction "shovels" trade; Broadcom weakness on multi-sourcing headlines is the debate to have into Nvidia's print.
  • Memory (Micron, SK Hynix, Samsung), a digestion scare, not yet a break. The below-whisper Anthropic/OpenAI run-rates revived a narrative that engineers will "engineer around memory" (a Cathie Wood line), and memory names took an extra hit on the day. Watch this as the first real digestion signal of the cycle, the memory group is where an "air pocket" would show up first, offsetting the cash-return story (SK Hynix's $29 billion buyback) from earlier this week (The AI Investor Podcast, Aug 20; Buy Hold Rant, Aug 19).
  • Financing / private credit, the systemic read got bigger and more concrete. With Broadcom now arranging up to $100 billion (Nvidia already at $500 billion via SPVs), the SEC easing AI-securitization rules, and Meta's Hyperion showing the mechanics in the open (20/80 Meta/Blue Owl, $27 billion of bond-financed debt, Blue Owl under redemption pressure), the first sign of stress will almost certainly appear in a repriced compute-linked bond or a private-credit mark, not in a hyperscaler's headline earnings (Bloomberg Businessweek, Aug 20; The Banker Next Door, Aug 20; Monetary Matters, Aug 20).
  • Power / thermal (VRT, ETN), utilities and nuclear, quiet on the tape this cycle. No named commentary on Vertiv, Eaton, ERCOT, Texas SB6, or nuclear PPAs surfaced in this window, a coverage gap, not a signal. The one energy episode was about using AI inside utilities, not powering AI. The structural bottleneck (transmission, grid connection, behind-the-meter build) from prior issues stands unchanged.

What changed vs last issue

Yesterday's issue (Thursday, Aug 20, "Nvidia turns chips into collateral. Google's cash ran dry.") laid out the mechanics of Nvidia's $500 billion financing plan and Google's first-ever negative free cash flow. This cycle the story broadened on two fronts, a second chip giant joined the debt wave, and the bears gained a heavyweight voice:

  • The vendor-financing model spread to Broadcom. New this cycle: Broadcom is arranging $60–70 billion of senior secured debt (up to $100 billion with a $30 billion Apollo/Blackstone junior piece) to fund 20 gigawatts of compute for Anthropic and others, an extension of the June Broadcom, Blackstone and Apollo pact. The "borrow to build" structure is no longer Nvidia-only.
  • A former Fed president put a valuation scorecard on the bubble. New: Bill Dudley, Shiller CAPE 41 (vs 44 in Dec 1999, ~17 average), real equity risk premium 1.1%, Buffett indicator 240%, a $2 trillion revenue gap against a ~$5 trillion capital base, and the "change in investment, not the level" argument for why a still-huge-but-slower 2027 compresses margins.
  • The bear chorus broadened and got specific. New: Andy Constan's "capital markets just close" thesis and his read that Nvidia's shift to indirect SPV financing means it is "out of capacity" for direct deals; Luke Gromen's five-boom history (canals-to-telecom, all busts) and flag on the SEC easing AI securitization; Michael Burry's raised shorts and "shades of Enron" label plus the 2–3 to 5–6 year depreciation stretch; Marc Faber's "devastating" bookend.
  • Demand delivered its first miss. New: leaked run-rates for Anthropic ($65B) and OpenAI ($40B) came in below whisper numbers (~$80B for Anthropic), triggering a ~10% one-day drop in momentum/AI-infrastructure names.
  • Off-balance-sheet detail got granular. New per-company color on the (previously covered) ~$3 trillion WSJ figure: off-balance-sheet obligation growth of ~800% (Alphabet), ~780% (Meta), ~570% (Oracle), ~250% (Nvidia) year over year; Alphabet purchase commitments $811 billion (up from $332 billion a quarter earlier, energy deals to 2054); Meta Hyperion = 20/80 Meta/Blue Owl, $27 billion of debt, $12.3 billion initial lease.
  • Chip-deal detail hardened. New: Google's Marvell warrants require $120 billion of spending over six years (~$20–25 billion/year by 2028); BofA TPU market map (Broadcom $250–350B / MediaTek $100–150B / Marvell $50–100B, Broadcom >60%); AI networking TAM ~$10B (2023) to $245B by decade's end.
  • Quiet vs yesterday: no fresh Vertiv (VRT) or Eaton (ETN), ERCOT/SB6, or nuclear-PPA commentary; no new AMD product print; no fresh Microsoft or Amazon company print beyond their roles in the financing/off-balance-sheet story; no update on SK Hynix beyond yesterday's buyback.
  • Deltas on the big numbers: aggregate 2026 hyperscaler capex ~$697B (unchanged); off-balance-sheet commitments ~$3T (unchanged, now itemized); Nvidia backstop ≤25% / ~$125B (unchanged). New numbers: Broadcom $60–70B senior secured / up to $100B total; Marvell $120B over six years; Shiller CAPE 41; 30-year Treasury ~5.3%; hyperscaler "$2T revenue needed vs ~$5T base."
  • Still open: whether Nvidia's Aug 26 quarter shows demand it is not itself financing; whether the funding market stays open through the next wave of deals; whether the Anthropic/OpenAI miss was noise or the start of an air pocket; and whether a compute-linked bond or private-credit fund (watch Blue Owl) cracks first.

Next catalysts to watch: Nvidia's Q2 FY2027 earnings, Wednesday Aug 26 (after close), the demand referee for the whole complex; the Jackson Hole symposium (~a week out) and the new Fed chair's tone into September; terms/take-up of Broadcom's $60B+ debt deal; and any repriced compute-linked bond, neocloud stumble, or private-credit mark as the first sign the financing machine is straining.