Newsletter · · Ashutosh Agarwal

The AI Buildout Bill Comes Due in 2029 as Money Floods In Anyway - The AI Capex Tracker - Week of September 23, 2026

The AI Capex Tracker for the week of September 23, 2026, synthesizing investor podcasts published September 22. This issue covers SK Ventures' Paul Kedrosky putting a 2029 maturity-wall date on the AI financing reckoning, Alibaba's $53 billion AI plan and a new training chip aimed at Nvidia, SoftBank's oversubscribed junk-bond sale to fund OpenAI, and Meta's Muse hitting number one on the App Store as the first real capex-to-revenue proof point, with Micron's print eight days out.

The AI Capex Tracker

Week of September 23, 2026: The AI Buildout Bill Comes Due in 2029 as Money Floods In Anyway

Issue: Wednesday, September 23, 2026

(Covers podcasts published since the last issue, the batch dated Monday, September 22. After a quiet, power-focused Tuesday, the conversation swung hard back to the two questions that actually decide this trade: who's paying for the build, and what has to be true for it to pay off. A systems-minded investor put a specific date on the reckoning, 2029, and walked through why. Meanwhile the money kept flooding in from new directions: Alibaba committed more than $53 billion and unveiled a chip aimed straight at Nvidia, and SoftBank's junk-bond sale to fund OpenAI drew over $20 billion of demand. And for the first time, one of the hyperscalers showed a product that actually turns the capex into cash: Meta's Muse hit number one in the App Store. Three podcasts carried real substance; a fourth gave the memory read into next week's Micron print.)

TL;DR

  • The clearest bear framework in weeks, with a date on it. Paul Kedrosky of SK Ventures, a systems analyst who called the mechanics of the 2008 crisis, argued that "30% to 70% of US GDP growth" has been driven by AI capex alone, that "more than 60%" of it is now funded by borrowing rather than company cash flow, and that this debt hits a "maturity wall" around 2029 that "no freaking way" gets fully refinanced, and if it does, "at prohibitive terms." His one-line ROI test: with token prices falling "60 to 80% a year," a model company needs "in excess of 1000% growth a year" just to justify today's valuations (Between Two COO's with Michael Koenig, Sept 22).
  • The demand side went global and got a new challenger to Nvidia. Alibaba unveiled "China's most powerful AI chip built to take on NVIDIA," part of a broader plan with "more than $53 billion committed to AI" and "20 gigawatts of compute," aiming at a 5-to-10-trillion-parameter frontier model, while President Trump used his UN speech to say the US will "encourage" AI, "not rein it in," ahead of a Thursday meeting with Xi (Bloomberg Tech, Sept 22).
  • The money is still there, it's just increasingly borrowed. SoftBank's junk-bond sale to fund its OpenAI bet "already drew more than $20 billion in early demand" against a target to "raise more than $11 billion," pricing Thursday, and Meta finally showed the market a way to earn back the spend: its Muse AI agent hit "number one on the App Store free downloads," sending the stock to its "highest level since October," up "11%" (Bloomberg Tech, Sept 22).

What's new

A word on the window and the order. This is a Wednesday issue, sweeping the podcasts dated Monday, September 22. After Tuesday's grid-and-power cycle, the material this time is about financing and returns, the money map, not the megawatts. I've ranked by what moves a book: the fresh, dateable bear framework first (it reprices the whole trade if you believe it); then the global demand shock and policy backdrop; then the funding-market and product proof points; and finally the memory read into next week's Micron print.

A quick plain-English key for the terms below. Free cash flow is the cash a company has left after it pays for its buildings and equipment. External financing just means money raised from outside (bonds, loans, leases) rather than from that internal cash flow. A maturity wall is a moment when a big pile of borrowing all comes due at once and has to be paid back or re-borrowed ("refinanced"); if it all lands in the same year, a lot of borrowers get squeezed together. A token is the unit AI models are billed in (roughly a chunk of a word) so a falling "token price" means the core product is getting cheaper by the month. A junk bond is a bond from a borrower rated below the safe "investment-grade" band, so it pays a higher interest rate. A gigawatt (GW) is a billion watts, very roughly the output of one large power plant.

1. A systems investor put a date on the AI reckoning, 2029, and showed his work. Between Two COO's with Michael Koenig, Sept 22: Paul Kedrosky, of SK Ventures, an MIT fellow and CNBC contributor who describes himself as "a systems guy" and says he "approached this" the same way he "approached the global financial crisis." (This is specialist-investor analysis, not operator testimony, but it's unusually rigorous and specific.)

This is the freshest and most actionable bear case we've heard in weeks, because it does three things the usual "it's a bubble" chatter doesn't: it sizes the exposure, it names the mechanism, and it puts a date on it.

The size. Kedrosky says he noticed something odd about the economy about 18 months ago: "depending on the quarter, anywhere from 30% to 70% of US GDP growth was being driven by a single thing. And that single thing was AI CapEx." With the whole US economy growing around "2%" and a single category driving up to half of that, he calls it "historically so wildly unusual." His warning on the pattern is blunt: "there's never been an episode where we've had this spending on this scale [that] hasn't ended in a major recession, if not a depression." His historical analogs are the canals, the railroads, 1920s rural electrification, WWII rearmament, and the fiber-optic bubble, all moments when "a single category of economic activity" ate "a huge chunk of the economy." He frames today the same way: "a single sector is 40% of the S&P 500, more than half of S&P growth driven by 15 stocks."

The mechanism. The change "in the last year," he says, is where the money comes from: "more than 60% is coming from what's euphemistically called external financing… anywhere but free cash flow." That matters because outside investors "are looking for a specific yield… tied to the credit quality of the issuer," with "a duration… renewals, take-or-pay clauses." So the build creates "a fixed basket of obligations," your interest bill doesn't fall just "because token prices have fallen."

The date. Here's the part a PM should write down. Kedrosky draws the direct 2008 parallel: back then, an "explosion of mortgages… all hit a maturity wall at around the same time." He argues the same setup exists now, "an explosion of issuance in the last 18 months" of "exotic financing structures." His conclusion: "in about 2029, we'll hit a maturity wall, a five-year maturity wall… where all of this stuff has to be refinanced, especially anything that's tied to the… classic Oracle-style five-year GPU-backed lease." Will it all get refinanced? "No freaking way." And what does refinance will come "at prohibitive terms." In financial-crisis terms, he says, "we're kind of sitting in 2025, 26, looking at 27, 28 and saying… I already see this coming."

The ROI test. This is the number that reframes the demand debate. AI compute, he argues, is "a new industrial commodity" whose price is falling "anywhere from 60 to 80% a year." So, holding new premium models aside: "if I'm seeing a 70% year-over-year price decline in tokens… to stand still, I need a roughly 500% year-over-year growth." And "to deliver the kinds of returns that justify the valuations we're seeing… you probably have to deliver… something in excess of 1000% growth a year" for "at least the next decade." His verdict: "It would be the first commodity in history to do it… It seems wildly unlikely." He also flags that the heaviest current users (coders) are misleading, because coding is "expansive": "a small prompt explodes into thousands of lines of code," so extrapolating token demand from developers "is really misleading."

The operator's takeaway. Kedrosky's practical advice names names. "Be very careful about being overcommitted to companies who are most likely to see the most aggressive repricing of their debt loads because in four years they may not be there. CoreWeave is my favorite example." His counsel to any CEO: operate "with a scarcity mindset," because "the recent past is not representative of what's coming."

Why it matters: this is the bear case with a calendar attached. It says the risk isn't that AI is useless, it's that the financing structure (borrowed money, five-year GPU leases, a synchronized refinancing wall) turns a normal boom into a 2008-style air pocket around 2029, and that the return math needs a miracle to clear. It sharpens the single most important thing to watch on every hyperscaler and neocloud: how much of the build is funded with debt, and when that debt comes due.

2. Alibaba floored the accelerator, a $53 billion AI plan and a chip aimed at Nvidia, as Washington said "hit the gas." Bloomberg Tech, Sept 22: reported by Bloomberg's Peter Elstrom, with market reaction from Nancy Tengler (CEO/CIO, Laffer Tengler Investments). (Elstrom's segment is straight reporting; Tengler's is investor opinion.)

While the American bears do the debt math, China just widened the demand side of the ledger. At an event in Hangzhou, Alibaba "unveiled what it calls China's most powerful AI chip built to take on NVIDIA," an accelerator (the "GenWoo V900," per the segment) aimed at training frontier models, which Elstrom said is "three times better" than a chip Alibaba shipped only "three, four months ago." The chip is "just one piece of a much bigger push, with more than $53 billion committed to AI and plans for a massive expansion in computing capacity," specifically "20 gigawatts of compute." CEO Eddie Wu said his "primary objective now is going to be AGI," and the company is targeting "an LLM with 5 to 10 trillion parameters." Elstrom's framing of the competitive stack matters for Nvidia's China story: domestically "Huawei has really been ahead," with Cambricon "considered the next best alternative," and Alibaba "wants to be in that mix."

The policy backdrop cuts the same way. President Trump used his UN General Assembly address to say the US "rejects attempts of a global scheme to control AI," will "encourage" what he's now calling "super intelligence," and views existential-risk warnings as "a hoax": Bloomberg's Mike Shepard summed up the stance as wanting "to hit the gas as much as possible." Markets barely flinched (the 10-year sat at "4.95%," equities "largely flat" with tech outperforming), but the direction is clear ahead of a Trump-Xi meeting Thursday where AI is "expected to be part of the conversation." Tengler's read on the one thing that matters for US chipmakers: "will we be able to ship more NVIDIA chips [to China]? Will that be allowed? I think that's the only thing I'm really paying attention to."

Why it matters: Alibaba's $53 billion and 20 GW are a genuine demand datapoint (the AI build is not just four American hyperscalers), but the chip is a two-edged sword for Nvidia. A credible Chinese training accelerator, on top of Huawei and Cambricon, chips at Nvidia's China total addressable market exactly as US export policy stays unresolved. The bull reads it as "the build is global and accelerating"; the bear reads it as "Nvidia's China share is being competed and regulated away at once."

3. The money is still flowing, it's just borrowed, and Meta finally showed how the spend gets paid back. Bloomberg Tech, Sept 22

Two data points here speak directly to Kedrosky's thesis: one confirming the "borrowed money" worry, one pushing back on the "no returns" worry.

First, the funding. SoftBank's giant junk-bond sale to fund its OpenAI investment "already drew more than $20 billion in early demand" against a plan to "raise more than $11 billion" ("$10 billion dollar-denominated, €1 billion euros-denominated"), pricing "this Thursday." That is the external-financing engine Kedrosky is warning about, running at full throttle, and, for now, hungrily oversubscribed. The debt market hasn't blinked yet; it's leaning in.

Second, and more important for the bull case: Meta's Muse. The stock hit its "highest level since October," an "11% jump," because Muse (its consumer AI agent) is "number one on the App Store free downloads for Apple." Tengler, who admits she'd "been bearish on Meta" and only just bought it, called the shift material: "It is a way for them finally to monetize the CapEx spend. And I think that's very important." She now frames Meta as "trading as a value stock," with revenue "compounded at 20%," and adds that "the sell side is saying… it justifies Meta's spend." She also flagged, as a sign of how aggressively talent and tech are being bid, that "SpaceX bought Cursor… they paid $60 billion for a company that employed 20 people and was four years old." (Meanwhile SpaceX's own agent, Grok Bot, "topped 400,000 weekly users," "418,000… a 24% week-on-week jump.")

Why it matters: this is the first concrete "the capex converts to revenue" story in a while, and it landed on the lightest-capex name of the big four, which is exactly why the market rewarded Meta and not, say, Amazon. It's the bull's rebuttal to Kedrosky in miniature: if the agents monetize, the return math improves and the maturity wall gets easier to climb. The tell to watch is whether Muse's download spike turns into revenue, a number, not a chart position.

4. Micron broke out into its print, the memory verdict lands in eight days. Stock Market Today With IBD, Sept 22: IBD's market hosts (chart technicians, so read this as positioning color, not fundamental analysis).

Micron was IBD's "stock of the day," up "5% on the day" and finally clearing the "round number of 1,000" it had repeatedly failed at, "definitively getting out of this trading range." The hosts flagged the setup into the print: earnings "in eight days" (September 30), "triple digit growth expected on both the top and bottom line," and a "forward P/E of eight" that they called "just incredible for as much as this price has moved," because "the earnings are moving so much faster," keeping the forward valuation low. Chips broadly "sent Nasdaq to new highs," with risk-on tells like ARK Innovation "leading the market out into new high ground."

Why it matters: memory is the cleanest read on whether the AI hardware cycle is still tightening or rolling over, and Micron's Sept 30 print is the next hard catalyst on the calendar. A forward P/E of 8 on triple-digit growth is the bull's whole argument, cheap on this year's numbers, but it's also the setup that stings most if guidance signals a peak. (One useful cross-check against the noise: last issue's Telltales round-up pegged Micron consensus at "about $51 billion" revenue and "24 times free cash flow" going in.)

The debate

Steel-manning both sides of the $700B-plus 2026 hyperscaler capex thesis. This cycle didn't move the headline capex number (the weekend's Catz "$5 trillion, 2025-2030" and "$830 billion this year → $1.1 trillion next" still stands as the anchor). What it did was hand the bear case its most rigorous framing in weeks, and, at the same time, hand the bull two fresh data points.

Bull: the demand is global, the money is still eager, and the returns are starting to show up. Alibaba's "$53 billion" commitment and "20 gigawatts" of planned compute say the build is bigger than four American names and still accelerating (Bloomberg Tech, Sept 22). SoftBank's OpenAI bond drawing "more than $20 billion in early demand" says the financing channel Kedrosky fears is, for now, oversubscribed, not seizing up. And Meta's Muse hitting "number one on the App Store" is the first clean "capex → product → revenue" proof point, the bull's answer to "where's the ROI." Washington is pressing the accelerator, not the brake.

Bear: the growth is borrowed, the returns need a miracle, and the bill lands in 2029. Kedrosky's frame is the whole bear case, made concrete: "more than 60%" of the spend is now "external financing," creating "a fixed basket of obligations" that hits a "five-year maturity wall" around 2029 that "no freaking way" fully refinances, and if it does, "at prohibitive terms" (Between Two COO's, Sept 22). His ROI math is the killer: token prices falling "60 to 80% a year" mean a company needs "in excess of 1000% growth a year" to justify its valuation, "the first commodity in history to do it." Read the same SoftBank bond his way and the "$20 billion of demand" isn't comfort; it's the "explosion of issuance" that builds the wall. And a credible Chinese training chip means even the volume leader, Nvidia, faces a shrinking, contested China market.

The synthesis. The two sides are now arguing about the same fact from opposite ends: the flood of borrowed money. Bulls see an eager, oversubscribed debt market and a first real monetization win (Muse) and conclude the build is self-funding and starting to pay. Bears see the identical borrowing and conclude it's stacking a synchronized 2029 refinancing wall on top of a commodity whose price falls faster than any plausible volume growth. The tie-breaker is time: if the agents (Muse and its rivals) convert downloads into real revenue before the debt reprices, the bull wins; if token deflation outruns monetization into a 2029 refinancing crunch, Kedrosky's calendar wins. Meta just cast the first vote for the bulls; the debt market casts one every day the SoftBank book stays oversubscribed.

Sell signals to watch:

  • The SoftBank/OpenAI bond pricing wide or the book thinning. ">$20 billion of early demand" for a ">$11 billion" raise is the current all-clear; if that oversubscription fades or the coupon prices above expectations at Thursday's pricing, it's the first crack in Kedrosky's "external financing" engine (Bloomberg Tech, Sept 22).
  • The mix shifting from cash flow to debt across the hyperscalers. Kedrosky's "more than 60% external financing" is the number that matters; watch each name's capex-vs-free-cash-flow gap widen and its reliance on GPU-backed leases grow (Between Two COO's, Sept 22).
  • Neocloud debt repricing. Kedrosky named CoreWeave as the archetype of a debt-laden name facing the 2029 wall; a widening in neocloud credit spreads would be the early warning that his mechanism has started (Between Two COO's, Sept 22).
  • Muse downloads that don't convert to revenue. If Meta's App Store spike doesn't show up as monetization in the next print or two, the bull's freshest data point deflates and the "no ROI" bear thesis reasserts (Bloomberg Tech, Sept 22).
  • A credible Chinese training chip at volume. Alibaba's GenWoo V900, on top of Huawei and Cambricon, is the tell to watch on Nvidia's China TAM; volume adoption would confirm the share loss, not just the ambition (Bloomberg Tech, Sept 22).

Stocks in play

NVDA. Bull: still the volume leader, and the whole global build, now including Alibaba's "20 gigawatts" ambition, ultimately wants the best silicon it can get. Bear: the sharpest fresh negative this cycle: Alibaba "unveiled what it calls China's most powerful AI chip built to take on NVIDIA," a training accelerator "three times better" than its version from a quarter ago, joining Huawei and Cambricon, just as US export policy to China stays unresolved (Bloomberg Tech, Sept 22). Next catalyst: any concrete outcome from Thursday's Trump-Xi meeting on chip exports; the pace of Chinese custom-silicon adoption.

AVGO. Bull: the custom-ASIC arms dealer that gets paid whichever model wins; untouched by this cycle's financing-and-China news. Bear: sits squarely inside Kedrosky's "borrowed build, unproven ROI" math like the rest of the complex. Next catalyst: the next hyperscaler custom-ASIC order update. Quiet on the podcasts this cycle.

AMD. Bull: the credible number-two in accelerators, still riding the trillion-dollar re-rating from last week. Bear: no fresh company signal this cycle, and as the higher-beta AI-hardware name it's the most exposed to Kedrosky's rate/refinancing air pocket. Next catalyst: a company-issued MI450X/Helios deployment update and a named frontier anchor customer. Quiet on the podcasts this cycle.

MSFT. Bull: a "build-it-to-rent-it" landlord with the balance sheet to carry debt through a repricing. Bear: named indirectly in Kedrosky's warning: the "Oracle-style five-year GPU-backed lease" structure is exactly the financing he says hits the 2029 wall; last issue's depreciation stretch (server life 15→25 years) still flatters EPS. Next catalyst: Azure growth against the capex line; whether peers copy the depreciation change. Quiet on the podcasts this cycle.

GOOGL. Bull: a cash-machine landlord that funds more of its build from its own free cash flow than most, the cleanest name against Kedrosky's "60% external financing" critique. Bear: last issue's ad-tech remedy (six-year compliance officer, "60 days to comply") is still an implementation overhang; nothing new this cycle. Next catalyst: the ad-tech remedy timeline; free-cash-flow trajectory against a rising capex line. Quiet on the podcasts this cycle.

AMZN. Bull: AWS is the offtake engine behind the build. Bear: the AI-agent land grab has a competitive edge for it: Amazon is "not really wanting to play ball with [Meta's] Muse on behalf of customers going to merchant websites," while Shopify is willing, ceding first-mover agent-commerce ground (Stock Market Today With IBD, Sept 22); and it carries the heaviest absolute capex to justify. Next catalyst: proof AWS demand is broad beyond the frontier labs; its own agent-commerce answer.

META. Bull: the standout this cycle: Muse hit "number one on the App Store free downloads," the stock its "highest level since October," +"11%," and a former bear (Nancy Tengler) bought it, calling Muse "a way for them finally to monetize the CapEx spend" (Bloomberg Tech, Sept 22). Bear: it's a download rank, not revenue yet, and Meta is still the poster child for the off-balance-sheet build (the "$27 billion" Hyperion JV bond from prior issues). Next catalyst: whether Muse's spike converts to disclosed revenue; how much more of the build moves off the books.

Also in play: BABA (Alibaba): "more than $53 billion committed to AI," "20 gigawatts of compute," a 5-to-10-trillion-parameter frontier-model target, and a new training chip (GenWoo V900) aimed at Nvidia; CEO Eddie Wu's "primary objective now… AGI" (Bloomberg Tech, Sept 22). MU (Micron): broke out above "1,000," up "5%," "triple digit growth expected," "forward P/E of eight," reports September 30 (~"$51 billion" consensus from last issue) (Stock Market Today With IBD, Sept 22). SoftBank / OpenAI: junk-bond sale for the OpenAI bet drew ">$20 billion in early demand" against a ">$11 billion" target, pricing Thursday, the external-financing engine running hot (Bloomberg Tech, Sept 22). CoreWeave: named by Kedrosky as his "favorite example" of a debt-laden name most exposed to the 2029 refinancing wall (Between Two COO's, Sept 22). SHOP (Shopify): up "15%" in two sessions on willingness to partner with Meta's Muse for agent-driven purchases, the winner of the same news that pressures Amazon (Stock Market Today With IBD, Sept 22).

Read-throughs

  • Memory / networking: the calendar is the catalyst, and it's next Wednesday. Micron's September 30 print is the cleanest read on whether the memory tightness the bulls love is holding. The setup into it is a breakout above "1,000" on "5%," "triple digit growth expected," at a "forward P/E of eight," cheap on this year's numbers, which is the bull's whole case and the bear's whole risk if guidance signals a peak (Stock Market Today With IBD, Sept 22). No fresh HBM/SK Hynix/networking-silicon operator color this cycle.
  • Rates / credit: this cycle's whole story, and the signal is the SoftBank book. The most actionable trade isn't a stock, it's a spread: watch Thursday's SoftBank/OpenAI bond pricing. ">$20 billion of demand" for a ">$11 billion" raise is the market voting that the financing engine is fine; a wide print or a thinning book is the first evidence for Kedrosky's 2029 wall. And watch neocloud credit, Kedrosky's named CoreWeave as the tell (Bloomberg Tech, Sept 22; Between Two COO's, Sept 22).
  • Chips / China: the two-way trade on Nvidia's China TAM. Alibaba's training chip, on top of Huawei and Cambricon, is the concrete datapoint that China's domestic accelerator stack is catching up. The actionable read: it caps the upside case on Nvidia's China revenue exactly as Thursday's Trump-Xi meeting could reset export rules either way. Own the global-demand story; hedge the China-share leg (Bloomberg Tech, Sept 22).
  • Software / agents: the first "capex pays off" read, and it picks winners and losers. Meta's Muse is the proof-of-concept that a hyperscaler's spend can convert to a consumer product people actually download. The read-through already has a winner (Shopify, willing to plug into agent commerce, +"15%") and a laggard (Amazon, "not… wanting to play ball"), the agent-commerce layer is becoming a competitive battleground, not just a cost center (Bloomberg Tech, Sept 22; Stock Market Today With IBD, Sept 22).
  • Power / grid: quiet this cycle. After Tuesday's grid-heavy issue (Edison's "$1.4 trillion," Amazon-Generac, Oracle's PPA), no fresh operator color on Vertiv, Eaton, Generac, ERCOT, or nuclear PPAs (Vistra, Constellation, Talen) in the Sept 22 podcasts. The power story didn't reverse, it just wasn't where the microphone was this cycle.
  • Optics: quiet this cycle. No new signal on Coherent, Lumentum, Fabrinet, Marvell, Astera Labs, or Credo.

What changed vs last issue

Last issue (Tuesday, Sept 22, "$1.4 trillion for the grid, and it cuts your bill.") was about the physical build: Edison International's Pedro Pizarro putting the electric industry's own five-year investment at "$1.4 trillion" and arguing data-center load can cut customer rates; Amazon's "$2.4 billion" Generac backup-generator deal; Oracle's equity-not-bonds funding and "433-MW" PPA; and AMD touching a trillion. This edition covers the Monday, Sept 22 podcasts, and the story swung from megawatts back to the money map and returns.

  • The bear case got its most rigorous framing in weeks, and a date. New this cycle: Kedrosky's "30-70% of GDP growth," "more than 60% external financing," a "2029… five-year maturity wall," and the "1000% growth a year" ROI test. Last week's bears (Wellum's "debt bubble," the weekend's historians) gestured at the risk; Kedrosky put a mechanism and a calendar on it (Between Two COO's, Sept 22).
  • China entered the capex ledger, with a chip aimed at Nvidia. New this cycle: Alibaba's "$53 billion," "20 gigawatts," 5-to-10-trillion-parameter model, and a training accelerator "built to take on NVIDIA," the first major non-US hyperscaler capex datapoint in this letter in a while (Bloomberg Tech, Sept 22).
  • The funding engine got a live example, and it's oversubscribed. New this cycle: SoftBank's OpenAI junk bond ">$20 billion of early demand" for a ">$11 billion" raise, pricing Thursday. Last issue's funding story was Oracle avoiding the bond market; this cycle's is a debt market begging to lend (Bloomberg Tech, Sept 22).
  • META flipped from "quiet" to the standout, with the first real monetization tell. Last issue META was "quiet on the tape"; this cycle Muse hit "number one on the App Store," the stock is at its "highest level since October," +"11%," and a prior bear bought it on the "monetize the CapEx" thesis (Bloomberg Tech, Sept 22).
  • AMD cooled off. Last issue's trillion-dollar milestone had no follow-through in the podcasts this cycle.
  • Quiet this cycle: AVGO, AMD, MSFT, GOOGL (no fresh operator color); power/grid (Tuesday's whole story, no new color); memory/optics operator commentary (Micron's Sept 30 print is the catalyst); dedicated nuclear PPAs (Vistra, Constellation, Talen). No new aggregate hyperscaler capex-guidance number, the weekend's "$5 trillion, 2025-2030" anchor still holds.
  • Still open: whether the SoftBank book stays oversubscribed through Thursday's pricing; whether Muse's downloads convert to revenue; whether Chinese training chips reach volume; and whether Kedrosky's "60% external financing" mix keeps rising.

Next catalysts to watch: the SoftBank/OpenAI bond pricing (Thursday) as the credit-market tell; the Trump-Xi meeting (Thursday) on chip exports; Micron earnings (Wed, Sept 30, ~$51B consensus) as the memory verdict; and any follow-through on Meta's Muse monetization or Alibaba's chip adoption.