Newsletter · · Ashutosh Agarwal

Memory Sells Out Through 2028 as Nvidia Heads Into Earnings - The AI Capex Tracker - Week of August 24, 2026

AI capex newsletter for the week of August 24, 2026. Heading into Nvidia's Wednesday print, high-bandwidth memory sold out through 2028 while the bear case narrowed to a credit-cycle worry over how the build-out is financed.

The AI Capex Tracker

Week of August 24, 2026: Memory Sells Out Through 2028 as Nvidia Heads Into Earnings

TL;DR

  • Nvidia reports Wednesday after the close, and it's the referee for the entire AI trade. Bernstein's top chip analyst, Stacy Rasgon, expects "very, very good" numbers as Nvidia enters its next chip cycle (called Rubin), but says the two things that actually matter are whether Nvidia can keep its profit margins in the mid-70s as memory and wafer costs climb, and how much detail it gives on the financing deals where it is, in his words, "donating their credit rating" to help customers borrow. Nvidia has teased $1 trillion of revenue over three years, which would mean roughly $500 billion of data-center sales next year alone. (CNBC's "Fast Money", Aug 21)

  • Memory just flipped from "digestion scare" to "sold out." High-bandwidth memory, the specialized chips that feed the GPUs, is booked solid through the end of 2027 and likely 2028, ordinary memory contract prices are up 80-95%, SK Hynix just reported revenue up 250% and profit up 500%, and Micron's CEO says customers want about 50% more supply than the company can commit to. If you don't own the memory makers, you don't own the tightest link in the AI chain. (Grey Beards on Systems, Aug 22; Behind the Ticker, Aug 23)

  • The bear case sharpened into one idea: this is a credit story, not a demand story. A wall of macro voices this weekend reframed the whole build-out as a financing bet, Jeff Snider's "gray swan" comparison to the Nortel/Lucent telecom-vendor collapse, private-credit funds quietly "gating" (blocking) withdrawals, and neocloud CoreWeave burning through $13.66 billion of cash in a year. Their shared point: the risk isn't whether AI is useful, it's whether the debt keeps flowing long enough for revenue to catch up. (Eurodollar University, Aug 21; The SharePickers Podcast, Aug 22)

What's new

A note on the window: this is the Monday issue, sweeping the weekend plus Friday evening (Aug 21-23). It was a quiet weekend for company news but a loud one for opinion, the sweep was heavy on macro strategists, value investors and bear-side commentators, plus two genuinely useful specialist voices (a Bernstein chip analyst and a memory-industry analyst). Almost no hyperscaler operators spoke. I've flagged who's an insider or specialist and who's an opinion. Everything below is ranked by where the dollars land hardest for a book heading into a huge week.

1. Nvidia's print is Wednesday, and the specialist scorecard says "good number, watch the margin and the financing." CNBC's "Fast Money", Aug 21, Stacy Rasgon, Bernstein senior semiconductor analyst (a top-ranked sell-side specialist), plus technician Carter Worth.

Nvidia reports after the close on Wednesday, August 26, and Rasgon's preview is the cleanest institutional read of the cycle. His base case: "I think it's going to be a good print... the numbers broadly should be very, very good." The company is right at the start of its next chip generation, Rubin, and product-cycle starts are historically good for the stock. Nvidia has already guided to roughly $1 trillion of revenue over a three-year period, which, as Rasgon put it, "would suggest something close to $500 billion in data center revenues next year," and he thinks the build plans make "numbers that are higher than that start to become pretty plausible." He'd love to see management actually commit to that.

But the two swing factors are not the headline number. First, gross margin, the share of each sales dollar Nvidia keeps as profit. It sits in the mid-70s, and Rasgon wants confirmation Nvidia can hold it there "particularly as things like memory pricing and wafer pricing is coming up." (That is the direct link between today's #2 story, surging memory prices, and Nvidia's own P&L.) Second, the financing backstops. Rasgon was blunt about the mechanics:

"NVIDIA, in some sense, has been sort of donating their credit rating, kind of, and offering backstops for some of those deals. And so any more details we can get on that, I think, would be useful and helpful."

His broader framing is worth holding onto: "none of AI ultimately works if NVIDIA is not working." And there's a tape signal underneath, Nvidia is on a six-day losing streak (its longest since September 2022), yet over the past month the broad chip index (the SOX) is down about 5% while Nvidia is up 4-5%, an early reversal of a year of underperformance. Carter Worth, the show's chart reader, wants to "be long NVIDIA into the print." Why it matters: this is the single biggest P&L event of the next month for the whole complex. A clean beat that holds margins and adds financing detail confirms the bull case; a margin wobble or vague financing answer is the crack the bears are waiting for.

2. Memory went from "digestion scare" to "sold out through 2028", the supply data hardened dramatically. Grey Beards on Systems, Aug 22, Jim Handy, Objective Analysis (a veteran memory-industry analyst, reporting from the Future of Memory & Storage conference); Behind the Ticker, Aug 23, Howard Chan, Kurv Investment Management (a portfolio manager who runs a memory-focused fund).

This is the most important reversal of the cycle. Just last issue, the memory group was where a demand "air pocket" might first appear. This weekend the physical supply data pointed the opposite way. Handy, walking through the industry's biggest memory conference, laid out the picture: high-bandwidth memory (HBM, the stacked, high-speed memory bolted next to each GPU) is "sold out through the end of 2027," and there's "a high likelihood that... in the next couple of months, it's going to be sold out through 2028." Ordinary memory is caught in the wake: DRAM contract prices are up as much as 80 to 95%, and some forecasts show DRAM revenue "rising hundreds of... percentage points year over year in 2026." His killer point is that this isn't only an HBM story: "even if HBM didn't exist, we would have a DRAM and NAND flash shortage right now," because the hyperscalers are buying so much of everything. The shortage has now spread to NAND flash, hard drives (Western Digital, Seagate, Toshiba are capacity-constrained), SSDs, circuit boards, and even the small stuff, power-management chips, ceramic capacitors, resistors and inductors. He described the hyperscalers as measuring themselves purely by spending: "they're in a race to outspend each other... they talk about tens of billions of dollars and they talk about... gigawatts."

Chan, whose fund holds the memory makers, gave the investable version. About 90% of the world's memory is made by just three companies, Micron, Samsung, and SK Hynix, and the AI build-out has turned a once-commoditized business into one with pricing power. His proof point: "SK Hynix just reported earnings. Their revenue went up 250% and their profit went up 500%." They are "price setters instead of price takers." He calls memory "the mines" (not the picks and shovels), sees a two-to-four-year runway because a new fab costs $30-50 billion and years to build (and the lithography machines come only from ASML, with years-long waitlists), and notes a secondary effect already showing up: Apple, MacBooks, iPhones and Xbox are raising prices because memory costs more, as the big three crowd out commodity chip production. He treats the recent chip pullback as an entry point. And a corroborating data point from an AI-generated markets recap: Micron just committed $10 billion to a memory research lab in Boise, the same week its CEO said data-center customers are asking for about 50% more supply than Micron can currently commit to (Telltales, Aug 23). Why it matters: the tightest, highest-pricing-power link in the AI chain got tighter, not looser, a direct margin headwind for the GPU makers and a tailwind for the memory names.

3. The bear case stopped being about valuation and became about the credit cycle. Eurodollar University, Aug 21, Jeff Snider (macro strategist); The Competent Investor, Aug 21, Chris MacIntosh (macro portfolio manager); The SharePickers Podcast, Aug 22, an equity analyst running 40% cash.

Last issue the bears brought a valuation scorecard (Dudley) and an accounting label ("shades of Enron"). This weekend they brought something more specific: a credit-cycle mechanism, and a sharp new historical analogy. Snider's argument is the standout. He calls Nvidia a "gray swan", a risk hiding in plain sight that everyone can see but ignores because "it's bubble time", sitting at the center of the AI financing structure. His new frame is the dot-com telecom vendors: Nortel and Lucent lent their own customers the money to buy their equipment, betting the internet build-out would let those customers pay it back in time. The technology was right; the timing was wrong; both companies went bankrupt (Nortel in 2009). "It's not the technology that's bad... more often than not, it's just the timing."

The modern twist, per Snider: Nvidia isn't lending directly. Instead, Wall Street sets up special-purpose vehicles (SPVs, off-balance-sheet financing shells) seeded by shadow banks like BlackRock and Apollo, funded further by bonds sold to insurers and pension funds, which then lend to Nvidia's customers, collateralized by the very chips they buy. Because those chips depreciate fast, Nvidia sweetens the deal with a "residual value guarantee", a promise to make the lenders whole if the equipment's value collapses, that "doesn't show up on a balance sheet anywhere," which he likens to credit-default swaps a generation ago. His warning is that this is one credit cycle away from unwinding: he flags that "non-accrual loans" (loans the lender expects to lose money on) are "creeping up... pretty quickly in certain funds," and a "serious refinancing wall" for software borrowers next year is already driving redemptions. Pull one Jenga block, the debt stops flowing, and customers can't buy, collateral clauses trigger, and Nvidia has to step in.

MacIntosh sharpened the private-credit angle: there are now "hundreds of these funds that are all gated" (investors literally can't withdraw their money), and the toxic paper was sold to insurers and pension funds who "don't even know what they bought." His memorable line on the vendor financing: it's "like lending the neighbor next door. He's got a drinking problem and no job money." And SharePickers put concrete numbers on a neocloud, CoreWeave, the AI-cloud landlord, with a $50 billion market cap, roughly $100 billion enterprise value once debt is counted, and negative $13.66 billion of free cash flow over the past year, buying Nvidia chips with Nvidia's own equity injection and then borrowing against those chips to buy more. Why it matters: the debate has migrated from "is it expensive?" to "does the money keep flowing?", and that makes the health of the credit market, not just earnings, the thing to watch.

4. Power finally got a voice this cycle, and the picture is uglier than the grid headlines suggest. Motley Fool Hidden Gems Investing, Aug 23, Hanan Happy, co-founder and CEO of ExoWatt (an operator building on-site solar-plus-storage power for data centers, an interested party with a book to talk, so weight accordingly).

After a cycle of silence on power, one operator filled the gap with vivid numbers. The scale shift is staggering: a data-center "building block" has gone from 10-20 megawatts a few years ago to 300-700 megawatts today, individual campuses now average about a gigawatt (some reach 10 GW), and a single gigawatt consumes as much electricity as a million US households, "you're basically building a city from scratch for a million people." The national math is the alarming part: the US adds about 50 gigawatts of grid power a year and needs to close a 70-100 GW gap, while China adds 540 GW a year, more than ten times US capacity. Gas, the go-to fix, has its own wall: turbine backlogs of five to seven years, fuel-price volatility, and rising local opposition ("NIMBYism") even in gas-friendly Texas. His most quotable stat for a book: a one-year grid delay costs a hyperscaler about $12 billion in missed revenue (he cites SpaceX/Anthropic/Google deals running near $1 billion a month), and over $160 billion of data centers that were supposed to come online this year are not, because of community backlash and state moratoriums, a bipartisan problem. Customers have shifted from "give me a gigawatt or I'm not interested" to "any power you can give me, I'll take it." Why it matters: power availability, not chips, is increasingly the binding constraint on how fast the capex actually converts to revenue, bullish for anyone who shortens the path to electrons (behind-the-meter generation, turbines, grid gear), and a real execution risk for the hyperscalers.

5. Meta caught a courtroom and a soft spending guide in the same week. Telltales, Aug 23, an AI-generated markets recap, used here only for facts that are publicly verifiable or corroborated elsewhere.

The one big single-company development of the weekend was Meta's. A California jury found the company liable on 75,000 separate consumer-protection violations, a judge separately ruled its platforms a "public nuisance," and it now faces $375 million in civil penalties, with a bigger open case: 29 state attorneys general are in federal trial seeking up to $200 billion in damages plus limits on the algorithmic feed. On the money side, Meta guided third-quarter revenue to $61-64 billion, below where the street was sitting, and is cutting staff across WhatsApp, Instagram and Reality Labs for the second time this year. The recap pegs Meta's trailing free cash flow at $43 billion, down about 73% from the prior year, "not because the business slowed, because the build-out ate it." That free-cash-flow collapse was independently echoed by a venture investor who called it "the lowest it's ever been in the history of a company" (20VC, Aug 22). Why it matters: Meta is the clearest live example of AI capex eating the very cash flow that funds it, and now with a legal overhang on top.

The debate

Steel-manning both sides of the hyperscaler capex thesis (~$700 billion in 2026, still heading toward ~$1 trillion-plus in 2027). Last issue the fight widened to a second chip giant borrowing (Broadcom) and a former Fed president's scorecard. This weekend it narrowed to one question that Wednesday will start to answer: does real demand show through, in Nvidia's print and in the memory data, faster than the financing strain shows up in the credit market?

Bull, the demand is physically real, and the tightest link in the chain is getting tighter. The single best bull data point this cycle isn't a stock, it's a shortage: HBM sold out through 2027 and likely 2028, DRAM contract prices up 80-95%, and SK Hynix printing revenue +250% / profit +500% (Grey Beards on Systems, Aug 22; Behind the Ticker, Aug 23). You do not sell out a commodity two years forward on a bubble of air. Nvidia's own guide, $1 trillion over three years, ~$500 billion of data-center revenue next year, with Bernstein saying higher is "plausible", points the same way, and its stock has quietly started outperforming the chip index into the print (CNBC's "Fast Money", Aug 21). Even the bears concede the base demand: as the 20VC guest put it, "the demand for AI compute is not going to change... the issue is the ability to fund it in the short term," and the hyperscalers are "no one is better prepared to survive" a dislocation (20VC, Aug 22). And Microsoft's tell, AI cloud revenue growing 43%, is the kind of real, cash-converting demand that services the debt (Deep Values, Aug 21).

Bear, the whole thing rides on the credit cycle staying open, and the warning lights are on. The sharpest bear frame this cycle is Jeff Snider's: Nvidia is a "gray swan" at the center of a Nortel/Lucent-style vendor-financing structure, now dressed up in off-balance-sheet SPVs and residual-value guarantees, and "non-accrual loans... creeping up... pretty quickly" plus a software refinancing wall next year say the credit cycle is already turning (Eurodollar University, Aug 21). Private-credit funds are gating withdrawals, and the paper sits in pension funds that "don't even know what they bought" (The Competent Investor, Aug 21). Arthur Hayes' version: we've told ourselves "real estate equals technology when real estate is just real estate," slapping 100x-sales tech multiples on what are really debt-financed real-estate companies housing chips with a two-year useful life, "a massive capital misallocation" that cracks when the market starts rewarding the hyperscaler that cuts capex (Unchained, Aug 21). And the concrete casualty case already exists, CoreWeave, at negative $13.66 billion free cash flow, is what a neocloud looks like when demand is high; the 20VC guest expects half of all neoclouds to disappear within 36 months (The SharePickers Podcast, Aug 22; 20VC, Aug 22).

The synthesis. Both sides now agree on two things: the demand is real today, and the financing is fragile. The bull owns the present (sold-out memory, 43% cloud growth, a likely-good Nvidia print); the bear owns the tail (a credit cycle that "just closes," neoclouds that die, chips that depreciate faster than the loans against them). Wednesday is the first real referee, a clean Nvidia beat with held margins and clear financing detail tilts it bull; a margin miss or a vague financing answer hands the bears their catalyst.

Sell signals to watch:

  • A credit-market stumble. Snider's "non-accrual loans creeping up," more private-credit funds gating, a repriced compute-linked bond, or a widely-watched mark-down at a private-credit lender (Blue Owl, PIMCO's interconnected exposure) is the first crack (Eurodollar University, Aug 21; Deep Values, Aug 21).
  • Nvidia's gross margin slipping out of the mid-70s. With memory and wafer prices surging, a margin miss on Wednesday would confirm the bulls' own supply story is now taxing the GPU maker (CNBC's "Fast Money", Aug 21).
  • A hyperscaler rewarded for cutting capex. Per Hayes, the psychological break is the first time the market cheers a lower spending guide instead of a higher one, watch the reaction function, not just the number (Unchained, Aug 21).
  • A neocloud default. With half of them expected to fail inside three years, the first real CoreWeave-style default would test every residual-value guarantee behind it (20VC, Aug 22; The SharePickers Podcast, Aug 22).

Stocks in play

NVDA. Bull: entering the Rubin cycle with a likely-strong print, a $1 trillion / three-year revenue tease (~$500 billion of data-center sales next year, possibly more), and the stock already outperforming the chip index into Wednesday (CNBC's "Fast Money", Aug 21). Bear: it's the "gray swan" underwriting its own demand, offering residual-value guarantees on depreciating chips through off-balance-sheet SPVs, and "donating its credit rating" to customers who can't otherwise borrow; a mid-70s margin that's now under pressure from surging memory prices is the soft spot (Eurodollar University, Aug 21; CNBC's "Fast Money", Aug 21). Next catalyst: Q2 FY2027 earnings, Wednesday, Aug 26, after the close, the demand referee for the whole complex.

AVGO. Bull: still the custom-silicon and AI-financing leader from last issue (the $60 billion-plus debt deal, >60% of the custom-chip market per Bank of America), with real backlog behind it. Bear: no fresh operator color this weekend, and it carries the same vendor-financing and multi-sourcing risks (Google spreading work to Marvell and MediaTek) flagged last issue. Largely quiet on the tape this cycle. Next catalyst: earnings Wednesday, Sept 2, the print that resets custom-silicon and financing expectations (Telltales, Aug 23).

AMD. Bull: still the deliberate second source as hyperscalers diversify away from single suppliers, and named in passing as a beneficiary of smarter memory-tiering (its recent Maxxed acquisition came up at the memory conference) (Grey Beards on Systems, Aug 22). Bear: no fresh MI450X/Helios catalyst surfaced, and it remains squeezed between Nvidia's dominance and the hyperscalers' own custom silicon. Quiet on genuinely new product news this cycle. Next catalyst: its second-half data-center step-up; read-through from Nvidia's Aug 26 print.

MSFT. Bull: the "demand is real" poster child, AI cloud revenue grew 43%, and the stock recently jumped 9% on that (not on any accounting trick), making it the cleanest capex-to-cash story in the group; a venture investor's blunt verdict this weekend was "Microsoft is Mega" (Deep Values, Aug 21; 20VC, Aug 22). Bear: the off-balance-sheet build keeps growing, capital-lease obligations reportedly jumped from ~$27 billion to nearly $63 billion, so the question is whether cloud demand keeps converting fast enough to keep free cash flow positive (Deep Values, Aug 21). Next catalyst: continued proof that AI cloud growth outpaces the capex line.

GOOGL. Bull: rising capex is a direct positive for its suppliers (as the memory PM noted, hyperscaler cash outflows are the memory makers' cash inflows), and it keeps signing the deals that define the trade (Behind the Ticker, Aug 23). Bear: it went free-cash-flow negative for the first time in about 20 years on AI spending, and it's using derivative accounting to shrink a reported ~$43.8 billion of data-center loan guarantees down to an ~$815 million footnote, the poster child for how the true obligation stays off the books (The SharePickers Podcast, Aug 22; Deep Values, Aug 21). Next catalyst: whether accelerating cloud revenue turns into real free cash flow rather than more debt.

AMZN. Bull: AWS keeps compounding, and the demand backdrop (sold-out memory, cloud backlogs) supports its build-out. Bear: it sits inside the off-balance-sheet tally with roughly $106 billion of signed-but-not-yet-commenced data-center leases that stay off the active balance sheet until the buildings finish (Deep Values, Aug 21). Otherwise quiet on Amazon-specific news this cycle. Next catalyst: evidence AWS AI demand is broad and durable enough to convert the lease commitments to cash.

META. Bull: still the most ad-funded of the group, with a core business throwing off "hundreds of billions" of cash, it doesn't strictly need to borrow (20VC, Aug 22). Bear: a rough weekend, a California jury verdict (75,000 violations, $375 million penalties, a "public nuisance" finding) and a looming 29-state trial seeking up to $200 billion; a soft Q3 revenue guide of $61-64 billion; a second staff-cut round; and trailing free cash flow reportedly down about 73% to $43 billion as the build-out eats it, "the lowest it's ever been in the history of a company" (Telltales, Aug 23; 20VC, Aug 22). Next catalyst: which attorney general files next, and whether ad-revenue growth keeps outrunning capex.

Read-throughs

  • Memory (Micron, SK Hynix, Samsung), the highest-conviction "shovels" trade got stronger. This is the sharpest reversal since last issue: what looked like a digestion scare is now a hard sell-out. HBM booked through 2027 (likely 2028), DRAM contract prices +80-95%, SK Hynix revenue +250%/profit +500%, and Micron's customers wanting ~50% more than it can supply. A memory-focused PM frames it as a two-to-four-year pricing-power story and calls the recent pullback an entry point (Grey Beards on Systems, Aug 22; Behind the Ticker, Aug 23). The one caveat to hold: the same tightness that's bullish for memory is a margin headwind for Nvidia, watch that tension resolve on Wednesday.

  • Power / thermal (VRT, ETN), utilities, nuclear, a real signal at last, and it's a bottleneck, not a boom. After a cycle of silence, an on-the-ground operator quantified the wall: US grid additions of ~50 GW/year against a 70-100 GW gap, China adding 540 GW/year, gas-turbine backlogs of 5-7 years, and over $160 billion of data centers stalled this year by community opposition and moratoriums (Motley Fool Hidden Gems Investing, Aug 23). Actionable: this supports anything that shortens the path to electrons, power/thermal gear (Vertiv, Eaton), turbine makers, and behind-the-meter generation, while flagging grid delay as a genuine execution risk on capex-to-revenue conversion. No named commentary on ERCOT, SB6, or the nuclear names (Vistra, Constellation, Talen) surfaced this cycle, a coverage gap, not a signal.

  • Financing / private credit, the systemic read got a name (the credit cycle) and a mechanism. The clearest new lens: this is a credit-cycle story running underneath the equity story. Watch non-accrual loans and fund gating in private credit, the residual-value guarantees behind neocloud loans (interest rates on compute-backed loans reportedly fell from ~15% to ~5.9% in a year on the perceived Nvidia backstop), and stress markers like a spike in a hyperscaler-adjacent name's default-insurance cost, the recap flagged Oracle's credit-default-swap cost jumping to ~203 basis points and an S&P downgrade to BB- (Eurodollar University, Aug 21; Deep Values, Aug 21). The first sign of stress will show up here, in a repriced bond or a private-credit mark, before it shows up in a hyperscaler's headline earnings.

  • Neoclouds (CoreWeave and peers), the fragile middle of the chain. With one investor expecting half of them to fail within 36 months and CoreWeave already at negative $13.66 billion free cash flow, this is where a demand-plus-financing shock would land first. What separates survivors is capital efficiency and management, not headline revenue (20VC, Aug 22; The SharePickers Podcast, Aug 22).

What changed vs last issue

Friday's issue (Aug 21, "Broadcom's $60B debt deal. The bears find a Fed voice.") widened the story to a second chip giant borrowing and a former Fed president's valuation scorecard. This weekend narrowed the focus onto Wednesday's Nvidia print and delivered two genuinely new fronts, a hard memory-supply update and a credit-cycle reframing of the bear case:

  • Nvidia's print moved from "next week" to "this Wednesday," with a specialist scorecard attached. New: Bernstein's Stacy Rasgon expects a strong print, watches gross margin (mid-70s, now pressured by rising memory/wafer costs) and financing-backstop detail as the swing factors, and notes the $1 trillion / three-year tease implies ~$500 billion of data-center revenue next year. The technicals say "long into the print."

  • Memory flipped from digestion-scare to sold-out. New (and a reversal of last issue's "first digestion signal" framing): HBM booked through 2027 and likely 2028; DRAM contract prices +80-95%; SK Hynix revenue +250% / profit +500%; Micron customers wanting ~50% more than it can commit, plus a $10 billion Boise research-lab commitment. The shortage has spread to NAND, hard drives, SSDs, circuit boards and passive components.

  • The bear case became a credit-cycle case with a new analogy. New: Jeff Snider's Nortel/Lucent vendor-financing "gray swan" frame (beyond last issue's canal/telecom and "Enron" analogies), plus concrete credit-cycle tells, non-accrual loans creeping up, private-credit funds gating withdrawals, a software refinancing wall in 2027, and pension funds holding paper they don't understand.

  • Power got its first real voice this cycle. New: US 50 GW/yr vs China 540 GW/yr; a 70-100 GW gap; gas-turbine backlogs of 5-7 years; >$160 billion of data centers stalled by community backlash; a one-year grid delay ≈ $12 billion of missed revenue. Fills the power gap that last issue explicitly flagged as quiet.

  • New concrete casualty math. New: CoreWeave at negative $13.66 billion free cash flow, ~$100 billion enterprise value on a $50 billion market cap; a forecast that half of all neoclouds fail within 36 months.

  • New single-company news, Meta. New: a California jury verdict ($375 million, "public nuisance," a looming 29-state trial seeking up to $200 billion), a soft Q3 revenue guide ($61-64 billion), a second staff-cut round, and trailing free cash flow down ~73% to $43 billion.

  • Financing-mechanics detail hardened. New color on the (previously covered) off-balance-sheet story: Nvidia's residual-value floor of ~25%, compute-backed loan rates falling from ~15% to ~5.9% in a year, an H100 losing ~55% of resale value by year three, and Oracle's default-insurance cost spiking to ~203 bps with an S&P downgrade to BB-.

  • Macro backdrop. New: US national debt crossed $40 trillion; Treasury Secretary Bessent doubled long-bond buyback limits (from $2 billion to $4 billion per operation), funded by issuing more short-term bills; the 30-year yield ~5.3% (unchanged from last issue).

  • Deltas on the big numbers: aggregate 2026 hyperscaler capex ~$700 billion (unchanged); off-balance-sheet obligations ~$2-3 trillion (unchanged, re-tallied at ~$2.13 trillion in one weekend walk-through); Nvidia residual-value backstop ~25% (unchanged). New numbers: HBM sold out to 2027/2028; DRAM +80-95%; SK Hynix +250% rev / +500% profit; China 540 GW vs US 50 GW; $160 billion of data centers stalled; Meta Q3 guide $61-64 billion; US debt $40 trillion.

  • Quiet vs last issue: no fresh Broadcom or Amazon operator print; no new AMD product catalyst; no named ERCOT/SB6, Vistra/Constellation/Talen, or direct Vertiv/Eaton commentary (power came via a startup CEO, not the utilities); the ~$3 trillion off-balance-sheet figure was re-litigated, not newly expanded.

  • Still open: whether Wednesday's Nvidia print shows demand it isn't itself financing and holds its margin; whether the credit cycle stays open through the next wave of deals; whether the memory sell-out proves the bull case or simply squeezes the GPU makers' margins; and whether a neocloud or a private-credit fund cracks first.

Next catalysts to watch: Nvidia's Q2 FY2027 earnings, Wednesday, Aug 26 (after the close), the demand referee for the whole complex (Salesforce reports the same day); Broadcom, Snowflake and others on Wednesday, Sept 2; the aftermath of the Jackson Hole symposium and the Fed's September tone; and any repriced compute-linked bond, neocloud stumble, or private-credit mark as the first sign the financing machine is straining.