Newsletter · · Ashutosh Agarwal

A $20B Fund Blew Up, Chips Cracked, and Microsoft Soared - The AI Capex Tracker - Week of August 3, 2026

A synthesis of what investor podcasts said about AI capital spending for the weekend of August 1 to 3, 2026, including the forced liquidation of a roughly $20 billion levered AI hedge fund that helps explain the chip crash, blowing-out credit-default swaps on the biggest AI spenders, and Jim Chanos on hyperscaler incremental returns halving.

The AI Capex Tracker

Issue: Monday, August 3, 2026: A $20B Fund Blew Up, Chips Cracked, and Microsoft Soared

TL;DR

  • The chip crash the whole complex has been blaming on "fundamentals" turns out to have had a name. A roughly $20 billion, wildly over-leveraged AI hedge fund, Situational Awareness, run by 24-year-old Leopold Aschenbrenner, got margin-called and force-liquidated over the past month, and that forced selling is a big part of why memory, neocloud and semiconductor stocks fell 40-60% from their highs. Ken Griffin's Citadel stepped in and bought the margined portfolio; the stocks then snapped back, and Microsoft posted the biggest one-day market-cap gain in the history of any company. (Big Technology, Jul 31; The Wall Street Skinny, Jul 31)
  • The real story underneath the fireworks is credit, not stocks. The cost to insure the big AI spenders' debt is blowing out. Oracle's is at a multi-year high (~215 basis points, up from ~145 at the end of last year), SpaceX's has surged, CoreWeave's is near a record, and even Microsoft, Google, Amazon and Meta have crept up 15-20 basis points. AI-linked debt issuance hit ~$270 billion by early July, nearly twice all of 2025, and both Amazon's and BlackRock's recent mega-bond deals drew unusually weak demand. (Eurodollar University, Aug 1; Merryn Talks Money, Jul 31)
  • And Jim Chanos, first-person, dropped the sharpest bear number yet. The return the hyperscalers earn on each extra dollar they invest has roughly halved in 18 months, from ~100% to ~25%, with a couple of names now in the teens, and an accounting quirk is flattering the whole market, letting S&P profits grow ~28-29% versus a trend of 8-9%. (Prof G Markets, Jul 31)

What's new

A note on the window: this is a Monday issue, so it sweeps the weekend plus Friday evening, roughly Friday July 31 through Sunday August 2. The prior issue (Jul 31) graded "the prove-it week" and split Microsoft from Meta. This weekend's podcasts add the missing piece: why the tape whipsawed so violently, a forced fund liquidation, and what's actually cracking underneath, the credit market. Ranked by where the dollars of P&L impact are biggest.

1. The chip crash had a mechanical cause: one 4x-levered fund being liquidated. Big Technology, Jul 31, with Reed Albergotti, technology editor at Semafor; The Wall Street Skinny, Jul 31.

Situational Awareness was, as one host put it, the purest single-fund expression of the AI trade you could build. Its biggest longs were the "bottleneck" names, the neoclouds Nebius and CoreWeave, and the memory makers SanDisk and Micron, and it was short software (reportedly names like Adobe). That is a bet that demand for AI compute will explode while traditional software gets eaten. For two years it worked spectacularly: the fund was up more than 1,000% since its July 2024 launch and peaked above $45 billion in assets (Chit Chat Stocks, Jul 31).

The problem was leverage, reportedly around 4x, with no hedge. When the memory and neocloud names rolled over and software rallied, both sides of the book went the wrong way at once, and the prime brokers (the bank desks that lend the leverage) started issuing margin calls. Aschenbrenner had to sell. As Albergotti explained, in a note to investors Aschenbrenner said he sold "only to the point where effectively he covered his shorts," reporting the fund down about 67% on the month but still up ~80% on the year. When the fund shopped its portfolio to the street, Jane Street and Millennium passed; Citadel stepped in and bought only the margined slice.

"It's the market can stay irrational longer than you can stay solvent. If you are right, it doesn't matter… this wasn't an Archegos situation. We're not accusing this guy of having done anything wrong other than just not being a great trader." host, The Wall Street Skinny, Jul 31

Why it moves numbers: it reframes the last two weeks. The prior issues read the semiconductor sell-off as fundamentals turning. A large part of it was one fund being unwound. That matters for positioning, the mechanical seller may now be largely done (some think the forced selling actually made the bottom), which is why the stocks snapped back and why Microsoft's Thursday print detonated into the vacuum.

2. The Korean epicenter: this was systemic where the fund's positions were biggest. All-In, Jul 31; The Wall Street Skinny, Jul 31.

Because SK Hynix and Samsung's memory arm make up roughly half of Korea's KOSPI index, the unwind of a fund this large in those names became a national event. Something like 1.2 million leveraged Korean retail accounts were hit with margin calls, roughly one in 30 adults, and on All-In, Dave Friedberg said around 350,000 of them had already been fully liquidated, with that (already stale) count still climbing. The iShares MSCI South Korea ETF fell ~34% in a month; a 3x-levered semiconductor ETF fell ~69%. Why it matters: it's a live illustration of how much of this rally has been leveraged retail and hedge-fund money, the fuel that amplifies both the melt-up and the crash.

3. The credit market is the real signal, and it's flashing. Eurodollar University, Aug 1, with Jeff Snider; Merryn Talks Money, Jul 31, with credit strategist Neil Chadwick.

Snider's framing is the one to internalize: "Everyone believes the AI bubble is mainly about stocks… However, this one is all about credit, unlike 1999." A credit-default swap (CDS) is essentially insurance against a company failing to repay its debt; when its price rises, investors are paying up for protection. This week Oracle's CDS hit ~215 basis points (up from ~145 at the end of last year), SpaceX's jumped to ~185, CoreWeave's neared a record, and Nvidia, Meta, Amazon and Alphabet all made new highs at lower levels. The behavioral tell: for the SpaceX bond, "dealers were circulating prices for five-year default protection before the bond offering had even been formally announced. Investors wanted a hedge before there were bonds to hedge."

Two concrete cracks in the plumbing:

  • Amazon's $25 billion, eight-tranche bond drew initial orders of ~$62 billion, but after the banks tightened the spread, orders fell to ~$41 billion, only ~1.5x covered versus the ~4x that high-grade deals have averaged this year, and its longest bonds needed an extra 18-21 basis points of yield to clear.
  • A $12.5 billion data-center bond arranged through BlackRock (financing a Texas project built for Meta) took nearly a week and priced at a 7.53% yield, one of the highest for a blue-chip data-center deal in this cycle.

Chadwick's addition is the vendor-financing loop: Alphabet's forward spending commitments rose ~$500 billion in three months while future revenue grew nowhere near that, and the hyperscalers are now the biggest issuers of investment-grade corporate debt, a job that used to belong to banks. AI-linked debt issuance reached ~$270 billion by early July, nearly 2x all of 2025, with the top six tech firms accounting for ~$194 billion. Snider's one-liner: "Stocks priced the dream. Credit asks whether the dream can make the payments."

4. Chanos, in his own words: the returns are quietly collapsing. Prof G Markets, Jul 31, with Jim Chanos, the short-seller who called Enron.

Last issue quoted Chanos secondhand. Here he is at length, and the standout is a number he says his firm tracks closely, return on incremental invested capital, meaning how much extra operating profit each new dollar of spending buys:

"Most of the return on incremental invested capital… has been pretty much cut in half for the hyperscalers over the past year and a half. Companies that were earning 100% incrementally on their capital are now earning 25%. And a couple of the companies are now earning in the teens." Jim Chanos, Prof G Markets, Jul 31

His second point is an accounting one that explains why the market has felt so strong even as returns fade. When a hyperscaler spends $100 on a data center, it doesn't count that as a cost right away, it spreads it over 5-10 years as "depreciation." But the companies receiving that money (Nvidia, the utilities, the equipment makers) book it as revenue and profit immediately. That mismatch, Chanos argues, is why S&P profits are growing ~28-29% when the economy would justify ~8-9%. He notes that in the dot-com bust, S&P profits fell 40% from mid-2000 to mid-2001 as order books collapsed, and that his personal "foolproof" late-cycle indicator, heavy equity issuance, is now flashing, with 2026 on track for near-record stock sales. Why it matters: if the incremental-return decline continues 12-18 months, he expects the C-suites that swear they "can't lose the race" to start having very different conversations about the spend.

5. Microsoft's record snap-back proved the bull case is still alive. The Real Eisman Playbook, Jul 31, with Steve Eisman; Squawk on the Street 10AM, Jul 31, with Gil Luria of D.A. Davidson.

Into the deleveraging vacuum, Microsoft printed the quarter of the year and had the biggest one-day market-cap gain of any company ever. Eisman's numbers: earnings per share of $4.74 (up 30%), revenue up 18%, and Azure, the cloud business the whole thesis hangs on, accelerating to 43% growth from 40% the prior quarter. Free cash flow was $19.6 billion, down 23%, but firmly positive, and management notably didn't raise capex. Amazon backed the bull case up: on Squawk, Gil Luria said Jassy gave "the best articulation yet" of why these companies spend, AWS grew 37%, a $25 billion AI revenue run-rate, and "two-to-three-year payback on six-year chips," with 2027 capacity already sold. Luria's demand anchor: OpenAI and Anthropic will together do more than $100 billion of revenue this year, up from roughly zero a couple of years ago, "that is true demand." Why it matters: the mechanical crash and the credit stress are real, but so is accelerating cloud growth at the disciplined spender. This is a genuine two-sided market now, not a one-way unwind.

The debate

Steel-manning both sides of the roughly $800-billion-over-the-next-12-months hyperscaler capex thesis (that's Jefferies' Jason Greenberg's framing on Squawk on the Street, Jul 31, with capex "trending towards a trillion dollars"). This weekend the debate sharpened around one question: was the crash a fundamentals warning, or just a levered fund blowing up?

Bull: the demand is real, the leaders are disciplined, and the crash was mechanical. Microsoft accelerated Azure to 43%, kept ~$19.6 billion of free cash flow, and held capex flat. Amazon showed a 2-3 year payback on 6-year chips with 2027 sold out, and OpenAI plus Anthropic are generating $100 billion-plus of very real revenue that flows straight through the hyperscalers. And the violent sell-off was substantially one over-leveraged fund being force-liquidated, not the market re-rating the fundamentals. An AI-capex analyst on The Rest Is Money, Aug 2 added the bullish data point: AI revenue (adjusted to strip out companies paying each other) hit ~$110 billion over the trailing 12 months, up 3.5x year-over-year and running at a ~$175 billion annualized pace by June, faster than the early internet, cloud or mobile reached the same mark.

Bear: the returns are halving, the financing is going circular, and the plumbing is stressed. Chanos's incremental-return collapse (100% to 25%, some in the teens) is the cleanest bear number of the cycle, and his accounting-mismatch point suggests reported profits are flattered. The credit market is corroborating: CDS blowing out, weak bond demand, and ~$270 billion of AI debt issued by July. The same analyst on The Rest Is Money laid out the sustainability test with numbers: annual depreciation could reach ~$250 billion by 2028, so to fully "pay its own way," depreciation, power, salaries, and a profit, the sector would need roughly $400-450 billion of revenue by 2028, versus ~$185-190 billion expected for full-year 2026. That is a very steep ramp, and the average US business still spends only $11 per employee per month on AI tools. Ben Hunt went furthest on Excess Returns, Aug 2: the US will spend as much on AI capex, inflation-adjusted, as it did fighting World War II ($4-5 trillion), half of 2026 GDP growth is coming from it, and it can no longer be self-funded, so if the "hamster wheel" stops, "we have a full-blown systemic financial system crisis," because the lenders this time are private credit and private equity, not the banks.

The synthesis. Both camps got fed this weekend. The bear case is now mostly about credit and financing, not stock valuation, and it is grounded in real, observable signals (CDS, weak bond books, falling incremental returns). The bull case is that the mechanical seller is largely flushed, the demand is genuine, and the disciplined operator (Microsoft) is compounding. As the Canadian Investor hosts put it, there may well be "a Lehman Brothers moment at some point during this… but we're a lot earlier in this trade than most people think" (The Canadian Investor, Aug 1). The tie-breaker has shifted from "who can spend" to "who can keep financing the spend as the credit window tightens."

Sell signals to watch:

  • A failed or ugly bond/equity raise. Amazon's 1.5x-covered book and BlackRock's 7.53% data-center print are the early tremors. The first hyperscaler or neocloud that can't place paper at a sane price is the crack.
  • Widening CDS on the spenders and the neoclouds. Oracle (~215bps), SpaceX (~185bps) and CoreWeave are the canaries; watch whether the mega-caps' 15-20bps creep keeps going.
  • Incremental returns falling further. Chanos's metric is the one to track: if the giants keep earning less on each new dollar, the boards start asking questions.
  • The 30-year Treasury holding above ~5%. A debt-funded build into a 5%-plus long bond gets more expensive by the week.
  • A second leg of forced selling. Michael Burry reportedly used Thursday's rally to add to bearish bets on Micron, Nvidia and a semiconductor ETF (Smart Investing, Aug 1), a reminder the deleveraging may not be finished.

Stocks in play

  • NVDA. Bull: still the compute bottleneck, and the OpenAI/Anthropic $100 billion-plus of real revenue flowing through the cloud validates the demand. Bear: now the face of the circular-financing worry, its CDS is at new highs, its vendor guarantees are "blowing up the bond market," and Michael Burry reportedly added to his short on Thursday's bounce. Next: its own late-August quarter, the demand read for the entire complex.
  • AVGO. Bull: the "own-the-silicon" custom-ASIC theme keeps riding the group's strength as hyperscalers lean on in-house chips. Bear: no standalone Broadcom signal this weekend, so it trades on the complex's beta, and semis just lived through a violent forced-selling episode. Next: read-through from Amazon's and Google's custom-chip ramps.
  • AMD. Bull: the MI455/Helios story from last week still stands as the challenger case. Bear: rode the semiconductor complex down in the deleveraging with everything else. Next: MI455 production timing and a second anchor customer.
  • MSFT. Bull: the quarter of the year, Azure accelerating to 43%, EPS $4.74 (+30%), ~$19.6 billion free cash flow, capex held flat, and the biggest one-day market-cap gain of any company in history; "the adult in the room" reasserted. Bear: on the call, management "couldn't give an answer" when asked for concrete return-on-invested-capital numbers on its $175-200 billion annual spend (The Canadian Investor, Aug 1). Next: whether the capex-discipline-plus-accelerating-cloud formula holds as spending ramps.
  • GOOGL. Bull: a real, multi-stream business (search, cloud, YouTube, Android) with a genuine hyperscaler moat, the kind of balance sheet that can fund the build internally. Bear: it posted its first-ever negative free cash flow (-$5.9 billion, first since its 2004 IPO), capex of $45 billion in the quarter (2x a year earlier), a 2026 budget of $195-205 billion set to rise "significantly" in 2027, and forward spending commitments up ~$500 billion in three months. Next: the 2027 capex number, and whether core earnings can grow into it.
  • AMZN. Bull: it earned the right to spend, AWS +37%, a $25 billion AI run-rate, a 2-3 year payback on 6-year chips, and 2027 capacity already sold; the stock rose ~15%. Bear: it raised 2026 capex to ~$220 billion (blaming memory costs), and its $25 billion bond drew unusually soft demand, orders falling from ~$62 billion to ~$41 billion after the spread tightened, only ~1.5x covered versus a typical ~4x. Next: whether AWS growth holds as debt service climbs.
  • META. Bull: the ad engine still grew ~28%, and the stock now looks cheap, leaving optionality if Zuckerberg finds discipline. Bear: the flunk of the week, EPS $6.18 vs $7.14 a year ago (a ~13% drop and a miss), expenses up 55%, R&D jumping from $13 billion to $22 billion, free cash flow collapsing to $784 million, and $279 billion of future lease commitments not yet on the balance sheet (up 53% in three months), with, as several hosts noted, no plan to monetize all the compute it's building. Next: an actual monetization plan for the compute, the market wants specifics.

Read-throughs

  • Memory: this is where the crash lived, and where the next air pocket hides. SK Hynix reported a six-fold jump in quarterly profit with margins above 80%, and fell ~19% anyway, because it announced ~$31 billion of capex (up ~50%), which bulls read as demand and bears read as a supplier expanding at the top of the cycle. Micron traded down to ~$7.40 from ~$12.55 on its earnings (The Macro Trading Floor, Jul 31). The MSCI World Semiconductor index fell ~16% in July, its worst month since 2022. And a genuinely useful caution from Smart Investing, Aug 1: those multi-year memory contracts the bull case leans on may not hold if prices crack, because if you locked in at $100 a unit and spot falls to $20, "I'll honor that commitment because I have to, but I will not do business with you after this ends." Most actionable: memory is now the single highest-beta way to express the whole AI trade in either direction, the mechanical unwind may be spent, but the 2028 supply cliff is the fundamental risk under it.
  • Power and grid: still the hard physical constraint, and it's getting more expensive. The AI-capex analyst on The Rest Is Money, Aug 2 put scale on it: a single latest-generation Nvidia rack draws ~100 kilowatts (like running 50 kettles flat out), individual data centers have gone from ~50 megawatts five years ago to 500-750 megawatts up to a gigawatt today, "like a big aluminium factory," there's now a ~24-month backlog to buy gas turbines, the US grid has been "moribund for about 15 years," and the nuclear revival is "wading through molasses." Neil Chadwick sized the demand at up to 300 gigawatts of additional power by 2030, enough for 225 million homes. Most actionable: the power picks-and-shovels keep printing, with Quanta Services (which utilities hire to build capacity) posting EPS of $4.24, up 71%, on revenue of $9.6 billion, and Bloom Energy revenue up 166%.

Deltas on the big numbers: aggregate capex framed at ~$800 billion over the next 12 months, "trending towards a trillion"; Situational Awareness NAV ~$45B to ~$10B, ~4x levered, +1,000% since a July 2024 launch, ~67% down on the month; 1.2 million Korean margin calls (~350k already fully liquidated); Oracle CDS ~215bps (from ~145), SpaceX ~185bps; AI debt issuance ~$270B by early July (~2x 2025), top six ~$194B; Amazon bond orders $62B to $41B; BlackRock data-center bond 7.53%; CoreWeave raising $2.6B at 9%-plus; Alphabet FCF -$5.9B (first since 2004), capex $45B in the quarter, forward commitments +$500B in 3 months; Meta off-balance-sheet leases $279B (+53% in 3 months); SK Hynix -19% (6x profit, 80% margins, +50% capex to ~$31B); Micron ~$12.55 to ~$7.40; MSCI World Semi -16% in July; hyperscaler incremental returns ~100% to ~25%.

Still open: whether the forced selling is truly finished or has another leg (Burry adding shorts); whether the September Fed meeting confirms a hike into a 5%-plus long bond; whether Meta produces an actual monetization plan; whether OpenAI and Anthropic stay healthy enough to hold up Oracle's and the neoclouds' backlogs; and whether the next big AI bond or equity raise clears cleanly.

Next catalysts to watch: Nvidia's late-August quarter (the demand read for everything upstream), the September Fed meeting (live hike risk into a 5%-plus long bond), and the next hyperscaler or neocloud debt/equity raise, since the funding market is now the swing factor.