Newsletter · · Ashutosh Agarwal

Nvidia Calls for 1.3 Trillion in AI Spending While Broadcom Beats and Falls - The AI Capex Tracker - Week of September 4, 2026

The AI Capex Tracker for the week of September 4, 2026, covering August 24 to September 3: Nvidia grew 106 percent and lifted the industry's 2027 spending forecast to $1.3 trillion, Broadcom's 221 percent AI growth still sent the stock down, and the bear case moved from valuation to debt financing and the power grid.

The AI Capex Tracker

Week of September 4, 2026: Nvidia Calls for 1.3 Trillion in AI Spending While Broadcom Beats and Falls

TL;DR

  • Nvidia didn't just beat, it moved the goalposts. Revenue grew 106% to $96.2 billion, and CEO Jensen Huang told the world the five biggest cloud companies will spend $1.3 trillion on AI next year, up from roughly $800 billion this year. Nvidia guided to ~70% growth next year when Wall Street expected 45%, and said the only thing holding it back is that it can physically make just 70% of what customers want. The catch: profit margins are slipping (from ~75% toward 71-73%) because memory chips have gotten so expensive. (Motley Fool Hidden Gems Investing, Aug 27; Tech Brew Ride Home, Aug 27)
  • Then Broadcom proved the market's mood has changed: a blowout is no longer enough. AI chip sales jumped 221% to $16.7 billion and the CEO teased a path toward $230 billion by 2028, and the stock fell about 3%, now down ~26% from its June high. The problem wasn't the quarter; it was that the near-term guide was merely good, and Google is quietly hiring cheaper backup chip suppliers (MediaTek, Marvell, AMD). (The Rundown, Sept 3; CNBC's "Fast Money", Sept 2)
  • The bears stopped arguing about price and started arguing about plumbing, the financing kind and the electrical kind. Venture investor Paul Kedrosky says more than half of all data-center spending is now paid for with borrowed money instead of company cash flow, a "Minsky moment," while power keeps emerging as the real ceiling: America needs about 70 gigawatts of new juice for AI by 2028 and can only plug in 30. (The Meb Faber Show, Aug 31; Squawk Pod, Aug 26)

What's new

A note on the window: this is a catch-up edition covering August 24 through September 3, the stretch since our last issue, which happens to bundle the two biggest earnings prints of the season, Nvidia (Aug 26) and Broadcom (Sept 2), into one recap. I've ranked everything by where the dollars land hardest for a book. Throughout, I've separated what companies actually disclosed (via the shows relaying their earnings calls) from what commentators think about it.

1. Nvidia's print was a demand bomb, and it raised the whole industry's spending forecast by half a trillion dollars. Motley Fool Hidden Gems Investing, Aug 27, Matt Frankel, Jon Quast, Tyler Crowe (Motley Fool analysts); Tech Brew Ride Home, Aug 27, Brian McCullough (relaying the earnings call and Bloomberg); Saxo Market Call, Aug 27, John Hardy and Ruben Dalfovo (Saxo Bank).

The numbers, straight from the report: revenue up 106% year over year to $96.22 billion, net income up 126% to nearly $60 billion, and data-center revenue up 117%. Nvidia then guided next fiscal year (which mostly covers calendar 2027) to about 70% growth, against a Wall Street expectation of 45%, and put next quarter at roughly $108 billion. CFO Colette Kress said the guide is limited by supply, not demand: "we are seeing demand acceleration even at our scale... customers' forecasts point to our growth doubling next year." The company said it can currently satisfy only about 70% of the demand it sees.

The single most important line for the whole sector was Jensen Huang's spending math. He framed the top five cloud builders (Microsoft, Google, Amazon, Meta, plus Oracle) as going from ~$500 billion of capital spending in 2025, to about $800 billion this year, to $1.3 trillion in 2027, a 60% jump. That $1.3 trillion is a real escalation from the "$1 trillion for 2027" number that had been the working assumption. And there's a sober counterweight the Motley Fool crew flagged: those same five companies generated almost $700 billion of operating cash flow over the past four quarters, so the spending, while enormous, isn't coming out of nowhere.

Now the wrinkle that connects to everything below, margins. Nvidia warned gross margin (the share of each sales dollar it keeps) will slip from ~75% to a low of 71-72% early next year before settling at 72-73%, specifically because memory has gotten so expensive. It's only raising its own prices about 15% to offset it. As the Motley Fool put it, you're "trading about three percentage points of gross margin in exchange for roughly doubling your revenue growth expectations." The stock rose about 7%, adding roughly $362 billion in market value in a day.

"The AI infrastructure build-out is at full steam. Vera Rubin, now in full production, was built to power exactly this moment." Jensen Huang, relayed on Tech Brew Ride Home, Aug 27

Vera Rubin, Nvidia's next-generation chip platform, is in full production with purchase orders "from every major hyperscaler," and Nvidia expects it to be the fastest product ramp in its history. Amazon alone is committing to deploy an additional 2 million Nvidia GPUs through mid-2029. Why it matters: this is the referee for the entire trade, and it came back saying demand is not the problem, supply and, increasingly, memory cost are.

2. Broadcom crushed its quarter and the stock fell anyway, the clearest sign yet that "beat" is now priced in. The Rundown, Sept 3, Zaid Admani (Public.com); CNBC's "Fast Money", Sept 2, Christina Partsinevelos and the desk; Bloomberg Intelligence, Sept 3, Matthew Bloxham (Bloomberg Intelligence senior tech analyst).

Broadcom's numbers were, in The Rundown's words, "absolutely insane": total revenue up 86% to $29.6 billion, AI chip revenue up 221% to $16.7 billion (beating its own forecast), profits more than tripled, and management guided next-quarter AI chip revenue to $21.7 billion. CEO Hock Tan's long-range framing pointed to roughly $115 billion of AI chip revenue next year, potentially doubling to ~$230 billion by 2028.

And the stock fell 3%. Here's why, and it's the actionable part. On the number investors cared about most, the fiscal-2027 AI forecast in the actual release, nothing new was added; management just repeated "north of $100 billion" when the Street wanted to hear "at least $115 billion." The next-quarter total revenue guide ($34.8 billion) came in below the ~$37 billion whisper, and Broadcom is now down ~26% from its June high. Underneath it all sits a competition worry: Google is bringing in cheaper backup suppliers for its custom chips, MediaTek, Marvell, and AMD, chipping at Broadcom's turf (Morgan Stanley thinks Broadcom keeps ~80% of Google's business, but the uncertainty is the point).

The bullish read came from Bloomberg Intelligence's Matthew Bloxham: Broadcom now expects Anthropic and OpenAI to become its most important customers, eclipsing Google, evidence, he argued, that "there's substance behind the plans and capital raising these companies are undertaking." Hock Tan said Broadcom has "six customers, and four of them are just simply going to be huge." Why it matters: same company, same circular-financing setup as Nvidia (Broadcom is helping fund Anthropic's data center and its custom chips), but when Nvidia does it the market shrugs, and when Broadcom does it the market frowns. The tell is that the bar for the "shovel-sellers" has quietly risen to blowout-or-bust.

3. The bear case moved from the P&L to the plumbing: more than half of AI's build-out is now debt-financed. The Meb Faber Show, Aug 31, Paul Kedrosky (venture investor); The Grant Williams Podcast, Aug 24, Gerard Minack (Minack Advisors); RenMac, Aug 28, the RenMac desk.

Kedrosky delivered the sharpest new frame of the period. For years the bull rebuttal was "relax, this is all paid out of company cash flow." As of the first half of 2026, he says, that flipped: more than half of data-center financing now comes from outside money, asset-backed securities, private credit, sovereign funds, and off-balance-sheet vehicles, which he calls a classic "Minsky moment," the point where the funding becomes "increasingly divorced from what's actually going on inside the data center." His most eye-catching detail: in one big GPU rental warehouse, chips are running at only 35-40% utilization even as everyone insists there's a shortage, a sign of hoarding and double-ordering. And he flagged a JP Morgan estimate that 15-18% of the entire investment-grade bond market is now data-center related. His deeper worry is that AI's product itself is deflating ~70-80% a year, so a frontier model company needs ~400% growth "just to stand still" while carrying new debt.

Minack's version is the cleanest one-liner of the cycle: this is an "earnings bubble, not a valuation bubble." When a hyperscaler buys chips, the chipmaker books the profit immediately, but the buyer only expenses the cost slowly over years, so profits look great today while the cash quietly drains. "Earnings are an opinion. Cash is a fact." He cited a Goldman Sachs figure of ~$1.5 trillion in hyperscaler lease commitments, of which ~$1 trillion isn't even on the balance sheet yet (leases don't count until the building opens), plus another ~$1 trillion of unrecognized commitments per Morgan Stanley. RenMac's desk put it bluntly, "a disaster waiting to happen", noting Google just did the largest Canadian bond issuance in history because it had "sold all it can" in US dollars, and that company earnings and cash flow "started diverging" the moment the AI build-out went on the card. Why it matters: the debate has migrated from "is it too expensive?" to "does the borrowed money keep flowing?", which makes the credit market, not just earnings, the thing to watch.

4. Power is now openly described as the binding constraint, and the physical grid is showing stress. Squawk Pod, Aug 26, the CEO of Emerald AI (a Nvidia-backed grid-flexibility startup, an operator with a book to talk, so weight accordingly); Factor This, Aug 28, the "This Week in Cleantech" panel; Alpha Exchange, Sept 1, Ulrike Hoffmann-Burchardi (Chief Investment Officer, UBS).

The Emerald AI chief put the gap in one sentence: "Between now and 2028, the demand for AI in America is 70 gigawatts, and we can only plug in 30 gigawatts of data centers onto the grid." Meanwhile "China just built over 500 gigawatts of new power capacity last year, 10 times what we did." His pitch is software that makes data centers flexible so utilities can connect them faster; his blunt macro point stands regardless of the pitch: "I don't think chips are going to be the bottleneck between now and 2030. It's power." UBS's CIO independently put the shortfall at ~100 gigawatts by 2030, and made the key point for anyone worried about rates choking the build: hyperscalers are already "paying three times the salary of an electrician" and "twice the base rate for electricity" to get data centers up, so the sensitivity of AI spending to interest rates is "very, very low", it would take far more than a 50-basis-point move to slow it.

The Factor This panel supplied the gritty read-through. The US is now building twice as much gas-fired power capacity as China; gas under construction jumped 76% in the first half to 378 gigawatts (a ~$647 billion pipeline), and about half of it is tied directly to data centers via one-to-one power deals. The strain is real: three of the four operational off-grid US data centers have already reported problems (cracked turbines at xAI's Colossus site in Memphis; a critical power failure at a Vantage campus in Virginia that forced a day on diesel backup), and PJM, the grid covering 13 states, recently had 4 gigawatts drop offline in seconds. For scale, Amazon's planned Pecos County, Texas site alone needs 7.65 gigawatts, the output of several nuclear plants, and Anthropic pays $1.25 billion a month for compute on Colossus. Why it matters: power availability, not silicon, increasingly governs how fast all that capex actually turns into revenue.

5. Even the referee's boss weighed in: the new Fed chair called AI "a hinge point in history." Squawk on the Street, Aug 28, coverage of Fed Chair Kevin Warsh's Jackson Hole speech, plus Dan Niles (Niles Investment Management).

New Fed Chair Kevin Warsh used his first Jackson Hole keynote to single out AI capex as a driver of a "remarkably resilient" economy and called AI a potential "new factor of production." He was also clearly hawkish on inflation, the odds of a September rate hike jumped to ~57% after the speech (and ~86% by December). That combination, a Fed that is both worried about inflation and impressed by the AI build, is exactly the crosswind the sector now trades in: higher-for-longer rates against a spending wave the same speech praised. Why it matters: it puts a central-bank stamp on the "AI is real" side of the debate, while raising the cost-of-capital risk the bears keep pointing at.

The debate

Steel-manning both sides of the hyperscaler capex thesis (now roughly $800 billion in 2026, heading to ~$1.3 trillion in 2027 on Nvidia's own math). Last issue the whole thing hinged on Wednesday's Nvidia print. It landed, bullishly, and yet the rest of the complex sold off on good news. So the question has sharpened into: does the demand (now proven in Nvidia's numbers) keep showing up faster than the financing strain and the power wall show up?

Bull, the demand is not just real, it's under-supplied, and the biggest spenders can afford it. Nvidia grew 106%, guided to 70% against a 45% expectation, and said it can only build 70% of what customers want, you don't get that combination in a demand bubble. The five biggest buyers threw off ~$700 billion of operating cash flow over the past year, so the base spending is cash-funded even if the incremental dollar is borrowed (Motley Fool, Aug 27). Investor Imran Khan argued the "circular" deals are actually a moat: by locking up the supply chain, Nvidia increases the durability and visibility of its own revenue, "we know AI is real, and we do know AI will generate positive ROI" (RiskReversal Pod, Aug 28). UBS's CIO made the subtle but crucial point that circular financing and rich valuations are not catalysts on their own: "It's not a risk in my mind as a catalyst. The catalyst will be if demand disappoints. So far, we don't see signs of that" (Alpha Exchange, Sept 1). And macro strategist Jon Turek explained why the spend is so hard to stop: hyperscalers treat it as existential, "if we don't do it, we could be out of business", which makes it remarkably insensitive to a 50-basis-point move in Google's bond yield (Macro Hive, Aug 28).

Bear, the money is now borrowed, the returns are unproven, and busts don't happen gently. Kedrosky's "half the financing is now external" and Minack's "earnings bubble, not a valuation bubble" are the two frames to internalize (The Meb Faber Show, Aug 31; The Grant Williams Podcast, Aug 24). Add the hidden liabilities, ~$1.5 trillion of hyperscaler leases with ~$1 trillion off the books, plus ~$1 trillion more unrecognized, and Dan Niles' warning that Nvidia's 70% annual guide is inflated by double-ordering, exactly like Cisco in 2000, "the Nvidia of its day," which promised 30-50% sustainable growth and then printed two straight years of negative revenue (Squawk on the Street, Aug 28). Even Turek, the near-term bull, concedes the tail: "You don't usually slow from a trillion dollars of spend to 800 billion. It usually goes from a trillion to 200" (Macro Hive, Aug 28). And RenMac points at the credit market already flashing: bond investors in these names have been hurting even as equity holders win, and spreads are "widening... getting worse" (RenMac, Aug 28).

The synthesis. Nvidia's print settled the near-term demand question in the bulls' favor, that's why even the loudest bears here (Niles, Kedrosky, Minack) own or concede the technology is real and useful. The fight now is entirely about the tail: whether the borrowed money keeps flowing, whether the grid can deliver the electrons, and whether returns show up before the debt comes due. The most important shift this period is that the market has started to act bearish (Broadcom and Marvell beating and falling) even while the fundamentals print bullish, the classic late-cycle divergence Minack noted from the 2000 playbook, when the AI names "kept motoring" right up until they didn't.

Sell signals to watch:

  • Credit spreads widening on data-center paper. RenMac is already short credit and says the spread between safe and risky corporate bonds is "widening... getting worse." A repriced compute-backed bond, or a private-credit fund marking down its AI loans, is the first crack (RenMac, Aug 28).
  • Nvidia's margin sliding below the 71-73% floor it just set. Management drew a line; a break below it would confirm the memory squeeze is eating the GPU maker faster than it can raise prices (Saxo Market Call, Aug 27).
  • A demand disappointment, the only catalyst that matters, per UBS. Not valuation, not circular financing. Watch for a single hyperscaler cutting its 2027 guide, or a frontier lab missing its revenue ramp (Alpha Exchange, Sept 1).
  • A power failure that costs real money. Off-grid sites are already cracking turbines and running on diesel; a multi-day outage at a flagship campus would put a number on the execution risk (Factor This, Aug 28).
  • A hyperscaler rewarded for cutting capex. Dan Niles expects a market wobble into the November midterms, when "data centers are bad" is the rare bipartisan agreement, even deep-red Texas stalled new builds (Squawk on the Street, Aug 28).

Stocks in play

NVDA. Bull: a genuine blowout, 106% growth, a 70% forward guide against a 45% expectation, only able to meet 70% of demand, and Vera Rubin ramping as the fastest product cycle in its history with orders from every major hyperscaler (Motley Fool, Aug 27). Bear: margins are now falling for the first time in ages (toward 71-73%) as memory costs bite; management stayed muted on the ~$500 billion financing web and its ~$105 billion Ohio data center leased to OpenAI; and Dan Niles thinks the annual guide is padded with double-orders, Cisco-style (Saxo, Aug 27; Squawk on the Street, Aug 28). Next catalyst: Vera Rubin volume ramp and any detail on the financing backstops; Jim Cramer is publicly agitating for a half-trillion-dollar buyback (CNBC's "Fast Money", Sept 2).

AVGO. Bull: AI chip revenue up 221% to $16.7 billion, next quarter guided to $21.7 billion, and Anthropic and OpenAI set to eclipse Google as its top customers, "six customers, four just simply going to be huge" (Bloomberg Intelligence, Sept 3). Bear: it beat and fell ~3% (now -26% from June highs) because the near-term guide was merely good and Google is second-sourcing to MediaTek, Marvell, and AMD; the desk's own take was "there's probably better places to play" (CNBC's "Fast Money", Sept 2). Next catalyst: whether the fiscal-2027 AI number gets formally raised above "$100 billion" toward the Street's $115 billion.

AMD. Bull: named repeatedly as a beneficiary of hyperscalers deliberately diversifying away from single suppliers, one of the "cheaper second sources" Google is bringing in for custom-chip work alongside MediaTek and Marvell (CNBC's "Fast Money", Sept 2). Bear: no fresh MI450X/Helios catalyst surfaced this period, and it remains squeezed between Nvidia's dominance and the hyperscalers' own in-house silicon. Largely quiet on genuinely new product news. Next catalyst: its second-half data-center step-up; read-through from Nvidia's supply-constrained demand.

MSFT. Bull: the cleanest capex-to-cash story, Dan Niles notes the cloud group (Microsoft, Google, Amazon) saw revenue growth accelerate from 35% to 43% quarter-on-quarter with operating margins expanding ~2 points, and Microsoft is now so confident it's restructuring to spotlight AI, disclosing Azure revenue for the first time: $29.4 billion last quarter, over $100 billion for the year, +42% growth, with 30 million-plus paid Copilot subscribers (Squawk on the Street, Aug 28; The Rundown, Sept 3). Bear: despite all that, the stock is up just 4% this year, the market wants proof the AI cloud keeps converting to cash faster than capex climbs. Next catalyst: the first quarter reported under the new "Agents and Infra" segment, which finally makes the Azure number a headline print.

GOOGL. Bull: a legal win, a federal judge ruled Google does not have to break up its ad-tech business (the disputed exchange is only ~8% of revenue and <1% of profit anyway), removing an overhang, and rising capex directly feeds its suppliers (The Rundown, Sept 3). Bear: it went so deep into AI spending that it did the largest Canadian bond issuance in history after exhausting its US-dollar borrowing capacity, a vivid marker that even the cash kings are now tapping debt, and the stock is up only 5% this year with lingering questions about whether its models are keeping pace (RenMac, Aug 28). Next catalyst: a genuine model or product splash (Gemini 3.8 Flash landed to lukewarm reviews) and evidence cloud revenue is turning into free cash flow.

AMZN. Bull: AWS is the standout demand signal from Nvidia's call, committing to deploy an additional 2 million Nvidia GPUs through mid-2029 and adopting Nvidia's full stack (Nemotron models on Bedrock/SageMaker, plus its robotics software for warehouse fleets); the cloud group's growth accelerated to 43% (Rob Black Show, Aug 27; Squawk on the Street, Aug 28). Bear: that 2-million-GPU commitment is also ~$20 billion-plus of spend concentrated in one relationship, and Amazon sits inside the same off-balance-sheet lease critique as its peers. Otherwise quiet on Amazon-specific developments. Next catalyst: proof AWS AI demand is broad and durable enough to justify the GPU commitment.

META. Bull: a genuinely favorable legal turn, it settled the teen-mental-health case with state attorneys general on terms one host called "an incredibly favorable deal" (a ~$18 million outcome and relatively benign product changes, versus the ~$200 billion "big tobacco" scenario that loomed over last issue), and its ad business remains the best real-world example of AI already paying off (Rob Black Show, Aug 27; RiskReversal Pod, Aug 28). Bear: it was the one hyperscaler that pointedly declined to give a 2027 capex guide, and it remains the poster child for AI spending eating the very free cash flow that funds it (Macro Hive, Aug 28). Next catalyst: whether ad-revenue growth keeps outrunning the capex line, and if/when it finally commits to a 2027 spending number.

Read-throughs

  • Memory (Micron, SK Hynix, Samsung), still the tightest link, but the easy money in the stocks may be over. Nvidia's own margin guide-down is the loudest possible confirmation that memory is scarce and expensive, and Dan Niles calls memory "arguably the most tight of all the sectors," with SK Hynix forecasting shortages into 2030 (Squawk on the Street, Aug 28). The nuance to hold, from investor Imran Khan: the memory story has gone "from 10 to 8", the explosive margin-expansion phase has largely played out, and investor questions have shifted almost entirely to share buybacks (SK Hynix just launched a ~$25 billion program), meaning from here the gains have to come from revenue and multiple, not margins (RiskReversal Pod, Aug 28). Gerard Minack frames the extreme: the three makers are forecast to earn ~$550 billion over the next year, yet trade at a forward P/E near 4, the market is screaming these earnings won't last (The Grant Williams Podcast, Aug 24).
  • Networking / custom silicon (Marvell, Broadcom, Credo), the "beat and fall" zone. Marvell posted 37% growth and raised its next-year guide from 45% to 50% (~$18 billion), with its data-center business growing over 60% and a Google warrant deal worth up to $12 billion of stock, and the stock still fell 8% on a "meet, not beat," now up ~143% year-to-date into a high bar (Motley Fool, Aug 28; Squawk on the Street, Aug 28). The actionable read: with GPU interconnect volumes set to explode, the Motley Fool flagged smaller interconnect names like Credo (CRDO) as leveraged plays, but the whole group now needs a blowout, not a beat, to rally, and Marvell's Analyst Day on October 6 is the next real catalyst.
  • Power / thermal / gas (VRT, ETN, turbines, behind-the-meter), the bottleneck trade got louder and more concrete. The numbers are stark: 70 GW of AI power demand by 2028 against 30 GW of grid connection capacity, a ~100 GW national shortfall by 2030, and 378 GW of gas under construction, with the twist that off-grid sites are already failing (cracked turbines, diesel days) (Squawk Pod, Aug 26; Factor This, Aug 28). Actionable: this supports anything that shortens the path to reliable electrons, grid gear (Vertiv, Eaton), gas turbines, and grid-flexibility software, while flagging power as the single biggest execution risk on capex-to-revenue conversion. Note the coverage gap: no named commentary this period on ERCOT/SB6 specifically, or on the nuclear names (Vistra, Constellation, Talen), that's a hole in the sweep, not a signal.
  • Financing / private credit, this is where the first crack will show. With more than half of the build-out now debt-financed, ~15-18% of the investment-grade bond market data-center-related, and Google reaching to Canada to issue bonds, the systemic exposure sits in credit, not equities (The Meb Faber Show, Aug 31; RenMac, Aug 28). Watch spreads on compute-backed paper and any private-credit mark-down before it reaches a hyperscaler's headline earnings. One stabilizer to respect: Nvidia's ~$500 billion financing fund (backed by every major Wall Street firm) plus Treasury's bigger bond buybacks are the "main actors" deliberately smoothing financing conditions, which, per Jon Turek, makes the near-term spend more durable even as it raises the risk of higher-for-longer rates (Macro Hive, Aug 28).

What changed vs last issue

Last issue (Aug 24, "Memory's sold out to 2028. Nvidia reports Wednesday.") teed up the Nvidia print as the referee for the whole trade and framed memory as sold out and the bear case as a credit story. This catch-up edition (spanning Aug 24–Sept 3) has the verdicts:

  • The Nvidia referee ruled, bullishly, and raised the spending bar. New: 106% growth, a 70% forward guide (vs 45% expected), only 70% of demand met, and Jensen's capex forecast lifted to ~$800 billion for 2026 and $1.3 trillion for 2027 (from last issue's ~$700 billion / ~$1 trillion working numbers). The margin worry we flagged last issue materialized, but modestly: guidance to 71-73%, not a collapse.
  • Broadcom's Sept 2 print (which last issue was watching) beat big and fell. New: AI revenue +221% to $16.7 billion, a path teased to ~$230 billion by 2028, and a ~3% drop, now -26% from June highs, on an unremarkable near-term guide and Google second-sourcing. The market's reaction function has shifted: beats are no longer enough.
  • Memory reframed from "sold out" to "the margin trade has played out." New nuance vs last issue's pure sell-out story: Imran Khan's "10 to 8" framing, the pivot to buyback-driven theses (SK Hynix ~$25 billion), and Minack's forward P/E-of-4 warning that the market expects these record earnings to roll over.
  • The credit-cycle bear case got harder numbers. New: Kedrosky's "more than half of financing is now external / Minsky moment," a JP Morgan estimate that 15-18% of the IG bond market is data-center-related, GPU utilization at only 35-40%, and Google's record Canadian bond deal. Minack's "$1.5 trillion of leases, $1 trillion off the books, plus ~$1 trillion more unrecognized" put a figure on the hidden liabilities.
  • Power moved from "quiet" to a quantified, stressed bottleneck. New: 70 GW demand vs 30 GW connectable by 2028; ~100 GW national shortfall by 2030; 378 GW of gas under construction; off-grid failures (cracked turbines, a diesel day in Virginia); Anthropic's $1.25 billion/month compute bill; Amazon's 7.65 GW Pecos County site.
  • New single-company shifts: Google won its DOJ ad-tech case (no breakup), a positive delta; Microsoft will now disclose Azure revenue ($29.4 billion/quarter, $100 billion-plus/year, +42%) under a new "Agents and Infra" segment; Nvidia is buying Hugging Face for $12.9 billion; and Meta got a favorable attorney-general settlement (~$18 million), largely clearing the legal overhang that dominated last issue's Meta entry.
  • Macro backdrop: New Fed Chair Warsh's Jackson Hole debut was hawkish (September hike odds ~57%, December ~86%) while praising AI as "a hinge point in history"; the 10-year sits ~4.79% and the 30-year ~5.18% (roughly flat to last issue); US government debt crossed $40 trillion.
  • Deltas on the big numbers: 2026 hyperscaler capex ~$700B → ~$800B; 2027 ~$1T → $1.3T; top-five operating cash flow ~$700B/4 quarters (new); cloud backlog >$2 trillion (new); Nvidia gross margin ~75% → 71-73% guided; memory forward P/E ~4 (Minack); US power shortfall ~100 GW by 2030.
  • Quiet vs last issue: no named ERCOT/SB6 or nuclear (Vistra/Constellation/Talen) commentary; no fresh AMD product catalyst; no direct Vertiv/Eaton operator color (power came via a startup CEO and a cleantech panel, not the equipment makers themselves).
  • Still open: whether the borrowed money keeps flowing; whether the grid delivers the electrons fast enough; whether the memory sell-out proves the bull case or simply taxes the GPU makers' margins; and whether the "beat and fall" pattern in Broadcom and Marvell is early distribution or just a crowded-trade reset.

Next catalysts to watch: the August jobs report (due today/tomorrow), which could decide the September Fed hike; Marvell's Analyst Day, October 6, for clarity on the Google custom-chip split; the September FOMC; and any repriced compute-linked bond, private-credit mark, or off-grid power failure as the first sign the machine is straining.