Newsletter · · Ashutosh Agarwal

Oracle Beats and Backlog Hits 664 Billion Dollars, but the Stock Pop Underwhelms - The AI Capex Tracker - Week of September 11, 2026

The AI Capex Tracker for the week of September 11, 2026: Oracle beat on almost every line and grew its backlog to $664 billion, yet the stock rose only about 7% after hours; a Vaughan Nelson portfolio manager argued the custom-chip boom, not AI itself, is the real bubble; NewEdge Wealth's Cameron Dawson said accounting is flattering Big Tech growth; Micron laid out its plan to break the memory wall; and power returned to the conversation as the 10-year yield closed near 5%.

The AI Capex Tracker

Week of September 11, 2026: Oracle Beats and Backlog Hits 664 Billion Dollars, but the Stock Pop Underwhelms

Issue: Friday, September 11, 2026

(Covers podcasts published Thursday, September 10: the evening Oracle finally reported, the day the 10-year Treasury yield closed within a whisker of 5%, and the day diesel hit $6 a gallon for the first time ever. The verdict everyone had been waiting for came in: Oracle beat on almost every line and its order backlog swelled to $664 billion, and the stock rose about 7%, which several people on the desk thought was a letdown given how much was riding on it. Underneath that headline, three quieter but arguably bigger ideas landed: a portfolio manager made the sharpest chip call of the week, that the custom-chip boom, not AI itself, is the real bubble; a Micron executive laid out the plan to break the "memory wall"; and the power grid came roaring back into the conversation. All of it against a macro backdrop that turned meaningfully more hostile to expensive, long-dated bets.)

TL;DR

  • The Oracle verdict is in, and it was a clean beat that the market only half-cheered. Oracle's cloud-infrastructure revenue more than doubled (up 121%), its order backlog (the contracted work it hasn't delivered yet) climbed to $664 billion from $638 billion the prior quarter, it signed over $30 billion of new contracts, and (critically for the bears) it said it has no new plans to raise capital and has finished its $20 billion stock sale. Even its cash burn came in half as bad as feared (negative $5.4 billion versus the ~$10 billion the street expected). The stock rose ~7% after hours, but as one host put it, on numbers this good "this should have been a 15% pop" (CNBC's "Fast Money," Sept 10; Schwab Network, Sept 10).
  • The freshest idea of the week: the real bubble isn't AI, it's the custom-chip boom. Adam Rich, deputy chief investment officer at Vaughan Nelson, argues that of the roughly $1.1 trillion being spent on AI, a large slice is flooding into 20-plus companies racing to build custom AI chips (ASICs) that mostly aren't earning their cost of capital. His call: hyperscalers will eventually prune that spending, which "makes free cash flow disappear," and the money will rotate into Nvidia and the hyperscalers themselves, because in chips, scale wins, not the best benchmark (Vaughan Nelson, Sept 10).
  • The macro turned harder against the whole trade. The 10-year Treasury yield closed near 5% (its highest since October 2023), the 30-year hit its highest since 2007, diesel touched $6 a gallon for the first time ever, oil ran an eight-day streak to ~$103, and 30-year mortgages crossed 7%. Markets now price a ~70%+ chance the Fed hikes next week, with CPI landing this morning. On top of that, a respected strategist warned that Big Tech's reported earnings are being flattered by accounting: strip out lease reclassifications, stretched payables and one-time securities gains and "50% growth drops to about 30%" (Fast Money, Sept 10; On The Tape with Danny Moses, Sept 10).

What's new

This issue covers Thursday, Sept 10, Oracle-earnings night. The biggest dollar event is Oracle. The most useful new idea is the custom-chip-bubble call. And the loudest background noise (rising rates) is the thing that quietly raises the bar under every AI valuation.

Quick jargon key, since these terms recur below. Capex is capital spending: money laid out on long-lived assets like data centers and chips. Hyperscaler means the giant cloud owners: Microsoft, Google, Amazon, Meta. Free cash flow is the cash left after a company pays for its capex, the money truly left over. RPO / backlog is "remaining performance obligations": contracts a company has signed but not yet delivered, i.e. future revenue it's booked but not earned. ASIC is a custom chip built for one specific job (here, AI), as opposed to Nvidia's general-purpose GPUs. HBM is high-bandwidth memory, the specialized memory stacked next to AI chips. ROIC is return on invested capital: how much profit a company squeezes from each dollar it puts to work.

1. The biggest event: Oracle cleared the bar, and the muted reaction is the tell. CNBC's "Fast Money," Sept 10 (Melissa Lee with Tim Seymour, Karen Finerman, Dan Nathan, Guy Adami, plus Dan Ives of Yorkville) and Schwab Network, Sept 10 (Kevin Green, senior markets correspondent).

For a month this newsletter has flagged Oracle as the single cleanest test of whether the market still believes the AI-infrastructure story. Here's what it delivered. Revenue $19.35 billion (versus $19.13 billion expected). Earnings $1.92 a share (versus $1.75). Operating income $8.15 billion at a 42% margin, both better than expected. Cloud-infrastructure revenue $7.39 billion, up 121%. The backlog jumped to $664 billion from $638 billion, with more than $30 billion of new contracts signed in the quarter. Oracle delivered 850 megawatts of new data-center capacity in the quarter alone. And on the two things the bears cared about most: management said it has "no new plans to raise capital" and has completed its $20 billion equity issuance, while free cash flow (which had been "a disaster") came in at negative $5.4 billion, roughly half as bad as the ~$10 billion of burn the street had penciled in (Fast Money; Schwab Network).

The reaction, though, was the story. The stock popped about 7% after hours, but the options market had priced a 12% move in either direction, and Guy Adami's read was blunt: "if you're on the long side of this, this should have been a 15% pop… the price action to me is disappointing." Dan Nathan pressed the structural worry: this is a company with $167 billion of debt, and "there is no guarantee they are going to be able to fulfill those obligations," especially if its anchor customer, OpenAI, "falters a bit in their demand." Karen Finerman flagged that the market will eventually want to know the capex number (expected around $92 billion this fiscal year) and whether "no new capital raise" survives another quarter.

Dan Ives (Yorkville) supplied the bull counter, and it's the number to hold onto: the market is currently assuming 50-60% of that $664 billion backlog "never happens", i.e. that Oracle can't build the data centers or convert the revenue. He called Oracle "a penalty box stock," said this was "a big step in the right direction," and put the broader AI trade in "the third inning," while noting chip demand-versus-supply in Asia is running 13-to-1. His verdict was more patient than euphoric: "Rome wasn't built in a day… the street's not going to give them credit for just one quarter." Why it matters: Oracle passed its test on the fundamentals but not on sentiment. A 7% pop on a 12%-priced move, into a tape falling for a fourth straight day, says the market is quietly discounting the very backlog Oracle keeps growing, exactly the bear worry that this whole tower rests on future promises.

2. The sharpest chip call of the week: the custom-ASIC boom is the real bubble. Vaughan Nelson, Sept 10, Adam Rich, portfolio manager and deputy chief investment officer, walking through the firm's new research paper "Artificial Intelligence: More Compute, Fewer Architectures."

This was the most differentiated view in the entire sweep, because it refuses the lazy "is AI a bubble, yes or no" frame. Rich's argument: of the roughly $1.1 trillion being spent, "most of it's going into the Nvidia ecosystem… but there's also a pretty big part that's going into this custom ASIC world, and we're talking hundreds of billions." His claim is that "the custom ASIC market is where the bubble is right now." The historical rhyme: in every prior tech cycle "you get one or two platforms that really work" (PC and smartphone), but "right now we've got 20 different companies, U.S. and Chinese, building custom ASICs," and "not all of them are going to work."

Why did hyperscalers start building their own chips? Rich's answer is refreshingly unromantic: foundry scarcity. "TSMC is the only one who is capable of building these chips" after Intel's and Samsung's stumbles, "so everybody wants more Nvidia chips, but… TSMC can't make any more Nvidia chips," and hyperscalers built ASICs as a workaround and a way to claw back pricing power from Nvidia. The problem: no one can get enough TSMC capacity to reach the scale that makes custom silicon economical.

Then the killer detail on who actually captures the money. In the recent wave of deals ("Google partnering with Marvell, Amazon with Qualcomm"), the hyperscalers negotiated warrant structures: they get equity in the chip designers if volume targets are hit. "When you run the math," Rich said, "effectively these hyperscalers are getting custom ASIC chips for basically free if all the targets are hit." That, he argues, tells you exactly where the bargaining power sits: with the buyers, not the ASIC vendors. Meanwhile Nvidia isn't a chip company anymore, it's "an architecture company": GPU plus CPU plus networking plus the CUDA software layer, which keeps getting faster for free as millions of developers optimize it (he cited a company speeding throughput "three to five times… by just doing a software update").

His most persuasive evidence is behavioral: what the smartest spenders actually do. "Elon Musk came out and said they're standardizing on Nvidia's infrastructure" for SpaceX/xAI despite Musk's deep custom-silicon experience, and the neoclouds are building on standardized Nvidia too, because merging multiple chip architectures "doesn't get you cost benefits," it slows you down, "and with AI, speed really matters." The trade that falls out: as hyperscalers realize custom chips underperform, "they're going to start pruning," which frees up "hundreds of billions of free cash flow," lifts their return on capital, and drives a rotation out of the custom-ASIC cohort and into Nvidia and the hyperscalers. Why it matters: this reframes the AI-bubble debate as an intramural one, not "sell all of AI," but "sell the ASIC middlemen, own the platform." It's a direct, actionable challenge to the Broadcom/Marvell custom-silicon bull case.

3. A respected strategist puts a number on the accounting problem, and warns the market has already peaked in quality. On The Tape with Danny Moses, Sept 10, Cameron Dawson, chief investment officer at NewEdge Wealth.

Dawson isn't a permabear (she's spent years constructive on this market), which is what made her caution land. Her core point: Big Tech is playing "certain games with their accounting" that are "a very late-cycle dynamic." The specifics she named: reclassifying operating leases as financing leases to flatter free cash flow; extending payables (paying suppliers later) at the Mag 7; Nvidia's extension of receivables; and one-time gains on non-marketable securities: "we know that's SpaceX and that's Anthropic, at least in the last quarter, with Amazon and Google." Net it all out and "50% growth drops down to about 30% growth when you take those numbers out." Her warning about consensus: analysts expect second-quarter 2027 earnings to be "effectively flat" with 2026, but since a chunk of the 2026 number was a one-time gain, "you have to have a lot of real operating earnings" just to stay flat.

She also traced the mechanism the bulls under-weight. Back in the summer of 2025, analysts expected hyperscaler capex to be flat; now it's likely to grow 90-100% this year, with consensus at 30-40% growth for 2027. That swing is why semiconductor earnings (18% of the S&P 500) have exploded. But it has changed the character of these companies: they used to be "agnostic to the capital cycle" because they were self-funded; now, spending more than they generate, "they absolutely have to rely on outside markets," both debt and equity. That makes them newly sensitive to rates and spreads, and "the market is saying, we'll fund you, but at a bigger price." Her framework line, worth taping to a screen: "Price leads data." The old mantra (if 12-month forward earnings make a new high, so will the market) "is starting to fray," and valuations have already compressed from ~23x to just over 19x. Why it matters: this is the bridge between the accounting and the tape. If reported growth is really 30% not 50%, and the multiple is falling while capex still has to be funded externally at a rising cost of capital, the margin of safety is thinner than the headline numbers suggest.

4. The operator antidote: Micron's plan to break the "memory wall." Embedded Insiders, Sept 10, Jeremy Werner, senior vice president and general manager of Micron's Core Data Center Business Unit (a genuine operator, not a pundit).

Amid a bearish feed, here was a builder talking supply. Werner's framing: with AI, "the bottleneck really shifts from computation to memory": chips have gotten faster far quicker than memory can feed them data, and that widening gap "is what we call the memory wall." Micron's answers: HBM4, which "doubled the bandwidth from the prior generation… and brings down the energy meaningfully," plus a five-layer hierarchy of memory and storage so data can move up and down intelligently and keep "very precious GPU resources" from sitting idle.

On the question every device-maker is asking (when do sky-high memory prices ease?), Werner was candid that it takes time and steel. Micron is building five fabs at once: two 600,000-square-foot fabs in Boise, Idaho; the first of what will eventually be four in Clay, New York; plus extensions in Taiwan, Singapore and Hiroshima, Japan. For scale, one 600,000-sq-ft building is "10 football fields" of cleanroom held to "10 particles per million atoms of air." His honest constraint: "there's just not enough of the supply chain, of the labor force, of the know-how to build these things as fast as we'd like." Why it matters: this is the supply side of the HBM-tightness thesis in the builder's own words: prices stay high while these fabs ramp "over the next five years," a pricing tailwind for Micron and a cost headwind for everything with memory in it (Werner separately flagged consumer devices getting more expensive). It's confirmation, from inside the fab, that the memory shortage is a multi-year story, not a quarter.

5. Power came roaring back, with a real answer on the "data centers spike my electric bill" fight. Catalyst with Shayle Kann, Sept 10 (Shayle Kann, energy-infrastructure investor) and Hub Podcasts / "Alberta Edge," Sept 10 (Nate Glubish, Alberta's Minister of Technology and Innovation).

The grid was back in the conversation, and the framing was unusually clear-eyed. Kann's counterintuitive top line: locally, big data centers now mostly keep electricity rates flat or push them down. Utilities have entered "phase three," where "they hold all the cards" and are increasingly negotiating deals that guarantee (often in the announcement itself) that a new data center "is at a bare minimum not going to increase rates, and oftentimes now going to reduce rates." He cited an EPRI study finding data centers have, on average, lowered customer bills about 6% so far, and noted the first utilities are now going to regulators requesting rate decreases because of large loads. (A Gallup poll he referenced shows the public has nonetheless turned sharply against data centers; the top three objections are water, energy use, and higher bills: a trust problem more than a math problem.)

But the global picture is the opposite, and it's where the read-through lives. Data-center growth is "a demand shock" that has bottlenecked "basically every step in the supply chain," so equipment costs have roughly doubled or more: "conductor is like twice as expensive… transformers are two-plus times more expensive… switchgear is two times more expensive… gas power plants are two to three times more expensive," and renewables cost more to interconnect just as their tax credits begin expiring. On top of that, two-thirds of all U.S. utility spending is just maintenance and hardening, non-negotiable, and now inflating too. Kann's macro worry ties it back to the day's theme: the power system "is a big capital expenditure we make generationally," and "we seem to be moving into a higher-interest-rate environment", the worst possible moment to be building it at this scale.

The Alberta interview supplied a concrete operator case study on the Meta build. Minister Glubish detailed the province's "bring-your-own-generation" and "cost-causation" rules: data centers pay 100% of their own grid-connection cost (transmission, transformers, substations), unique among industries. For Meta's site in Sturgeon County, that meant over $100 million paid up front for a 970-megawatt grid allotment, alongside a plan to build an 1,800-megawatt power plant on site with Kineticor/Pembina. The kicker: Meta's transmission fees will produce "approximately a 6% reduction in every Albertan's transmission costs in perpetuity." The grid operator had ~3,000 MW of spare capacity and allocated 1,200 MW to data centers. Why it matters: this is the cleanest example yet of the "build your own power" model that keeps a hyperscaler's expansion from becoming a local political landmine, and a reminder that Meta's capex isn't just chips, it's gigawatts of self-funded generation. The watch-item Glubish flagged: solving the electricity bottleneck could create a natural-gas-pipeline bottleneck on Alberta's NGTL system.

The debate

Steel-manning both sides of the hyperscaler capex thesis, still running near $700 billion in 2026 and heading toward $1 trillion-plus in 2027-2028. Oracle-night sharpened the argument rather than settling it: the numbers were fine, the reaction wasn't, and the macro got worse.

Bull, the demand is real, the money is being spent rationally, and the market is under-crediting the backlog. Oracle just doubled its cloud-infrastructure revenue, grew its backlog to $664 billion, signed $30 billion of new contracts, and said it doesn't need more capital, yet Dan Ives estimates the market is pricing in that 50-60% of that backlog never converts (Fast Money, Sept 10). Adam Rich's version is subtler and, for once, a bull case on the biggest names: strip out the wasteful custom-ASIC spending and the hyperscalers' free cash flow and return on capital snap back, and Nvidia (as an architecture, not a chip) keeps compounding its lead; the smartest capital allocators (Musk's SpaceX, the neoclouds) are voting for standardized Nvidia with their checkbooks (Vaughan Nelson, Sept 10). And the supply side is genuinely constrained, not glutted: Micron's memory wall is real and its five-fab build-out will take five years to catch demand (Embedded Insiders, Sept 10). Even the grid math is turning bull-friendly at the local level: data centers that build their own power are lowering neighbors' bills (Catalyst with Shayle Kann, Sept 10; Hub Podcasts, Sept 10).

Bear, the reported growth is inflated, the spending now depends on a hostile capital market, and a big chunk of it is being wasted. Cameron Dawson's accounting math is the sharpest bear weapon of the week: strip the lease reclassifications, stretched payables and one-time securities gains and "50% growth drops to about 30%"; the multiple has already compressed from 23x to 19x; and hyperscalers have gone from self-funded to "very sensitive to the capital cycle" right as the cost of capital rises for everyone (On The Tape, Sept 10). Adam Rich (again cutting both ways) says hundreds of billions poured into 20-plus custom-chip projects is dragging down returns and will be pruned, a source of funds that could crater the ASIC cohort (Vaughan Nelson, Sept 10). PIMCO's Anastasia Bujalskaya adds the dot-com caution: the internet was transformative, "but much of it accrued to consumers or to a relatively small number of eventual winners," while "many companies that financed the transition disappeared"; and government subsidies and the "single AI race" framing may simply "extend the cycle, postpone the recognition of losses, and transfer some of those losses from shareholders to taxpayers," a BIS paper she cites warning that debt-financed competition "can produce over-commitment" (PIMCO Pod, Sept 10). And the tape agrees the reaction, not the print, is the signal: Oracle's muted 7% on a 12%-priced move.

The synthesis. Oracle-night crystallized the standoff. Both sides now accept the same facts (backlog is huge, capex is enormous, growth is decelerating and increasingly debt-funded) and fight over whether the glass is 40% full or 60% empty (literally: Ives says the market prices 50-60% of the backlog as vapor). What tipped this cycle wasn't the print; it was everything around it. The 10-year at ~5% raises the discount rate under every long-dated AI bet. Dawson's accounting work chips at the "E" in the P/E. And the freshest structural idea (that the ASIC boom is a bubble inside the boom) means even an AI bull might now want to be short the middlemen and long the platform. The next tie-breaker isn't a podcast; it's this morning's CPI and next week's Fed.

Sell signals to watch:

  • Oracle needing to raise capital again. Management said "no new plans." The moment that changes (a fresh shelf offering, another equity raise), it validates the bear worry that the $664 billion backlog can't be self-funded (Fast Money, Sept 10; Schwab Network, Sept 10).
  • A hot CPI this morning that the market respects. With hike odds already ~70%+ and the 10-year at ~5%, a hot core print lifts the discount rate on every AI multiple (Fast Money, Sept 10).
  • The custom-ASIC names cracking. If hyperscalers start signaling they'll prune custom-silicon spend (or the warrant terms keep tightening), the Broadcom/Marvell/Qualcomm ASIC cohort is the "source of funds" that gets sold first (Vaughan Nelson, Sept 10).
  • Earnings-quality flags multiplying. More lease reclassifications, payable extensions or one-time securities gains propping up free cash flow would confirm Dawson's "late-cycle games" thesis and pressure multiples further (On The Tape, Sept 10).
  • Widening credit spreads / rising CDS on the spenders. Dawson noted Nvidia's credit-default swaps have "certainly jumped"; if hyperscaler spreads keep widening, the "we'll fund you but at a bigger price" dynamic starts to bite the build (On The Tape, Sept 10).

Stocks in play

NVDA. Bull: the week's most striking chip call was, in effect, a Nvidia bull case: Adam Rich frames Nvidia as "an architecture company" (GPU + CPU + networking + CUDA software) whose ecosystem compounds efficiencies "for free," gives "some of the best paybacks in the business," and stands to absorb the capital that rotates out of failing custom-ASIC projects; he notes even Musk's SpaceX is standardizing on Nvidia (Vaughan Nelson, Sept 10), and Dan Ives called it "a safer place to be than companies like AMD or Intel" with Asian chip demand/supply at 13-to-1 (Fast Money, Sept 10). Bear: Cameron Dawson flagged Nvidia's "extension of receivables" as an earnings-quality game and noted its credit-default swaps "have certainly jumped" (On The Tape, Sept 10); the stock closed about 2.3% lower on the day (Schwab Network, Sept 10). Next catalyst: any Vera Rubin roadmap detail (Rubin systems were pitched at 260 terabytes/second of interconnect); this morning's CPI and next week's Fed.

AVGO. Bull: Broadcom sits at the center of the optics scale story: its Tomahawk 6 switch "can move more than 100 trillion bits of information every single second," the kind of throughput that forces the whole industry toward light (The MoneyFlows Show, Sept 10). Bear: it's the poster child of Adam Rich's "custom-ASIC is the bubble" thesis, a big beneficiary of the hundreds of billions he thinks will get pruned, and its Google-Marvell-style warrant deals show the hyperscalers, not the chip designers, capture the economics ("getting custom ASIC chips for basically free if all the targets are hit") (Vaughan Nelson, Sept 10). Next catalyst: Marvell's Analyst Day (Oct 6) as the cleanest custom-ASIC read-across.

AMD. No fresh MI450X/Helios catalyst in the Sept 10 batch, and the only mentions were unflattering: Dan Ives said Nvidia is "a safer place to be than companies like AMD or Intel right now, given where their positioning is" (Fast Money, Sept 10), and Adam Rich's "scale wins, not the best benchmark" argument cuts against every Nvidia challenger (Vaughan Nelson, Sept 10). Next catalyst: second-half data-center step-up and any MI450X deployment color.

MSFT. Bull: it's the self-funded scale case: the software trade got an "all clear" from Dan Ives on the back of Oracle/Adobe/Salesforce, with Microsoft "front and center" as investors rotate back into software (Fast Money, Sept 10); the stock closed higher on a red day (Schwab Network, Sept 10). Bear: Cameron Dawson's capital-cycle warning hits every hyperscaler now spending beyond its cash flow, and her "stretched payables at the Mag 7" flag is a Microsoft-adjacent earnings-quality concern (On The Tape, Sept 10). Next catalyst: next capex commentary and Azure growth.

GOOGL. Bull: Alphabet closed up ~0.6% on a down day, and it's the buyer-with-leverage in the custom-silicon deals: its Marvell partnership comes with warrant economics that, per Adam Rich, hand Google the upside if volumes hit (Vaughan Nelson, Sept 10; Schwab Network, Sept 10). Bear: Cameron Dawson named Google specifically among the companies booking "one-time gains on non-marketable securities" (its Anthropic/SpaceX stakes) that flatter earnings, the "50% becomes 30%" adjustment (On The Tape, Sept 10). Next catalyst: whether free cash flow stabilizes next quarter.

AMZN. Bull: Amazon is the hyperscaler flexing buyer power in custom silicon: its Qualcomm chip deal (which Adam Rich noted "happened this past week") is another warrant-style arrangement where Amazon, not the vendor, captures the economics (Vaughan Nelson, Sept 10). Bear: Amazon closed fractionally lower, and Dawson singled it out for "one-time gains" on its Anthropic and SpaceX stakes propping up reported growth (On The Tape, Sept 10; Schwab Network, Sept 10). Next catalyst: proof AWS demand is broad enough to justify the spend.

META. Bull: the most concrete real-world build detail of the week, Meta's Sturgeon County (Alberta) site: 970 MW of grid access secured, $100 million+ paid for its own connection, and an 1,800 MW on-site power plant under construction, a self-funded model that will actually cut local transmission bills ~6% (Hub Podcasts, Sept 10); Meta also announced its "Muse" AI agent (Schwab Network, Sept 10). Bear: Meta gave back some of the prior day's gains and closed lower, and it's squarely inside Dawson's capital-cycle warning as a company now spending well beyond its cash flow (On The Tape, Sept 10; Schwab Network, Sept 10). Next catalyst: whether ad growth keeps outrunning the capex line.

Also in play: ORCL (the night's main event: beat on revenue/EPS/margins, backlog $664B, no new raise, ~7% pop that underwhelmed vs a 12%-priced move). AAPL (foldable-phone reviews drove a ~3.5% gain, leading the Mag 7, on a $2,000 starting price). MU (Micron's Werner detailed HBM4 and a five-fab build-out; memory stays tight for years). Optics: LITE, CIEN, AAOI (a fresh "oversold photonics" theme; see Read-throughs). (Fast Money, Sept 10; Embedded Insiders, Sept 10; The MoneyFlows Show, Sept 10)

Read-throughs

  • Financing / rates, this is the read-through that matters most, and it points up. The dominant force this cycle wasn't a stock, it was the cost of money: the 10-year closed near 5%, the 30-year at its highest since 2007, a $22 billion bond auction saw the weakest dealer takedown Rick Santelli says he's ever seen, and the Treasury's own buyback operation "disappointed": offer-to-cover of just 2.02 versus a normal ~2.5, which Santelli read as holders refusing to sell (Fast Money, Sept 10). Cameron Dawson's "crowding out" point is the trade: as hyperscalers move from self-funded to market-dependent, they compete for capital with governments, and "the cost of capital is going up for everybody" (On The Tape, Sept 10). Actionable: hedge duration risk alongside semiconductor risk; this morning's CPI and next week's Fed sit directly on top of every AI multiple.
  • Power / thermal, back with force, and the equipment inflation is the durable trade. Shayle Kann's supply-chain read is the one to keep: conductor ~2x, transformers 2x+, switchgear 2x, gas plants 2-3x more expensive, with two-thirds of all utility spend locked into non-negotiable maintenance that's inflating too (Catalyst with Shayle Kann, Sept 10). The pricing-power tailwind runs to the equipment makers closest to the constraint (the Vertiv/Eaton/transformer/switchgear complex), even though no operator named them this cycle. The Alberta case study shows the winning political model: hyperscalers building their own generation (Meta's 1,800 MW plant) so the grid, and ratepayers, come out ahead (Hub Podcasts, Sept 10). Watch the second-order bottleneck Glubish flagged: gas-pipeline capacity.
  • Memory, the shortage is a five-year story, straight from the fab. Micron's Werner confirmed HBM4 doubles bandwidth and that Micron is mid-build on five fabs to chase demand it admits it "was caught off guard" by, constrained by labor and know-how (Embedded Insiders, Sept 10). Actionable: a multi-year pricing tailwind for the memory sellers (Micron, and by extension SK Hynix/Samsung) and a persistent cost headwind for anything with a lot of memory in it; Werner and others noted consumer-device prices rising on memory inflation.
  • Optics / networking, a new "oversold photonics" dislocation worth a look. With Broadcom's Tomahawk 6 at 100 trillion bits/second and Nvidia's Rubin interconnect at 260 TB/s, copper can't carry the load and the money moves to light. Three names were pitched as beaten-down on a forced-seller unwind (tied to the "situational awareness" fund liquidation this newsletter has tracked via the Leopold Aschenbrenner short thread), not a fundamental crack: Lumentum (LITE): revenue +109% to >$1B, operating margin 15%→37%, EPS $0.88→$3.23, with Deutsche Bank ($1,200) and Evercore ($1,100) initiations; Ciena (CIEN): revenue +37% to $1.67B, earnings +215%, co-packaged optics ramping 2027-28, ~62% implied upside to a $520 consensus target; and Applied Optoelectronics (AAOI): revenue +86% to $192M, fifth straight record quarter, 800G/1.6T demand said to exceed capacity through mid-2027 (The MoneyFlows Show, Sept 10). Treat the price targets as promotional, but the setup (record fundamentals, forced-seller drawdown) is the kind of dislocation worth screening.

What changed vs last issue

Last issue (Thursday, Sept 10, "Eisman joins the bears. Oracle's the verdict tonight") covered the Sept 9 batch: Steve Eisman calling an AI bust recessionary, three hosts describing a debt-financed tower on OpenAI/Anthropic, Matt Klein's "AI could push rates up" thesis, a hot Fed backdrop, and Apple's $13B-vs-$700B pitch, with Oracle teed up as the coming verdict. This edition covers Sept 10, and the marquee cliffhanger resolved.

  • Oracle reported, and the verdict was "fine, not thrilling." Last issue: Oracle was the unknown, options implying an ~11-12% move. New this cycle: a clean beat (revenue $19.35B, EPS $1.92, 42% margin), backlog up to $664B from $638B, $30B of new contracts, no new capital raise, and cash burn half as bad as feared, yet only a ~7% pop that the desk called disappointing. Dan Ives says the market still prices 50-60% of the backlog as never happening (Fast Money, Sept 10; Schwab Network, Sept 10).
  • A new, more surgical bear (and bull) idea: the custom-ASIC bubble. Last issue's bear case was macro and holistic: "the whole tower is fragile." New this cycle, Adam Rich narrows it: AI broadly is fine, but the custom-chip cohort (20+ ASIC makers) is the mispriced piece, and the trade is to rotate into Nvidia and the hyperscalers as that spend gets pruned (Vaughan Nelson, Sept 10). This is the first constructive-on-the-mega-caps argument in several issues.
  • The accounting critique got a hard number. Last issue flagged evaporating cash flow qualitatively (Eisman). New this cycle, Cameron Dawson quantifies it: strip lease reclassifications, stretched payables and one-time securities gains and reported ~50% growth is really ~30%, with the multiple already down from 23x to 19x (On The Tape, Sept 10).
  • Rates went from "might rise" to "rising, hard." Last issue: Klein's theory that AI could lift rates. New this cycle, the 10-year closed near 5% (highest since Oct 2023), the 30-year hit a post-2007 high, diesel touched $6/gal for the first time (some California stations near $9.99), WTI ran to ~$103, 30-year mortgages crossed 7%, and Fed-hike odds firmed to ~70%+ (Fast Money, Sept 10).
  • Power returned to the conversation, with data, not just anecdote. New this cycle: Shayle Kann's equipment-inflation numbers (conductor/transformers/switchgear ~2x, gas plants 2-3x) and the EPRI finding that data centers have lowered local bills ~6% so far, plus a concrete Meta/Alberta build case (Catalyst with Shayle Kann, Sept 10; Hub Podcasts, Sept 10).
  • Memory and optics returned as concrete themes. New this cycle, Micron's Werner gave first-person HBM4-and-five-fabs detail (Embedded Insiders, Sept 10), and a fresh "oversold photonics" theme (Lumentum/Ciena/AAOI) surfaced, tied to the same "situational awareness" forced-seller unwind we've tracked (The MoneyFlows Show, Sept 10).
  • Still open: whether Oracle avoids another capital raise; whether the custom-ASIC pruning Adam Rich predicts actually starts (and cracks Broadcom/Marvell); whether the 10-year breaks decisively through 5%; and the unresolved core question: is this offensive, ROI-positive spending, or debt-funded over-build whose reported growth is one-third accounting?

Next catalysts to watch: CPI this morning (Fri, Sept 11), the single biggest near-term print; the September FOMC (~Sept 16-17) and whether a hike (25 or 50bp) lands; the S&P index rebalance Sept 21 (Bloom Energy added); Marvell's Analyst Day, Oct 6 as the custom-ASIC read; Oracle's next-quarter capex update and any capital-raise signal; and OpenAI's / Anthropic's IPO filings whenever they drop.