Newsletter · · Ashutosh Agarwal
Qualcomm Lands Amazon and Nvidia's Grip Cracks Again - The AI Capex Tracker - Week of September 9, 2026
The AI Capex Tracker for the week of September 9, 2026: Qualcomm signed Amazon as a custom-chip customer against a 60 billion dollar ceiling, a specialist put clean unit economics on AI demand (Anthropic earns more than 50 million dollars per gigawatt against 25 million in cost), and the bears confirmed that hyperscaler free cash flow has already gone negative, all into a market bracing for a possible Fed hike.
The AI Capex Tracker
Week of September 9, 2026: Qualcomm Lands Amazon and Nvidia's Grip Cracks Again
Issue: Wednesday, September 9, 2026
(Covers podcasts published Monday–Tuesday, September 8, the first full trading day back after Labor Day. Fourteen episodes cleared the filter, all dated Sept 8. The theme of the day: a new name muscled into the custom-chip business Nvidia and Broadcom have had mostly to themselves, one specialist finally put clean unit economics on the demand side, and the bears kept sharpening the same knife, free cash flow. All against a market suddenly bracing for a Fed that might hike, not cut, next week.)
TL;DR
- Qualcomm just signed Amazon as a custom AI-chip customer, a deal with a $60 billion purchase ceiling through 2030, plus warrants letting Amazon buy up to $4 billion of Qualcomm stock. It validates Qualcomm as a real fourth entrant against Nvidia, AMD and Broadcom, and it is one more sign that Amazon (and everyone else) is racing to diversify away from paying Nvidia's prices. Qualcomm shares jumped ~4-5%. (Bloomberg Tech, Sept 8; The Financial Exchange, Sept 8)
- The cleanest bull number of the week: Anthropic is willing to pay about $25 million per gigawatt of compute because it generates more than $50 million per gigawatt, a roughly 2-to-1 return that says demand is real and constrained, not speculative. Gil Luria of DA Davidson argued the local pushback and "bubble talk" are actually managing the cycle, keeping Jensen Huang's ~70% growth guide well below the demand that "would justify far higher." (Bloomberg Surveillance, Sept 8)
- The bear case moved from a countdown to a fait accompli: S&P Global's team said Google has already posted its first-ever negative free cash flow and is projected to stay cash-flow-negative for eight straight quarters, a "stunning reversal" for a company that used to throw off ~$70 billion a year. And a separate teardown argued that once you adjust hyperscaler capex for how fast GPUs go obsolete, the buildout is running near 8% of US GDP, higher than the 1870s railroad mania. (Next in Tech, Sept 8; Capital Decanted, Sept 8)
What's new
A note on the window and the ranking: this is the first full session back from the long weekend, so headlines outran fresh operator interviews in places. I've ranked by where the dollars actually land for a book, and flagged who is a specialist or operator versus a commentator. The single biggest event is the Qualcomm-Amazon deal; the single most useful number is Anthropic's compute economics; and the loudest risk is a free-cash-flow drain that one team says has already begun.
1. The biggest development: Qualcomm crashed the custom-silicon party, and Amazon wrote the check. Bloomberg Tech, Sept 8, Kunjan Sobani (Bloomberg Intelligence) and Joanne Feeney (Advisors Capital Management), portfolio manager; corroborated by The Financial Exchange, Sept 8 and Halftime Report, Sept 8.
Qualcomm, best known for the chips in your phone, has spent a year trying to convince the market it can also design custom AI chips for data centers, competing "so late" against Nvidia, AMD and Intel. This deal is the validation. Bloomberg Intelligence's Kunjan Sobani laid out the terms: a $60 billion ceiling on what Amazon can buy through Qualcomm over five to six years (running past 2030), against Qualcomm's own near-term targets of $5 billion and then $15 billion of data-center revenue by 2029. Separately, and this is the circular-financing wrinkle the market keeps flagging, the deal gives Amazon warrants to buy up to $4 billion of Qualcomm stock. Qualcomm shares rose about 4-5%.
Why it matters for a book, in one line from portfolio manager Joanne Feeney: this "gives them better pricing power down the line for acquiring these chips… and more protection against NVIDIA's price levels, which ultimately probably soften a little bit with this increase in competition." Amazon already works with Marvell on custom silicon; leaning on that single partnership "could have been seen as too risky," so Qualcomm is diversification. She also noted the tell for the whole sector, "Broadcom is another player here that Amazon doesn't currently work with", i.e. Amazon has room to add another custom-chip vendor still.
On the Nvidia question, Feeney was measured rather than alarmist: "We should expect [Nvidia's] market share to fall as more players come in. But that doesn't mean that NVIDIA is doomed here… the market itself, we believe, is growing so quickly that all of these players… are playing into a very quickly growing market." The traders on Halftime were more direct: one called it "a structural change for Qualcomm," another noted this is the third data-center announcement Qualcomm has made this year that hasn't yet stuck in the share price, and that Qualcomm "already have an existing relationship with Meta." The read-through: the custom-ASIC land grab is no longer a two-horse race (Broadcom + Marvell), and the "diversify away from Nvidia" trade now has a fresh, cheap way to play it.
2. The most useful number: Anthropic pays $25M a gigawatt and makes $50M+, and the "bubble" chatter is quietly extending the cycle. Bloomberg Surveillance, Sept 8, Gil Luria, DA Davidson (specialist analyst).
Gil Luria made the cleanest bull argument of the week, and it's a unit-economics one, not a vibe. "Anthropic is willing to pay 25 million a gigawatt right now because they're generating more than 50 million a gigawatt. So that means they'd buy far more compute if it was available." In plain terms: for every dollar of power capacity Anthropic rents, it earns roughly two, so the constraint on the buildout is supply, not demand.
His second point reframes the whole "is this a bubble?" debate. The local pushback against data centers, the permitting fights, even the endless bubble commentary are, in his view, a feature: "The good way to prevent a bubble from happening is by talking about a bubble. And everybody these days wants to predict a bubble… including Michael Burry. And that's helping create higher CDS rates… a more tempered market for lending." He tied it to Jensen Huang's guide: Huang expects to grow ~70% next year "but demand would actually justify far higher growth… he could have doubled next year. And that gap is the limits on data center construction." So the friction is slowing the build "in a good way… extending the cycle." Why it matters: it's the strongest rebuttal to the over-ordering fear, the claim is that we are under-building relative to demand, and that the very caution the bears are creating is what keeps this from blowing off the top.
Luria also gave the sharpest power read of the day (more in Read-throughs): because connecting to the grid is "really hard" and forces you to subsidize other ratepayers, "many of the data centers right now are actually building the energy infrastructure behind the meter", turbines, reciprocating engines, Bloom Energy fuel cells, solar. That, he said, "is creating a new bottleneck around the supply of energy and around the supply of turbines and turbine blades and rotors and bloom energy units… and the natural gas eventually will become a bottleneck too." His punchline on why they pay through all of it: a $50 billion data center is delivering "a two, three year return… a very good payback."
3. The bear escalation: Google's free cash flow has already gone negative, and may stay there for two years. Next in Tech, Sept 8, Eric Hanselman and Bren Daly, S&P Global (specialist analysts).
Last week the bears were running a countdown to negative hyperscaler free cash flow. S&P Global's Bren Daly says the clock already struck: "Google, for the first time in its history, reported negative free cash flow. This is a company that used to do $70 billion of free cash flow… And now they don't. And it's only going to get worse for the next two years. That's eight quarters. They're projected to be cash flow negative. This is a stunning reversal."
The framing around it is the sharpest part. Daly walked through how this cycle is structurally different from every prior tech buildout. Hyperscaler capex has gone from roughly $40-50 billion a quarter in the cloud era to about $250 billion a quarter collectively now. And unlike the M&A-driven growth of the 2000s and 2010s, where "you either buy or you build," and companies like Oracle and Microsoft (200-plus acquisitions, including the biggest tech deal in history) bought their way into adjacent markets, the AI buildout "has all been done off balance sheet… supported by capital markets." Microsoft "is an investor in open AI, but it is not an owner the way it owns GitHub." The problem: expenses are recognized upfront, but "all of that backlog is for 2027 and 2028," a multi-quarter revenue lag versus the one-to-two-quarter payback of the cloud era. The market has tolerated it, Daly reminded listeners that Google sold $85 billion of equity back in March, "by far the largest equity offering in history," the market absorbed it, "and they could have sold another probably 15 or 20." But his warning is that the cash engines (Google's ~90% of search advertising, Microsoft's Windows/Office franchise) are now being fully drained to feed the build, and there is no play money left over.
4. The depreciation bomb: adjust for how fast GPUs die, and this buildout is bigger than the railroads. Capital Decanted, Sept 8, "The Prisoner's Dilemma: Investing in the AI Build-Out" (commentary/analysis, not operators).
This episode did the historical math the bears keep gesturing at. Raw AI spending in 2025 (~$400 billion) is "slightly less than 2% of GDP," versus ~1.5% for the dot-com fiber buildout in 2000 and ~6% for railroads at their 1872 peak, so on the surface, modest. But adjust for useful life, and the picture flips. Rail track lasts decades; fiber, a decade-plus; the latest Nvidia GPU is effectively obsolete "in just a few years, and by the way, that cycle is only quickening." Using a still-conservative five-year GPU life, "the depreciation-adjusted percentage of GDP skyrockets from that 2% to nearly 8% of GDP… 2% higher than the railroads and the highest by far in history." The kicker: because clusters are being bought ahead of construction, GPUs are "sitting on shelves… obsolete depending on the construction lifecycle before this project is even up and running", and some companies are stretching depreciation schedules past six years to soften the P&L hit, exactly the wrong direction.
The framework the hosts offered is worth keeping. This is a "prisoner's dilemma", a Cold War-style build where each hyperscaler spends "regardless of the near-term ROI" because pulling back means ceding ground to a rival. They cited Joaquin Clement's FT piece that the ROI math is "mathematically improbable, if not impossible," Deutsche Bank's observation that AI capex is "single-handedly preventing the US from tipping into a recession" (so AI "is creating its own systemic risk"), and Morgan Stanley's estimate that private credit could supply over half of the $1.5 trillion needed for data-center buildouts through 2028. Their most actionable conclusion is the "adjacency play": power, cooling, data-center real estate and networking are "technology-agnostic by design", you don't have to pick the AI winner, you just need the build to continue. But even there they flagged froth: 2025 saw 113 data-center transactions worth >$69 billion (a record, 84% private-equity funded, including the $40B Aligned Data Centers deal by BlackRock/MGX), and Goldman thinks the next leg shifts to "platform stocks and productivity beneficiaries," implying the pure infrastructure trade "may be less attractive than it was 18 months ago."
5. A genuinely new market: you can now hedge compute like a commodity. Energy Gang, Sept 8, energy-market analysts on CME's new compute futures.
Quietly, the plumbing of this trade is maturing: CME Group, with Silicon Data, is launching "compute futures", a standardized, regulated way to hedge the price of GPU compute, "turning unpredictable compute costs into a manageable asset class." The interesting insight for capex-watchers is who needs it. It's not just AI companies: it's the lenders financing data-center construction, who are "exposed in one way or another to the cost and also to the production value of GPU" and today "don't actually have a viable way to look into the future." And because power is "the most volatile component" of GPU economics, a compute contract lets a data-center operator hedge GPUs and the natural gas or grid power feeding them together. Why it matters: when a cost becomes a traded futures curve, financing and siting decisions get easier, which, at the margin, greases more buildout. It's also a fresh, real-time price signal a book can watch once it's live.
The debate
Steel-manning both sides of the hyperscaler capex thesis, now running near $760 billion in 2026 and heading toward $1 trillion-plus in 2027-2028. This cycle didn't change the levels; it hardened both cases. The bulls got a clean unit-economics number and a new supplier proving the demand is deep; the bears got confirmation that the free-cash-flow drain is no longer hypothetical.
Bull, demand is supply-constrained, the economics work, and competition is widening, not collapsing. Anthropic paying $25M/GW to earn >$50M/GW is a ~2x return that says the buyers would take far more compute if they could get it; Jensen's ~70% guide is a ceiling imposed by construction limits, not by demand (Bloomberg Surveillance, Sept 8). The Qualcomm-Amazon deal shows a growing market that can support a fourth serious silicon vendor without dooming Nvidia, "all of these players… are playing into a very quickly growing market" (Bloomberg Tech, Sept 8). And the cash engines are so strong that the market keeps funding the gap, Google placed the largest equity offering in history ($85B) and "could have sold another 15 or 20" (Next in Tech, Sept 8).
Bear, the cash is already gone, the accounting flatters the returns, and rates are turning against risk. Google has already printed negative free cash flow and is projected to stay negative for eight quarters, the countdown didn't count down, it already hit zero (Next in Tech, Sept 8). Adjust for GPU obsolescence and the buildout is ~8% of GDP, bigger than the railroads, with ROI math Deutsche Bank and an FT columnist call near-impossible, and it's the only thing keeping the US out of recession (Capital Decanted, Sept 8). And the macro tailwind just became a headwind: the market now prices a ~60% chance of a Fed hike next week, which raises the discount rate on exactly these long-duration bets and lifts the borrowing cost funding the whole build (Halftime Report, Sept 8; The Financial Exchange, Sept 8).
The synthesis. The two sides are now arguing about the same fact, collapsing free cash flow, and disagree only on the label. The bull says it's offensive spending against a 2x return that supply can't yet meet, so the drain is a feature. The bear says it's a treadmill funded increasingly with debt, into a rate environment that's tightening, with accounting that hides the obsolescence. The most important new data point is that the free-cash-flow question is settled at the fact level (Google is negative); the fight is now entirely about whether the revenue backlog for 2027-2028 shows up on schedule. The clean tell to watch is no longer a countdown, it's Oracle's print Thursday and its CDS.
Sell signals to watch: - The first hyperscaler stock that gets rewarded for cutting data-center plans. As The Financial Exchange put it, the narrative breaks when "the first company whose stock gets rewarded for saying that they are pulling back on data center plans and not seeing the ROI" appears, that's the moment "more capex = more value" inverts (The Financial Exchange, Sept 8). - Oracle's CDS widening through/after Thursday's print. Oracle's five-year credit-default swaps sit above 200 basis points; spreads across hyperscalers "peaked" in late July and have been calming, which is re-risking the trade, a reversal there is the early warning (Halftime Report, Sept 8; The Financial Exchange, Sept 8). - Anthropic's S-1. With the IPO filing imminent, watch the cash-burn/run-rate against that $50M-per-gigawatt claim, if the unit economics don't survive daylight, the cleanest bull number falls (Bloomberg Tech, Sept 8). - A Fed hike that markets actually respect. Two 25bp hikes may not "materially shift the needle," but a higher discount rate is a direct hit to the priced-to-perfection AI multiples (The Financial Exchange, Sept 8). - Depreciation schedules getting longer. If more hyperscalers push GPU useful life past six years, that's management flattering the P&L into the obsolescence wave, a quality-of-earnings red flag (Capital Decanted, Sept 8).
Stocks in play
NVDA. Bull: still the referee everyone routes through, near an all-time high at ~19x forward earnings, with a ~70% growth guide that's supply-constrained, it "could be higher if things on the supply chain become a little bit unstuck," and Anthropic's 2x compute economics say the demand behind it is real (Halftime Report, Sept 8; Bloomberg Surveillance, Sept 8). Bear: the Qualcomm-Amazon deal is one more customer engineering its way to "protection against NVIDIA's price levels," and every hyperscaler custom-silicon win chips at the pricing power the multiple assumes (Bloomberg Tech, Sept 8). Next catalyst: Rubin roadmap detail and how much of its OpenAI financing pool actually draws down.
AVGO. Bull: still the custom-ASIC leader off last week's $115B (FY27)/$230B (FY28) guide, and notably not yet an Amazon supplier, Feeney flagged Broadcom as a player "that Amazon doesn't currently work with," i.e. an open door (Bloomberg Tech, Sept 8). Bear: the custom-silicon field is crowding fast, MediaTek inside Google last week, now Qualcomm inside Amazon, so the scarcity premium on Broadcom's ASIC franchise is eroding at the edges (Bloomberg Tech, Sept 8). Next catalyst: whether Broadcom lands an Amazon program, and any lift to the FY27 ASIC number toward the ~$150B whisper.
AMD. Quiet on the podcasts this cycle, no fresh MI450X/Helios catalyst surfaced in the Sept 8 batch. The only read-through is that it's named among the chip designers portfolio managers still want exposure to in a "very quickly growing market," even as Qualcomm crowds the field it competes in (Bloomberg Tech, Sept 8). Next catalyst: its second-half data-center step-up and any MI450X deployment color.
MSFT. Bull: judged the most durable spender, its AI build is funded off the still-humming Windows/Office cash engine, and it's playing development, not risky M&A, "writing checks to NVIDIA" rather than buying startups (Next in Tech, Sept 8). Bear: it's an investor in OpenAI but "not an owner the way it owns GitHub," so its biggest AI bet sits off balance sheet and leans on the same capital-markets appetite that's now being tested (Next in Tech, Sept 8). Next catalyst: read-through from Oracle's print Thursday.
GOOGL. Bull: ~90% of search advertising still gushes cash, it placed the largest equity raise in history ($85B) into a market that "absorbed it" and wanted more, and it's building its own energy solutions (funding ~$50M of Xcel batteries in Minnesota) to power the build (Next in Tech, Sept 8; Electric Perspectives, Sept 8). Bear: it's the poster child for the drain, first-ever negative free cash flow, projected negative for eight straight quarters, a "stunning reversal" for a company that once did ~$70B a year (Next in Tech, Sept 8). Next catalyst: whether Q3 capex accelerates again and any sign FCF is stabilizing.
AMZN. Bull: the Qualcomm deal is smart supply-chain engineering, a second custom-silicon source alongside Marvell, better pricing leverage over Nvidia, and warrants that align Qualcomm to Amazon's roadmap; AWS remains the reinvest-the-cash flywheel that built the franchise in the first place (Bloomberg Tech, Sept 8). Bear: it's running its own capex at a scale that, like the group, converts a river of revenue into thinning free cash flow, and near-term the Qualcomm chips "won't impact Amazon significantly" or speed up capex deployment (Bloomberg Tech, Sept 8). Next catalyst: proof AWS demand is broad enough to justify the spend, and any additional custom-silicon vendor add (Broadcom?).
META. Bull: still the ad cash machine, and it already has a Qualcomm relationship to lean on as it diversifies silicon (Halftime Report, Sept 8). Bear: relatively quiet on the podcasts this cycle beyond that mention; the structural worry, funding a giant AI build plus Reality Labs losses off the Family of Apps, with its own credit spreads now in the mix, carries over from last week. Next catalyst: whether ad-revenue growth keeps outrunning the capex line.
Also in play: QCOM (the day's winner, +4-5%; three DC announcements this year, cheap, "structural change" per traders, but they're waiting out a heavy September). ORCL (Thursday's print: ~33% top-line growth expected, Abilene, Texas data center coming online with margins to watch, CDS >200bp). BE (Bloom Energy joining the S&P 500 later this month, +9%; the behind-the-meter power pick Gil Luria name-checked). ETN (Eaton upgraded, though the stock "hasn't traded all that well"). (Halftime Report, Sept 8; The Financial Exchange, Sept 8; Bloomberg Surveillance, Sept 8)
Read-throughs
- Power / thermal, behind-the-meter is now the bottleneck, and the ratepayer fight is being solved with contracts. Gil Luria's framing is the cleanest: connecting a big load to the grid is "really hard" and forces you to subsidize other ratepayers, so hyperscalers are building their own power, "turbines or reciprocal engines or bloom energy servers… or solar", which "is creating a new bottleneck around the supply of energy and around the supply of turbines and turbine blades and rotors and bloom energy units," and eventually natural gas (Bloomberg Surveillance, Sept 8). On the utility side, the "who pays" problem is being operationalized: AEP in Ohio pioneered upfront deposits and take-or-pay capacity requirements; data-center developers signed a ratepayer-protection pledge; and NV Energy and Microsoft just agreed the first real, structured ratepayer-protection tariff, the first time the DC-level principle has been "operationalized out in the field" (Electric Perspectives, Sept 8). Actionable: own the picks-and-shovels closest to the power constraint, on-site generation (Bloom, turbines, reciprocating engines) and the equipment names (Eaton upgraded, Vertiv), and treat ratepayer-protection tariffs as the signal that utilities can finally underwrite the load without a political backlash.
- Demand flexibility, the sleeper that could unlock interconnection faster than new supply. A striking claim from the EEI panel: back-of-envelope math suggests about half of the load-flexibility needed to interconnect new data centers faster could be met through software alone, tuning AI workloads in real time without breaking service-level agreements. The named vehicle is Emerald AI (Nvidia-backed, with an infrastructure-fund investor), running pilots with Silicon Valley Power, ComEd and SPP; on the off-site side, Voltus and Google announced ~100 MW of demand response and distributed batteries, and Aligned brought a 30-40 MW battery to a project in exchange for faster interconnection (Electric Perspectives, Sept 8). Actionable: flexible-load and grid-services software is an emerging category, watch the private names (Emerald AI) for eventual public read-throughs, and demand-response operators as a cheaper "capacity" than steel in the ground.
- Utilities / nuclear, gas today, nuclear for the back half of the decade. Black & Veatch's team framed the sequencing bluntly: clients say "I need gas today" to power data centers, while nuclear is "a longer-term, strategic play" filling the gap renewables (availability) and gas (carbon commitments) can't. What's changed is breadth, it's now "SMRs and large-scale nuclear plants" in every client conversation, backed by DOE/NRC policy support and private capital, with license extensions, uprates and even restarts doing the near-term work. The binding constraints are transmission, skilled workforce and nuclear supply chain (Electric Perspectives, Sept 8). Actionable: the nuclear trade (Vistra, Constellation, Talen and the SMR names) is a 2027-2030 story, not a this-quarter one; the near-term power P&L is gas turbines and on-site generation.
- Memory / networking / optics, quiet again, but the momentum is intact. No dedicated Micron/SK Hynix/Marvell/Coherent/Lumentum/Astera/Credo operator episodes surfaced in the Sept 8 window. The only tape-level color: traders flagged "the semis up… optical up… infrastructure higher" as the momentum factor rebounded, and TSMC/ASML/Samsung/Intel/SK Hynix announced a joint move to next-gen (High-NA-class) tools and larger 12-inch photomasks for cheaper leading-edge production starting ~2030, a long-dated cost-curve story, not a near-term catalyst (Halftime Report, Sept 8; Bloomberg Tech, Sept 8). Actionable: still waiting on fresh HBM/pricing operator color, the memory-tightness trade has no new fuel this cycle.
- Financing / market structure, spreads calming, and compute becomes a tradeable asset. The credit picture is, for now, improving: hyperscaler spreads "peaked" in late July and have calmed, "reinvigorating" the AI trade, and if Oracle's revenue looks de-risked Thursday, "everyone else's revenue stream effectively gets de-risked" (The Financial Exchange, Sept 8). Layer on CME's new compute futures, which give lenders a forward curve for GPU output for the first time (Energy Gang, Sept 8), and Morgan Stanley's estimate that private credit funds more than half the $1.5T buildout (Capital Decanted, Sept 8). Actionable: watch compute-backed and neocloud spreads as the first-crack indicator, but note the current signal is calmer, not more stressed, into Oracle.
What changed vs last issue
Last issue (Tuesday, Sept 8, "Broadcom named its number. The bears named a date") covered the Labor-Day batch: Broadcom's $115B/$230B ASIC guide, MediaTek as a new competitor, and a live countdown to negative hyperscaler free cash flow. This edition covers the first full session back, and it advanced the story on three fronts.
- The custom-silicon field widened again, and this time it's Amazon's turn. Last issue's new competitor was MediaTek inside Google. New this cycle: Qualcomm signed Amazon ($60B ceiling through 2030, $4B in warrants, $5B→$15B revenue targets by 2029). The "diversify away from Nvidia" theme now spans Broadcom, Marvell, MediaTek and Qualcomm, and Amazon still hasn't added Broadcom, so the field can widen further (Bloomberg Tech, Sept 8).
- The FCF countdown already hit zero. Last issue: a tracker showing hyperscalers ~50 days from negative free cash flow. New this cycle: S&P Global says Google has already reported negative FCF and is projected to stay negative for eight quarters, the countdown is arguably moot for Google (Next in Tech, Sept 8).
- The demand side got a clean number. Last issue's bull rebuttal to the stranded-asset fear was qualitative (robots and glasses as the next inference wave). New this cycle: a hard unit economic, Anthropic pays ~$25M/GW and earns >$50M/GW, i.e. it would buy far more compute if supply existed (Bloomberg Surveillance, Sept 8).
- The macro backdrop flipped. Last issue anchored on the September FOMC under new chair Kevin Warsh. New this cycle: the market now prices a ~60% chance of a rate hike next week (with oil at multi-month highs and CPI due Friday), a direct headwind to long-duration AI multiples that wasn't front-of-mind last week (Halftime Report, Sept 8).
- Credit read reversed direction, for now. Last issue: Oracle's CDS "trading like junk" and fresh spreads on Meta/CoreWeave/SpaceX, framed as amber. New this cycle: hyperscaler spreads "peaked" in late July and are calming, a partial re-risking, though Oracle's CDS still sit >200bp into Thursday's print (The Financial Exchange, Sept 8).
- The depreciation debate got a number. Last issue: a general accounting/growth-vs-waste debate. New this cycle: depreciation-adjusted AI capex ~8% of GDP, above the 1872 railroad peak, with companies stretching GPU lives past six years the wrong way (Capital Decanted, Sept 8).
- New plumbing: CME's compute futures launch, compute becomes a hedgeable, tradeable commodity for the first time (Energy Gang, Sept 8). And the power read moved from ERCOT battery economics to ratepayer-protection tariffs (NV Energy + Microsoft, the first operationalized) and software load flexibility (Emerald AI) (Electric Perspectives, Sept 8).
- Quiet vs last issue: AMD again (no fresh MI450X/Helios catalyst); no dedicated memory/optics operator episodes (Micron, SK Hynix, Marvell, Coherent, Lumentum, Astera, Credo all quiet); nuclear reframed as a 2027-2030 story rather than a near-term print.
- Still open: whether the 2027-2028 revenue backlog shows up on schedule to justify the negative FCF; whether Broadcom lands Amazon or the ASIC scarcity premium keeps eroding; whether Anthropic's S-1 confirms the $50M/GW economics; and the unresolved crux, offensive "fortress" capex or a treadmill into obsolescence.