Newsletter · · Ashutosh Agarwal

Anthropic's 518 Billion Dollar Bill Lands as CoreWeave Says Orders Aren't Slowing - The AI Capex Tracker - Week of October 1, 2026

A synthesis of what podcasts said about AI capital spending on Wednesday, September 30, 2026, for the week of October 1, 2026: Anthropic's leaked S-1 showing 518 billion dollars of compute commitments against 4.6 billion dollars of 2025 revenue, CoreWeave's CEO saying orders have not slowed, a16z's 780 billion dollar 2026 hyperscaler capex figure, the credit-and-rates bear case, Ben Pouladian on Nvidia and AMD, previews of Micron's results, and ERCOT's 500 gigawatt large-load queue.

The AI Capex Tracker

Week of October 1, 2026: Anthropic's 518 Billion Dollar Bill Lands as CoreWeave Says Orders Aren't Slowing

(Covers podcasts published Wednesday, September 30. Yesterday's issue was about the price of AI output falling: token prices halving and open-source models taking spend share. On Wednesday the argument turned to the bill. Anthropic's leaked IPO filing shows $518 billion of compute commitments against $4.6 billion of 2025 revenue. The podcasts spent the day arguing over whether that is a time bomb or a down payment. On the other side of the ledger, the CEO of CoreWeave said flatly that there "has absolutely not been any reduction in orders." Andreessen Horowitz put this year's big-five capex at about $780 billion. One important gap: Micron reported after Wednesday's close, but none of the podcasts in this sweep came out after the results. Everything on Micron below is a preview, not a reaction.)

TL;DR

  • Anthropic's leaked IPO filing (the S-1, the document a company files with regulators before going public) put a number on the lab-side bill: $518 billion. Per The Enterprise AI Show's reading of the leak, Anthropic carries "$518 billion of long-term infrastructure obligations" through about 2030–2032, plus "about $71 billion in off balance sheet" commitments through special purpose vehicles "mostly through Google and Broadcom," backed by Apollo and Blackstone. That is against 2025 revenue of "$4.6 billion" and an operating loss of "a little over $8 billion" (The Enterprise AI Show, Sept 30). The bear math: about "$350 billion for the debt market just to cover the half a trillion that we know about" (Eurodollar University, Sept 30).
  • The one operator on the podcasts says demand is fine, even with rates rising. CoreWeave CEO Michael Intrator: "There has absolutely not been any reduction in orders… the margins that we are earning are expanding faster than the rates are going up." He sees that holding "through 2027 and into early 28." CoreWeave's client Cognition is "the first client in the world to enter into production on Vera Rubin," Nvidia's next platform (Squawk on the Street, 9AM hour, Sept 30).
  • The capex number is $780 billion this year and over $1 trillion from 2027, and the credit bears are getting louder. a16z: Alphabet, Amazon, Meta, Microsoft and Oracle capex is "about $780 billion" in 2026, "up from $416 billion in 2025," with "over a trillion dollars annually from 2027" (The a16z Show, Sept 30). Jesse Felder counters: "something like $600 billion of total AI debt sold this year," Oracle debt in "stressed territory," and "Nvidia CDS spreads are up 100% this year" (Soar Financially, Sept 30).

What's new

A quick plain-English key. S-1: the registration filing a US company submits before an IPO; it discloses financials and risks. Compute commitment: a contract to pay for computing capacity (chips, data-center space) over future years, whether or not you end up needing it. Special purpose vehicle (SPV): a separate legal entity set up to hold assets and raise debt, often so the debt doesn't sit on the parent company's balance sheet. Operating loss: revenue minus the costs of running the business, before interest and taxes. Neocloud: a specialist company that rents out GPU capacity, like CoreWeave. Credit default swap (CDS): insurance against a borrower defaulting; a wider spread (a higher price) means more worry. Basis point (bp): one-hundredth of a percentage point. Forward P/E: share price divided by next year's expected earnings per share. HBM (high-bandwidth memory): stacked memory chips placed next to AI processors; the scarce part Micron and SK Hynix sell. Depreciation: the accounting charge that spreads the cost of equipment over its useful life. Curtailable load: a big power user that agrees the grid operator can cut its power during emergencies, in exchange for a faster hookup.

1. Anthropic's leaked S-1: $518 billion promised, $4.6 billion earned, and $71 billion parked off the balance sheet. The Enterprise AI Show, Sept 30: a solo host walking through the filing (not named in the transcript). Eurodollar University, Sept 30: the channel's host, a macro and credit commentator. Memes and Markets, Sept 30 and Tom Bilyeu's Impact Theory, Sept 30: commentators relaying the Reuters report on the leak. (All of these are commentators working from a leaked document. No Anthropic executive is quoted, and the hosts stress the financials are unaudited.)

The revenue line. The Enterprise AI Show: 2024 revenue was "somewhere in the $380 to $400 million range," 2025 was "about $4.6 billion," "roughly a 10x increase in a year." For 2026, the host cites "about $11.5 billion for Q2" and a claimed run rate (annualized revenue pace) of "somewhere around $60 billion." Multiplying the quarter by four gives about $46 billion, so he brackets 2026 at "$45 billion and… $60 billion."

The cost line. The 2024 operating loss was "roughly $3 billion"; 2025 was "a little over $8 billion." If losses double again, he gets "probably somewhere in the range of $15 to $20 billion for 2026. And that's on $60 billion of revenue." Memes and Markets fills in the 2025 detail: "spending is at $12.6 billion, and $7 billion of that went to computing power." The widely quoted "$42 billion net loss" includes "around $34 billion" that "is technically an accounting charge and not cash." Cash at the end of 2025 was "$20 billion."

The commitments, and where they sit.

"Right now they're on the books for $518 billion of long-term infrastructure obligations. I think that's through 2030 or 2032… on an annual basis, a little more than $100 billion of forward commitments." (The Enterprise AI Show)

Then the off-balance-sheet piece: "about $71 billion in off balance sheet… commitments… special purpose vehicles… mostly through Google and Broadcom… backed by other financial institutions like Apollo Global, Blackstone." Add them up and "that future commitment is roughly $600 billion." Eurodollar University adds a detail that matters for Broadcom holders: under one of the SPV deals, "Broadcom is backstopping some of the payments with their… residual value support." (Residual value support means Broadcom guarantees part of what the hardware will be worth later, so it carries some of the risk if that value falls.)

Customer concentration. Quoting Reuters, Memes and Markets says "nearly 1/4 of their revenue from last year came from just 2 customers," and that the filing's risk factors warn "many of its largest clients were not locked into long-term contracts and could stop… spending."

Renting, not building. The Enterprise AI Show makes a structural point. Anthropic is "still primarily sort of renting GPUs… from AWS, from Azure… from Google, and now from SpaceX," unlike "Oracle and OpenAI, who are making huge commitments to build and own their own data centers." That makes Anthropic "more of a renter," which is safer in a slowdown. The catch is that "the people that they are renting from… are also some of their largest competitors."

The bear arithmetic (Eurodollar University). Assume Anthropic raises "roughly $100 billion in equity" and generates "$50 billion in revenue" to be generous. "That leaves $350 billion for the debt market just to cover the half a trillion that we know about right now." He calls "$518 billion… the best case scenario," because "all of these companies are competing for a limited amount of equipment, which are driving up prices." His big-picture comparison: "The SPVs are back… central to this case, just like they were the mortgage bubble a generation ago."

Valuation. Tom Bilyeu says the IPO price tag being discussed is "north of $2 trillion… double what they were worth in May." (Yesterday's podcasts had "one to $1.5 trillion"; the gap reflects which reports each show is reading, so treat both as rumor.)

Why it moves numbers. $518 billion over roughly five or six years is about $100 billion a year of guaranteed demand landing on the hyperscalers' clouds and on the chipmakers. It is a bull data point for AMZN, GOOGL, MSFT and Nvidia's order books. It is also a concentration risk: one lab with $20 billion of cash and an $8 billion operating loss is underwriting a chunk of that backlog. The Broadcom residual-value guarantee is the most specific new risk for a public stock. Watch whether hyperscalers break out how much of their cloud backlog comes from Anthropic and OpenAI on the Q3 calls.

2. CoreWeave's CEO: no slowdown in orders, margins outrunning rates, and the first Vera Rubin customer in production. Squawk on the Street, 9AM hour, Sept 30: Michael Intrator, CEO of CoreWeave (the transcript spells it "Intrader"), interviewed by the CNBC anchors (Jim Cramer among them) from CoreWeave's Fully Connected conference in San Francisco. (Operator testimony, and the most important of the day. Keep in mind he runs a heavily indebted GPU landlord that has every reason to sound confident.)

Rubin is live. "We brought our first client. It is the first client in the world to enter into production on Vera Rubin architecture. Cognition." (Cognition is an AI coding startup.) Rubin pairs Nvidia's GPU with "NVIDIA's CPU" in one server. CoreWeave also has "an enormous number of CPUs that are outside of the server."

On the political backlash against data centers:

"There has absolutely not been any reduction in orders… the regulations that are being brought out around data centers, all they are going to do is they're going to relocate the infrastructure. They are not going to in any way reduce the demand."

On rates. An anchor asked whether the rate backdrop is "getting troublesome." Intrator: "Rates are rising, which are making things more expensive… What is important to take away from this discussion is that the margins that we are earning are expanding faster than the rates are going up. So the overarching profitability of the infrastructure that we are bringing online is higher than it was a year ago." On how long that lasts: "at least as far as I can see out through 2027 and into early 28."

On CPUs. Asked whether the ratio is moving toward more CPUs per GPU: "You are going to see substantial demand for CPUs across the board. They are going to be inside the GPU servers. They are going to be independent of the GPU servers." Conference attendance was "expecting 2,000 people. We got 5,000."

A Nvidia footnote from the same hour. Cramer, crediting SemiAnalysis: OpenAI's new GPT-6.1 model is not running on Cerebras chips, as many assumed. "It is an NVIDIA GPU, low batch size." He notes Nvidia "has been stuck" around "$2.30" (he means about $230) and says "today's the day" for the buyback to push it through.

Why it moves numbers. This is the direct answer to Felder and the other credit bears (item 4). The weakest borrowers in the chain are the neoclouds. The man running the biggest one says his returns on new capacity are rising faster than his funding costs, and that he can see demand into 2028. If that holds, rising rates squeeze CoreWeave's equity holders but not the order book behind Nvidia. If it doesn't, CoreWeave is where it shows first. Watch CoreWeave's next bond deal: the yield it pays is a live test of his claim.

3. a16z on the scale: $780 billion this year, $1 trillion-plus from 2027, $1.7 trillion of cloud backlog, and parts unavailable until 2028. The a16z Show, Sept 30: Andreessen Horowitz's "State of Markets" episode, a panel of the firm's investment partners (identified only as speakers in the transcript). (Venture investors with heavy AI exposure. They are talking their book, but these are the cleanest aggregate numbers of the day.) Supporting: MoneyWeek Talks, Sept 30 and The AI Daily Brief, Sept 30.

The capex path. "If you just look at the hyperscalers, Alphabet, Amazon, Meta, Microsoft, and Oracle, their CapEx in 2026 is about $780 billion. That's up from $416 billion in 2025. And all expectations point to them spending over a trillion dollars annually from 2027." The partner adds that "successive forecasts… have moved sharply higher in pretty short sessions. So spending that once looked like a ceiling… has become near-term." MoneyWeek's guest gives the last round of guidance changes: capex "expected to grow over 90% year over year… an acceleration from 73%… in 2025." Amazon "increased their 2026 capex outlook by 10%. Google raised it by 8%. Meta… more on the lower end of the range. Microsoft maintained their guidance."

The backlog. "Microsoft, Google, and Amazon have about 1.7 trillion of combined cloud backlog together." Free cash flow is "depressed during this build out," but "consensus forecasts are showing recovery from 2028."

Supply. "Everyone we talk to at every stage of the supply chain is telling us the same thing… demand outstripped supply. There are certain elements in the data center supply chain where you can't get access to materials or products until 2028." On GPU rental prices: "any GPU that you can bring online is being priced at an attractive rate."

The depreciation question, answered bullishly. Citing Amazon's last call, they say "the useful life of GPUs… or TPUs, is actually pretty long… those will have a very useful economic life for a long period of time. And it's been longer than I think any of us expected."

The proof-of-demand anecdote. OpenAI "had to pause new subscriptions on their pro plans" two weeks ago. The AI Daily Brief has the follow-up from OpenAI's DevDay: Pro has reopened, but "usage calculations have been changed, netting out to a 50% reduction in terms of API cost," in other words users get less for the same price. The host's conclusion: "compute constraints are real, present, and permanent… new compute isn't coming online at anywhere near the speed that people are increasing their use."

The price war continues anyway. Same show: OpenAI's new GPT-6.1 Sol was pitched as "near Astra Intelligence for a fifth of the price." In one benchmark it cost "a third the cost of 6 Sol and 13% the cost of Astra." Cisco President Jeetu Patel puts numbers on the volume side: token consumption by agents "go up 14x" since February, and "it is going to be an enormously supply constrained world for a long, long time" (AI Proving Ground Podcast, Sept 30).

Bubble check. a16z's case that this isn't 2000: "stocks are up about 20% while multiples are down about 20%… The S&P 500 earnings multiple is below 20 times," and "memory companies… are trading for, call it six times, seven times forward earnings."

Why it moves numbers. $780 billion is above the "$700B+" anchor this newsletter has used, but note it includes Oracle, so it is a five-company number, not four. The more useful fact is the pairing: per-token prices keep falling (a fifth, a third, 13%) while providers ration capacity (paused signups, a 50% cut in what a Pro subscription buys). That pattern means demand is outrunning supply even at lower prices, which is the bull case. What would break it is a quarter where token prices fall and hyperscaler cloud revenue growth slows at the same time.

4. The credit-and-rates bear case: $600 billion of AI debt, Oracle CDS at a record, and a 10-year heading for 6%. Soar Financially, Sept 30: Jesse Felder (founder, The Felder Report). Memes and Markets, Sept 30. Broken Pie Chart, Sept 30: an advisor-hosted markets show quoting Goldman Sachs. Counterpoint: What Bitcoin Did, Sept 30: Jordi Visser (macro investor). Rates: Squawk on the Street, 11AM hour, Sept 30. (Macro commentators and a strategist note relayed second-hand. No credit-desk operator.)

Felder's list of cracks. He argues that this year "the AI bubble crossed from being completely funded by… the cash flows of the hyperscalers, to now overwhelming the cash flows of the hyperscalers and demanding debt financing." Then the evidence:

"Goldman Sachs put out a thing, there's been something like $600 billion of total AI debt sold this year… we've seen Oracle debt enter stressed territory. We saw… bonds backing a Jane Street… data center project enter stressed territory. SoftBank was forced to sell debt at junk yields for the first time last week. Nvidia CDS spreads are up 100% this year."

He adds that "major IPOs for Anthropic [were] pushed back a month," "OpenAI said we're not going to come public till next year," and "data center companies like SB Energy and Holtec have delayed their IPO." The borrowers he worries about most are "the neo clouds… CoreWeave and… these guys who are already… having to sell debt at junk yields, or could be priced out of the markets."

Private credit. Citing PIMCO, "6, 6.5% of portfolio companies are in default." Counting companies that pay interest in kind (adding unpaid interest to the loan instead of paying cash), "it's closer to 20% already default rate in private credit." On scale: AI investment is "over 3% of GDP, which makes it bigger than any previous investment boom, bigger than the railroads."

His core argument is about moats. The hyperscalers "have backgrounds where they were essentially monopolies… Now they're all competing in a market to… offer compute," which is "a commodity product that has no moat." His prediction: "the cost of compute is going to collapse," as with dark fiber after 2000. That would be good for whoever builds on top, and bad for whoever paid for the build.

Oracle. Memes and Markets: Oracle's "5-year credit default swap spread just hit a record 230 basis points."

Depreciation hits the index. Broken Pie Chart reads out a Goldman Sachs strategist note: hyperscaler "depreciation expenses… will continue to increase as CapEx growth decelerates… We estimate a drag from hyperscaler depreciation expense on S&P 500 earnings growth of 5 percentage points… in 2027. Offsetting nearly half of the 11 [percentage point] boost to earnings from CapEx spending."

Rates. CNBC put the 10-year at "5.28" on Wednesday. Felder sees it "heading for 6%," and he expects "a full rate hike cycle" ("We just got our first hike"). On Squawk, the anchors noted the 10-year rose "almost 50 bips" in September, while New York Fed President Williams said there's "no urgency to hike" (Squawk on the Street, 9AM hour, Sept 30).

The rebuttal (Visser). "The Mag 7 CDS is widening out and all these scary things. Okay, but their equity's at all-time highs. Their equity is lower in the cap structure than their debt, guys." His read: people who think it's a bubble buy CDS as a cheap hedge, which widens spreads without signaling real default risk. On balance sheets: "no risk of default… borrowing a total of $2 trillion between them when they're worth $20, $30 trillion in the aggregate." On Nvidia: "Street was expecting 40% top-line revenue growth and they said it was going to be 70%… And the stock was unchanged 4 days later." He calls it "the most hated bull market around the greatest innovation technology the world has ever seen."

Why it moves numbers. Felder and Visser agree on the facts (CDS is wider, multiples are lower) and disagree on what they mean. Felder says lenders are pulling back. Visser says it's crowded hedging. The Goldman depreciation number is the most concrete new figure. A 5-point drag on 2027 index earnings growth means that once spending growth slows, the earnings boost reverses fast. The trade implication: the neoclouds and Oracle are where rates bite first. The hyperscalers' own bonds are where Visser's argument is strongest.

5. Ben Pouladian on chips: AMD's Helios weighs eight tons, Anthropic's billions go to Nvidia and TPUs, and GPUs are "Rodeo Drive" collateral. Monetary Matters with Jack Farley, Sept 30: Ben Pouladian (independent semiconductor investor and writer, a long-time Nvidia bull) with host Jack Farley, who says he is "overweight NVIDIA." (Investor commentary from a self-described Nvidia bull. Some of it is supply-chain chatter, not confirmed fact.)

AMD's rack problem. "When I saw the Helios rack, it's basically twice the size of a Blackwell system." Nvidia's Vera Rubin "went cable-less and… simplified the design." On AMD's timing: Lisa Su said Helios would "ramp at the end of this year into early next year, and already I'm hearing that there's potential delays in scaling it and manufacturing it, and the thing weighs like eight tons." (That is a rumor, not a disclosed delay.) His view is that AMD is "squeezed in the middle" between Nvidia and in-house chips like OpenAI's, Meta's "MTIA" and Microsoft's "Maia."

Where Anthropic's money goes. "With the new S1 that leaked for Anthropic, they've committed over a half a trillion dollars… I don't think that's Tranium. I think a lot of that is probably NVIDIA GPUs and Google TPUs." His reason: "a lot of Claude was trained on Tranium and you would have a lot of outages." (Trainium is Amazon's in-house AI chip. This is his inference, not a disclosed breakdown.)

Musk switching sides. Musk once talked about building "our own GPU… Dojo," then "comes out and says… We like the NVL 72 Vera Rubin platform, Tesla, SpaceX, exclusive NVIDIA."

Why GPUs get financed and custom chips don't. This was his most interesting argument:

"Your GPU is like your High Street, Rodeo Drive, Bond Street… Fifth Avenue type of location. And you know, once that space goes vacant, there's a list of people waiting to fill up that spot. The question is, what street is TPU on? What street is Tranium on?"

In plain terms, a GPU can be rented to anyone, so lenders treat it as collateral. A custom chip only works for the company that designed it. "You can try to finance TPUs, but you can't finance like a lot of these other custom startup types of chips."

Why Nvidia has lagged. Farley shows a TrendForce table (TrendForce is a Taipei semiconductor research firm) where every component category, "CPUs, GPUs, DRAM, NAND… ABF… MLCCs," is "tight to very tight except for GPUs, which was balanced." Pouladian agrees that the market is "chasing bottlenecks." Nvidia primed its own supply chain early, so the shortages are now in its suppliers' parts.

Token economics. He expects Wall Street to start valuing labs on "what is it costing company X to produce these tokens versus company Y," with "a discount on your valuation" for high-cost producers.

Why it moves numbers. The TrendForce point explains this year's rotation: the money follows whatever is in shortest supply, and right now that is memory, substrates and CPUs, not GPUs. The financeability argument matters for AVGO and GOOGL. If lenders keep preferring Nvidia hardware as collateral, custom-chip build-outs depend more on the buyer's own balance sheet or on guarantees like Broadcom's residual-value support (item 1).

6. Micron: the previews set the bar. No podcast had reacted to the results yet. Squawk on the Street, 10AM hour, Sept 30: Mehdi Hosseini, senior equity research analyst at Susquehanna (transcribed as "Susswahana"), Buy rating, "$2,000 share price target." Squawk on the Street, 11AM hour, Sept 30: CNBC reporter Seema Modi. Bloomberg Tech, Sept 30: host Ed Ludlow with Beth Kindig (IO Fund). Saxo Market Call, Sept 30: Saxo Bank's daily markets podcast host. Broken Pie Chart, Sept 30. (One specialist sell-side analyst; the rest are reporters and commentators. All recorded before the print.)

The bull bar (Hosseini). "Compute is actually becoming more commoditized. And it's memory that makes the difference." He expects "Micron earnings to reach $30, $40 on a quarterly basis… ultimately… a $50 quarterly earning or $200 annualized." The test isn't next quarter's guidance but "sustainability of margin profile… operating margin in the 70%." The difference from past memory cycles is "diversification in demand drivers… [and] in customer base." He notes CPU shortages also "drive incremental demand for DRAM."

The setup. The stock is "up nearly 300 percent so far this year," but flat for four months since early June. "95 percent of the analysts that cover Micron have a buy. The consensus price target is for 50 percent upside." Modi: Morgan Stanley, "using data from TrendForce," sees HBM "pricing going up by 15% to 20% in the fourth quarter." Micron "has pledged over $250 billion in US manufacturing… to produce about 40% of its DRAM stateside in 10 years." Melius says Micron "can buy back well over 10% of its market cap this year." Bloomberg cites "adjusted margins 87%." Kindig warns that "strategic customer agreements could mistakenly cause the market to think that Micron will not participate as HBM and DRAM continue to surge in pricing."

The valuation puzzle. The Saxo host: Micron's "current quarterly run rate is on the order of $33 billion in operating income. That's more than half of what NVIDIA does… and yet it's valued at less than a quarter of NVIDIA's $5.x trillion." Broken Pie Chart: Micron at "7.1 times forward earnings," Nvidia at "18.99 times" and Costco at "40.5." Its straddle math implies a break-even move of "more than 8%," versus the about 6.5% in yesterday's issue. (The host quotes the stock at "$10.53"; he means about $1,053.)

Timing. Micron reported after Wednesday's close. None of the podcasts in this sweep were published after the release, so this issue does not report the numbers or the stock's reaction. CEO Sanjay Mehrotra is booked on CNBC "in the 9 a.m. hour" Eastern today (Thursday).

Why it moves numbers. Hosseini's "70%" operating-margin claim and the "87%" margin baseline are the bars to grade the print against. A 15–20% HBM price increase in Q4 is the input to check against the guide. Kindig's warning about long-term contracts is the same risk this newsletter flagged in yesterday's checklist.

7. Power: ERCOT has 500 gigawatts of large-load requests, batteries can host data centers, and 60% of Americans now oppose new ones. Sustainability Currents, Sept 30: sustainable-finance bankers from Crédit Agricole (first names only in the transcript). Energy Capital Podcast, Sept 30: a Texas energy show interviewing a grid-battery developer about an ERCOT study (guest not named in the excerpt). Climate Rising, Sept 30: Harvard Business School's climate podcast. Trader Talk, Sept 30. (Bankers, a vendor and policy voices. No utility executive.)

The queue. A Crédit Agricole banker: "in ERCOT… they have received requests… from large loads that have totaled 500 gigawatts as of this summer. And that's more than 5 times the size of the current Texas peak demand record." Climate Rising adds the scale shift: utility CEOs once considered "a 50 megawatt facility" big. "Now they're talking about data centers that are 5,000 megawatts." US demand growth is "projected to rise at three or five percent in this decade… a rate we haven't seen since the 1970s." Trader Talk, citing S&P: data centers use "5%" of the grid now, going "to 15% to 20% by 2030."

Batteries as a substitute for transmission (Energy Capital). In the ERCOT study, a "100-megawatt data center" at Burleson Switch (Dallas area, Oncor territory) creates a string of transmission bottlenecks. "All of the constraints that it creates can be ameliorated by 80 megawatts of strategically sited batteries." The reason is location: a battery right next to the congested line gives full relief, while cutting the data center itself "only realizes, say, like half a megawatt" per megawatt. The batteries can be built "in less than 6 months." This connects to ERCOT's new "batch" study process, where data centers can opt to be curtailable to connect faster. That leaves "multiple hundreds of gigawatts of data centers electing to be curtailable" exposed to cutoffs costing "hundreds of thousands of dollars a megawatt hour."

The politics. Crédit Agricole: an August University of Pennsylvania study found "about 60% of US adults are now opposed to new data center development… up 12 percentage points from beginning of the year." Community opposition is "the fastest growing area of investor concern" because it "can really delay projects by years." Lenders now view data centers "as power infrastructure" and want to know each operator's "power strategy." a16z offers the other side: "for every 10% increase in data center capacity, residential rates went down by 40 bps" (The a16z Show, Sept 30).

Equipment. On Squawk, the CEO of Accelivation, a data-center infrastructure company that listed Wednesday, said building gear in factories and shipping it "complete" saves "over 80%" of installation time. The anchors noted that "AI industrials… have actually come well off their highs because of the backlog concerns." He named "medium-voltage gear" as the usual upstream delay (Squawk on the Street, 10AM hour, Sept 30).

Why it moves numbers. 500 GW of requests against roughly 100 GW of Texas peak means most of these projects will never get built as proposed. The queue measures demand intent, not demand. The investable point is that connection speed is now the scarce resource. Batteries, medium-voltage gear and factory-built power systems all sell time, and this is the third day running that a podcast has made that case.

The debate

Bull: the bill is big because demand is real, supply is rationed, and the strongest borrowers are paying.

  • Orders aren't slowing. "There has absolutely not been any reduction in orders," and returns on new capacity are "expanding faster than the rates are going up" (Squawk on the Street, 9AM hour, Sept 30).
  • Spending keeps beating its own forecasts. "$780 billion" in 2026, "over a trillion" from 2027, "$1.7 trillion" of cloud backlog, and parts unavailable "until 2028" (The a16z Show, Sept 30).
  • Rationing at lower prices. OpenAI cut prices to "a fifth" and still had to shrink Pro allowances 50% because "compute constraints are real, present, and permanent" (The AI Daily Brief, Sept 30). Agent token use is up "14x" since February (AI Proving Ground Podcast, Sept 30).
  • Lab revenue is compounding. Anthropic went from about "$400 million" to "$4.6 billion" to a claimed run rate of about "$60 billion" in two years (The Enterprise AI Show, Sept 30). OpenAI has "1.2 billion" users, and its CFO says "we could use more compute in 2027" (Bloomberg Tech, Sept 30).
  • Valuations aren't 2000. Nvidia at "18.99 times," Micron at "7.1" (Broken Pie Chart, Sept 30). Big tech owes "$2 trillion" against "$20, $30 trillion" of value (What Bitcoin Did, Sept 30).

Bear: the commitments far outrun the revenue, the debt is moving to weaker hands, and depreciation is coming.

  • The lab-side gap. "$518 billion" plus "$71 billion" off balance sheet, against "$4.6 billion" of revenue and an "$8 billion" operating loss (The Enterprise AI Show, Sept 30). About "$350 billion" has to come from debt markets (Eurodollar University, Sept 30). A quarter of revenue comes from "just 2 customers" who are "not locked into long-term contracts" (Memes and Markets, Sept 30).
  • Credit is tightening at the edges. "$600 billion" of AI debt this year, Oracle and Jane Street project debt "stressed," SoftBank at "junk yields," "Nvidia CDS spreads… up 100%," and private credit defaults "closer to 20%" including payment-in-kind loans (Soar Financially, Sept 30). Oracle five-year CDS is at a "record 230 basis points" (Memes and Markets, Sept 30).
  • Depreciation reverses the earnings boost. A "5 percentage point" drag on 2027 S&P earnings growth, "offsetting nearly half" of the capex boost (Goldman, via Broken Pie Chart, Sept 30).
  • Compute is becoming a commodity. It is "a commodity product that has no moat," and "the cost of compute is going to collapse" (Soar Financially, Sept 30). Even a memory bull says "compute is actually becoming more commoditized" (Squawk on the Street, 10AM hour, Sept 30).
  • The social license is shrinking. "About 60%" of US adults oppose new data centers, "up 12 percentage points" this year (Sustainability Currents, Sept 30).

Sell signals to watch (updated).

  1. Micron's print and Mehrotra's 9 a.m. interview today: margin guidance below the "87%" baseline, HBM pricing short of the "15% to 20%" Q4 increase, or long-term contracts capping price gains.
  2. Neocloud funding costs: CoreWeave or another GPU landlord issuing at wider spreads than its last deal. That would contradict Intrator's "margins… expanding faster than the rates."
  3. Anthropic's official S-1: commitments above "$518 billion," more SPV debt, or customer concentration worse than "1/4" from two customers.
  4. The 10-year moving through about 5.5% (Felder's path to "6%"; it was "5.28" Wednesday).
  5. Oracle CDS above 230 bp and an Oracle rating action (carried over).
  6. Hyperscaler cloud revenue growth slowing while token prices fall (carried over).
  7. Nvidia receivable days in its next filing (carried over).

Stocks in play

(No closing prices from Wednesday Sept 30 were quoted on today's podcasts except where noted.)

NVDA. Bull: OpenAI's GPT-6.1 runs on "an NVIDIA GPU," not Cerebras, per SemiAnalysis as relayed by Cramer, and Cognition is the first production customer on Vera Rubin at CoreWeave (Squawk on the Street, 9AM hour, Sept 30). Pouladian thinks much of Anthropic's half-trillion goes to "NVIDIA GPUs and Google TPUs," and Musk's companies are "exclusive NVIDIA" on Vera Rubin (Monetary Matters, Sept 30). The stock trades at "18.99 times" forward earnings (Broken Pie Chart, Sept 30). Bear: GPUs are the one "balanced" category in TrendForce's supply table, which is why the shortage trade has gone elsewhere (Monetary Matters, Sept 30). "Nvidia CDS spreads are up 100% this year" (Soar Financially, Sept 30). Breaking Points calls the $150 billion buyback financed by "very circular" deals (Breaking Points, Sept 30). Next catalyst: the Micron read-through today; hyperscaler Q3 calls in late October.

AVGO. Bull: Broadcom is named as a counterparty in Anthropic's "$71 billion" of off-balance-sheet compute deals (The Enterprise AI Show, Sept 30). That is a large, identifiable customer flow. Bear: under one SPV, "Broadcom is backstopping some of the payments" with "residual value support" (Eurodollar University, Sept 30). That puts Broadcom on the hook if the hardware's value falls. Pouladian argues custom chips are hard to finance: "what street is TPU on?" (Monetary Matters, Sept 30). Next catalyst: Anthropic's official S-1 disclosure of the SPV terms; custom-chip commentary on hyperscaler Q3 calls.

AMD. Bull: AMD has been outperforming Nvidia year to date, per host Jack Farley (Monetary Matters, Sept 30). The CPU boom is real: "substantial demand for CPUs across the board" (Squawk on the Street, 9AM hour, Sept 30). Bear: the Helios rack is "twice the size of a Blackwell system," weighs "like eight tons," and there are "potential delays in scaling it" (unconfirmed chatter). Nvidia's new agent-security sandbox "works on x86 ARM and NVIDIA chips, but not on AMD" (Monetary Matters, Sept 30). Next catalyst: Helios ramp confirmation "end of this year into early next year"; the Q3 print.

MSFT. Bull: part of the "1.7 trillion of combined cloud backlog" (The a16z Show, Sept 30). Azure is one of Anthropic's GPU landlords (The Enterprise AI Show, Sept 30). Bear: "maintained their guidance" while peers raised (MoneyWeek Talks, Sept 30). Enterprise AI exposure is "still mostly with Microsoft Copilot, which just shows you how far they have to go" (The a16z Show, Sept 30). Next catalyst: fiscal Q1 earnings, late October.

GOOGL. Bull: raised 2026 capex "by 8%" (MoneyWeek Talks, Sept 30). Google is one of the two named counterparties in Anthropic's SPV deals (The Enterprise AI Show, Sept 30), and TPUs are named alongside Nvidia as where Anthropic's spending likely goes (Monetary Matters, Sept 30). Climate Rising praises Google's long-term geothermal power deal with Fervo (Climate Rising, Sept 30). Bear: renting to Anthropic means funding "some of their largest competitors" (The Enterprise AI Show, Sept 30), and SPV exposure is tied to one unprofitable customer. Next catalyst: Q3 earnings, late October.

AMZN. Bull: raised 2026 capex "by 10%" (MoneyWeek Talks, Sept 30). Its last call explained "this sort of J-curve dynamic," the idea that cash flow dips during the build and recovers later because chips stay useful for a long time (The a16z Show, Sept 30). Bear: Pouladian says Trainium "fell behind" and doubts Anthropic's new commitments go to it (Monetary Matters, Sept 30). Next catalyst: Q3 earnings, late October; AWS's share of Anthropic's commitments in the official S-1.

META. Bull: Cramer calls Muse "a real product that's going to produce real money" for small businesses, and notes Meta "did recover… during the course of the day" (Squawk on the Street, 9AM hour, Sept 30). Bear: Meta's 2026 capex is "more on the lower end of the range" (MoneyWeek Talks, Sept 30). That is a data point, not a thesis. Next catalyst: Q3 call in late October and the first 2027 capex number; any mention of Muse demand on Micron's call.

Also in play: Micron (MU): reported after Wednesday's close. No podcast reaction yet; see item 6. CoreWeave (CRWV): operator bull case versus the credit bears (items 2 and 4). Oracle (ORCL): five-year CDS at a "record 230 basis points" (Memes and Markets, Sept 30). HPE: held a networking investor day and is "raising their guidance for networking revenue" (Squawk on the Street, 9AM hour, Sept 30). OpenAI (private): seeking "at least $30 billion" at a "$1.4 trillion" valuation after "pushing back plans to go public" (Bloomberg Tech, Sept 30).

Read-throughs

  • Memory (MU, SK Hynix, Samsung): the bar is 70% operating margins and 15–20% HBM price increases. Hosseini's "70%" operating-margin call and Morgan Stanley's "15% to 20%" Q4 HBM pricing (Squawk on the Street, 10AM and 11AM hours, Sept 30) are the numbers to grade against. Options priced a move of "more than 8%" (Broken Pie Chart, Sept 30). Cramer's setup for a rally: "if you get SK Hynix up tonight in Korea and you get a good number from Micron and you get a good bond number and you see oil down" (Squawk on the Street, 9AM hour, Sept 30). Most actionable: the reaction to the print, plus Mehrotra's comments on long-term-contract pricing and buybacks at 9 a.m., set the tone for the whole memory group today. Melius's "well over 10% of its market cap" buyback idea is the upside wildcard.
  • Power equipment (VRT, ETN, medium-voltage gear, batteries): connection speed is what's scarce. ERCOT's "500 gigawatts" of requests (Sustainability Currents, Sept 30), 80 MW of batteries hosting 100 MW of load (Energy Capital Podcast, Sept 30), and "medium-voltage gear" as the usual delay (Squawk on the Street, 10AM hour, Sept 30). Most actionable: "AI industrials… have… come well off their highs because of the backlog concerns" (same source). If CoreWeave's "no reduction in orders" holds, that pullback is the entry point. Watch how Accelivation trades as a sentiment gauge.
  • Utilities / ERCOT: politics is now a cost. "About 60%" of US adults oppose new data centers (Sustainability Currents, Sept 30), but Intrator says regulation "will relocate the infrastructure" rather than reduce it (Squawk on the Street, 9AM hour, Sept 30). Most actionable: favor utilities in states that want the load (Texas) over those adding restrictions. Data-center growth from "5%" to "15% to 20%" of grid use by 2030 (Trader Talk, Sept 30) is still the core growth driver.
  • Clean firm power: one data point. The clean-firm-power item this cycle was Google's geothermal deal with Fervo (Climate Rising, Sept 30).
  • Networking (HPE): one hard data point. HPE is "raising their guidance for networking revenue" (Squawk on the Street, 9AM hour, Sept 30). Most actionable: a guidance raise from a second-tier networking vendor suggests broad demand, not just at the leaders.
  • CPUs / substrates (INTC, ARM, AMD; ABF and MLCC suppliers): TrendForce shows everything but GPUs "tight to very tight" (Monetary Matters, Sept 30), and CoreWeave confirms "substantial demand for CPUs." Most actionable: the shortage trade remains the best place to be, but Pouladian warns "a lot of the good news is really already priced in on the forward curve."
  • Credit (CRWV, ORCL, SoftBank, AI project bonds): the weakest borrowers set the pace. Oracle CDS at a "record 230 basis points" (Memes and Markets, Sept 30); Jane Street project bonds "stressed" and SoftBank at "junk yields" (Soar Financially, Sept 30). Most actionable: if you want to hedge the AI book, these are the cleaner instruments. Visser's point stands that hyperscaler CDS reflects hedging demand more than default risk.

What changed vs last issue

Last issue (Wednesday, Sept 30, "Token prices halve, open models surge. Micron reports tonight.") covered Tuesday. It focused on the price of AI output: 50% token price cuts, open-source models going from 7% to 56% of spend on one platform, and Raja Koduri's $50–60 billion per gigawatt. This edition covers Wednesday. The story moved from what a token is worth to who has promised to pay for the capacity.

  • Aggregate capex anchor: $780 billion for 2026 (five companies), over $1 trillion from 2027. Last issue: no new aggregate; Koduri's "$24 trillion" demand-implied figure and earlier "$9 trillion" through 2031 (Bloomberg Intelligence). This cycle: a16z's "$780 billion," "up from $416 billion," and "over a trillion dollars annually from 2027" (The a16z Show, Sept 30). It includes Oracle, so it is above but consistent with this newsletter's "$700B+ for 2026 / ~$1T for 2027" four-company frame. New: "$1.7 trillion" of cloud backlog at MSFT, GOOGL and AMZN.
  • Demand side: from price cuts to price cuts plus rationing. Last issue: 50% token price cuts. This cycle: another step down (GPT-6.1 Sol at "a fifth of the price"), but also OpenAI cutting Pro value "50%" and pausing signups because of compute limits. Lower prices haven't freed up capacity.
  • Lab commitments: new, and the biggest number of the cycle. Last issue: Anthropic IPO timing "still debated." This cycle: the leaked S-1 shows "$518 billion" of commitments, "$71 billion" off balance sheet, and Broadcom "residual value support."
  • Operator voice: back. Last issue: no hyperscaler or neocloud executive. This cycle: CoreWeave's CEO on orders, margins versus rates, and the first Vera Rubin production customer. Cisco's president on "14x" growth in agent token use.
  • Credit: still contested, with new figures. Last issue: hyperscaler CDS "flat" (The Compound) versus "exploding" the day before. This cycle: "Nvidia CDS spreads… up 100%" and Oracle five-year CDS at a "record 230 basis points" on one side, Visser's "equity's at all-time highs" on the other. Still no actual spread level for a hyperscaler.
  • Rates: new print. Last issue: no yield level. This cycle: the 10-year at "5.28," up "almost 50 bips" in September, with Felder calling for "6%."
  • Depreciation: quantified. Last issue: "the life expectancy of a souffle" (Collum) versus long useful life. This cycle: Goldman's "5 percentage points" drag on 2027 S&P earnings growth, versus a16z's "longer than… any of us expected" useful life.
  • Power: ERCOT queue sized. Last issue: ERCOT reviewing large loads for cascading-failure risk. This cycle: "500 gigawatts" of large-load requests, "more than 5 times" Texas peak. Batteries hosting 100 MW with 80 MW. Public opposition at "60%."
  • Micron: reported, but no podcast reaction yet. Last issue: about 6.5% implied move and the "86%" gross-margin bar. This cycle: about 8% implied, "87%" adjusted margins as the baseline, a "70%" operating-margin bull case and "15% to 20%" Q4 HBM pricing. The actual results will be covered in tomorrow's issue.
  • Open items carried over: Oracle rating action; Nvidia receivable days; Meta's Beignet SPV; the unreconciled Nvidia FCF figures ($96–98 billion versus $200 billion).

Next catalysts to watch: Micron CEO Sanjay Mehrotra on CNBC, 9 a.m. ET today (Thu Oct 1), and the stock's first full session after the print. Jobs report Friday, Oct 2 ("working our way to jobs Friday," Squawk on the Street, 11AM hour, Sept 30), which matters for the 10-year. Anthropic's official S-1. OpenAI's "$30 billion" round. Hyperscaler Q3 earnings, late October. The November Treasury refunding (carried over).