Pure Inference Daily

261030 30 October 2026 18:08

Microsoft's first US bond sale in over nine years. Why?

Microsoft showed it could pay for AI out of its own cash, then borrowed about $25 billion the next morning. Nobody on the shows we follow argued this week that people are using AI less; they argued about the bill. So I asked the team what separates the payers, and the answer is the price of the money: about a point over the Treasury for Microsoft, about fifteen percent for new money at CoreWeave, a slice of the company for Anthropic's new shareholders, borrowed money near ten percent behind OpenAI, and a Fed that only put its next hike off to December. Plus a correction to our October 22 episode, and a few things we said that held up.

Chapters

  1. 00:00 Run of show
  2. 00:25 Microsoft paid for its quarter in cash, then borrowed $25 billion. Why?
  3. 03:38 Microsoft offered buyouts to 27% of its staff. Is that AI, or pandemic overhiring?
  4. 07:24 Anthropic earned $2.08 per compute dollar, up from 40 cents. Is AI a real business?
  5. 10:55 CoreWeave's bonds trade at 15%. A warning of what?
  6. 14:20 The Fed held, and Warsh tied rates to the AI trade. Is there a Fed put under AI?
  7. 17:18 Close

Transcript

TIM: Wednesday, Microsoft showed Wall Street it could pay for A.I. out of its own cash. Thursday morning, it went to borrow about twenty-five billion dollars anyway. Nobody on the shows we follow argued this week that people are using A.I. less. They argued about the bill. So who’s paying it, and how? My bet: the market’s grading A.I. on who pays from cash and who has to borrow. It’s Friday, October thirtieth. Let’s find out.

ZURI: Wednesday night, Reuters: “Cash flow is king again.” Microsoft could fund its AI “without endlessly tapping investors.”

MARGO: And the quarter backs it. Azure, Microsoft’s cloud, up forty-eight percent. On Tuesday’s show we had it at forty-three.

TIM: Here’s my take. Everyone agrees it was a clean quarter. Where the shows split is OpenAI.

ZURI: A Halftime Report panelist flagged “earnings quality.” The ten billion dollars “tied to OpenAI.”

TIM: OpenAI, the company behind ChatGPT. If that’s a third of Azure, it’s one customer that has to raise or borrow its money. Margo, our math.

MARGO: Ten billion from OpenAI, against thirty-three billion of Azure. At most about thirty percent.

TIM: At most?

DEV: Read the fine print. TechCrunch: “including revenue-sharing payments.” Microsoft doesn’t say how much of it is Azure.

MARGO: It’s about eleven percent of all of Microsoft. And one quarter is forty percent of what OpenAI paid all last year.

ZURI: OpenAI pays Microsoft almost twice what Xbox brings in. The console is the side hustle now.

TIM: So the concentration’s real, and nobody knows its size inside Azure. But Microsoft paid for the build-out from cash.

MARGO: Fifty-eight billion in from operations, thirty-nine out on property and equipment, mostly chips and servers. About eighteen and a half left.

DEV: That’s the cash. Capex was forty-five billion, including seven of leased data centers, booked like a purchase.

TIM: On Tuesday we said Microsoft expected over fifty.

HOLLIS: Its own guidance. It came in about five under.

MARGO: Amy Hood, the CFO, said it plainly: “There’s a form of financing called leasing.”

DEV: And it rents overflow from the neoclouds, the chip-rental clouds. Gil Luria, the D.A. Davidson analyst, on Prof G Markets Thursday, on why staying cash-flow positive matters:

CLIP: Gil Luria: “The reason that’s important is that that means they don’t have to go to the capital markets.”

HOLLIS: Same morning, Bloomberg: Microsoft “is selling high-grade bonds for the first time in more than nine years.”

TIM: How much?

MARGO: About twenty-five billion. More than the whole quarter freed up.

TIM: So it can pay from cash, and it borrows anyway.

MARGO: Because it’s cheap for Microsoft. Triple-A, one of two US companies with that rating from both big agencies. Beautiful.

HOLLIS: Lucent, the phone-equipment maker, once played bank. In two thousand it committed almost six billion dollars of credit to its own customers.

DEV: To buy Lucent’s gear.

ZURI: So, a store credit card.

HOLLIS: One customer, Winstar, a wireless upstart, went bankrupt in twenty oh one and sued, saying Lucent broke its financing deal. Lucent called that without merit.

MARGO: And Microsoft isn’t lending OpenAI anything. It owns about a quarter of it.

HOLLIS: Fair. But revenue that rides on one customer’s fundraising stops when the fundraising does.

TIM: And how does OpenAI raise it?

MARGO: Mostly by selling shares. Over a hundred billion this spring.

HOLLIS: And one of its biggest buyers, SoftBank, the Japanese investment group, sold junk bonds at up to nearly ten percent to pay for its stake.

TIM: So it’s not cash against borrowing. Microsoft pays from cash and borrows anyway.

MARGO: At a very different price from OpenAI’s backers.

TIM: Right. OpenAI sells shares, and one of its biggest buyers borrowed near ten percent to buy them. Microsoft borrows near the Treasury’s rate.

ZURI: And somebody inside Microsoft is paying too. More than a quarter of the staff just got offered a buyout.

TIM: That’s the other half of the bill. AI or pandemic overhiring is the wrong split.

HOLLIS: Then what is it?

TIM: The junior work is already being absorbed by AI. But this offer is mostly AI’s bill, and payroll is paying part of it.

MARGO: Ed Elson, who hosts Prof G Markets, says if Microsoft cuts all twenty-seven percent, it ends up “lower than before the pandemic.”

TIM: I wanted to check that.

MARGO: Sixty thousand offered. If every one of them goes, Microsoft is back to its June twenty twenty size, exactly.

TIM: Not before the pandemic?

MARGO: June twenty twenty was the first pandemic report. June twenty nineteen was nineteen thousand smaller.

HOLLIS: Ed Elson’s other number checks out, though. Almost sixty thousand added in the two years after COVID.

MARGO: So a full take-up undoes the pandemic spurt. Not more. And it won’t be full.

ZURI: Business Insider: Microsoft expects “three thousand or fewer” to take it.

MARGO: Five percent of the offers. About one percent of the company.

ZURI: And the CFO’s memo is pure brand. “Agency.” “Growth opportunities.” It’s giving exit interview with a scented candle.

HOLLIS: That’s Microsoft’s guess. The spring retirement program got more than thirty percent, on richer terms.

MARGO: Even at that rate, Microsoft’s still bigger than it was in twenty twenty-one.

DEV: Ed Ludlow, Bloomberg’s tech reporter, had the mechanism: “Wall Street cares about free cash flow now.”

MARGO: Then size the belt. The whole spring program cost under a billion to run. The bonds raise about twenty-five.

DEV: On Tuesday I said people out, chips in. Three thousand out, twenty-five billion of bonds in. I undersold the chips.

TIM: So the chips are the big number by a mile. Then what about “AI’s bill, not AI”?

HOLLIS: Ed Elson answers that one too.

CLIP: Ed Elson: “After all, the more they spend on data centers, the less they might be willing to spend on employees. And, you know, that makes sense, that might be true, but I would respond with the following question. What’s the difference?”

TIM: Fair question. Dev, what’s AI actually doing there?

DEV: Agents now open more than a third of all pull requests on GitHub, Satya Nadella, Microsoft’s CEO, said Wednesday. That’s the whole platform, not Microsoft.

ZURI: And inside Microsoft?

DEV: About a fifth to a third of its code, by Satya Nadella’s count. Eighteen months ago.

HOLLIS: And who got the offer? Copilot’s team, yes. The people who run the data centers, no.

DEV: Same week, Amazon told its engineers the opposite story. The job isn’t “defined by the use of AI.”

ZURI: And: “If your work does not benefit from current AI tooling, that is a finding, not a failing.”

HOLLIS: Yesterday we had the FT’s report from May: some Amazon staff running agents on meaningless tasks to pad their token stats.

DEV: Now Amazon calls token counts a proxy, and says used as a target they can do harm.

HOLLIS: IBM tried buyouts first. It was a no-layoffs company into the nineties.

ZURI: A tech company that just didn’t fire people? That’s fan fiction.

HOLLIS: It overhired for “a late-1980s boom that never happened,” Newsweek said, and ran round after round of early-retirement packages.

MARGO: And?

HOLLIS: Too many good workers left, Newsweek said. Then in ninety-three, IBM’s new chairman, Lou Gerstner, announced thirty-five thousand more cuts.

ZURI: More?

HOLLIS: Fifty thousand had already gone that year.

MARGO: IBM was losing eight billion a quarter. Microsoft’s operating margin is about fifty percent.

HOLLIS: Fair. My worry is who leaves. A buyout picks the people with somewhere better to go.

MARGO: And the ones left make a lot. Revenue per employee is up about seventy percent since twenty nineteen, to about a million and a half.

TIM: Okay. Still the wrong split, and the bill lands somewhere I didn’t expect.

DEV: Payroll?

TIM: A sliver of this bill. The chips are the rest of it. Tuesday I said watch the spending, not the payroll.

ZURI: And the spending needs payers. Anthropic wants new shareholders. It filed to go public on Monday.

TIM: Here’s my take. Anthropic, the lab that makes Claude, says a dollar of compute brought in forty cents of revenue in early twenty twenty-five.

DEV: And two dollars and eight cents this spring.

MARGO: About five times.

TIM: While it cut the price of Opus, one of its premium models, by two-thirds. That says the product works.

ZURI: Halftime’s reporter on CNBC asked the obvious one: “can usage offset the price cuts?”

TIM: My question is whether that ratio covers training.

MARGO: And who pays the gap?

TIM: The new shareholders, in the slice they buy.

DEV: A guest on Big Technology said where the game is.

CLIP: a Big Technology Podcast guest: “Is their spend side. It’s not the revenue side. It’s the spend side. The efficiency is kind of the ball game for them.”

TIM: Spend side. Does that spend include training?

DEV: Yes. Research, internal use, all of it. Less than a third of the bill goes to serving customers.

HOLLIS: So training’s already in the ratio.

DEV: Yes. And each dollar spent serving customers earned about six times more. That’s the whole rise, and then some.

MARGO: Almost seven dollars of revenue per serving dollar. Disgusting. I love it.

DEV: Anthropic’s own explanation: compute bought a year or more ahead, with demand running ahead of it. Plus cheaper serving.

ZURI: So usage does move it. The chips were already on the bill.

HOLLIS: Then watch what happens when new chips land. Anthropic told investors the third quarter was a dollar seventy. Down about eighteen percent.

DEV: Compute up about three-quarters, revenue up forty-five percent. The S-1 warned margins “may compress” when new compute comes online.

MARGO: On the twenty-second we said about a dollar eighty. That was the half-year average. The quarters swung inside it.

TIM: So the ratio covers training, and it can slip. Who pays the gap?

MARGO: Here’s the gap. Its compute ran about forty billion a year in the third quarter. It’s signed for about a hundred.

ZURI: So revenue has to grow into it.

MARGO: At a dollar seventy a compute dollar, about a hundred seventy-five billion a year. The third quarter, annualized, was about sixty-seven.

TIM: And today it loses money.

MARGO: A correction, mine: on the twenty-second I called about nine billion the real first-half loss. That was press arithmetic, not the filing.

ZURI: So what does the filing say?

MARGO: Running the business lost about seven hundred million, and cash came in. The eighty-billion headline loss is mostly an accounting charge on its convertible notes.

DEV: And the IPO?

MARGO: If it raises the hundred billion Fortune expects, at the valuations reported, that’s about five to seven percent of the company.

HOLLIS: A hundred billion, once. The bill is about a hundred billion a year, by 9fin’s math. 9fin’s a credit-news service.

TIM: I said sixty on the twenty-second.

HOLLIS: From the Wall Street Journal’s look at a draft. The filing’s commitments are bigger.

DEV: And it’s already borrowing. Anthropic guarantees about a hundred sixty billion of chip leases with Broadcom.

MARGO: Plus convertible notes, mostly to Amazon and Google. And 9fin expects bonds after the IPO.

TIM: So it sells five to seven percent of itself, and borrows behind it.

MARGO: The IPO sells shares, and shares never come due. Tesla sold stock three times in twenty twenty, then turned its first annual profit.

HOLLIS: With regulatory credits doing the lifting. Without them, it lost money.

ZURI: And it leads. Menlo Ventures, which is an Anthropic investor, put it at forty percent of companies’ AI model spend last December.

HOLLIS: And about eighty percent of revenue is usage customers can cut “with limited or no notice.”

TIM: So training was in the ratio all along.

ZURI: So what decides it?

TIM: The hundred billion a year it’s signed for.

HOLLIS: And if it grows into that?

TIM: Every wave goes the same way: the creators get the glory, and the companies that put it to work get the G.D.P.

MARGO: Lenders charge some borrowers more than others. Ask CoreWeave.

TIM: CoreWeave, the cloud that rents out Nvidia’s chips from its own data centers. Eight shows argued about it in the past two weeks.

HOLLIS: Canary in the coal mine?

TIM: Only three of thirteen mentions even used the word.

ZURI: So what’s your read?

TIM: A warning about one financing model: chips that lose value, rented out on money priced at about fifteen percent in mid-October. Not demand.

HOLLIS: Alex Altmann, a Barclays strategist, on Odd Lots: “there is evidently a financing problem within neoclouds.”

DEV: And he says the selloff wasn’t about CoreWeave itself. Cipher, another neocloud, had tried to tap financing.

MARGO: First, fifteen percent is what new money costs. CoreWeave’s notes pay nine and a quarter to nine and three-quarters.

ZURI: And on everything it owes?

MARGO: About seven percent across all of it, the cheap Meta-backed loans included. The market wants fifteen to lend it more.

TIM: Compared with who?

MARGO: The sorting. Microsoft this week, at most about a point and a third over the Treasury. CoreWeave in mid-October, about ten.

TIM: Calm in the middle, choosy at the edges. We said that on the twenty-second. Now it’s measured.

MARGO: That’s the market doing its job. The weak borrower pays up, the strong one borrows cheap.

DEV: And the other junk borrowers?

MARGO: SoftBank, junk-rated, sold its bonds under ten percent in September.

TIM: The Compound had a name for it.

CLIP: a host of The Compound and Friends: “that’s the AI bond market vigilantes taking the pricing into their own hands and delineating between good and bad risk.”

HOLLIS: He’s riffing on a note from Ed Yardeni, the Wall Street strategist.

TIM: And chips that lose value?

DEV: Well, actually, not this month. Silicon Data’s index has Nvidia’s B200 renting up about twelve percent in a week.

TIM: One week’s rent.

DEV: And the old H100 is still under three bucks an hour, like we said on the twenty-third. Rising, not falling.

ZURI: Says the man with a storage unit full of twenty twenty-one graphics cards.

DEV: Those are vintage.

HOLLIS: And CoreWeave says it keeps signing new capacity “at higher prices.” On contracts of three to six months, though.

MARGO: Meanwhile, in the June quarter its interest bill was five times its adjusted operating profit.

HOLLIS: Anyone heard of New Century? A big subprime mortgage lender. Bankrupt in April two thousand seven.

DEV: And the big banks?

HOLLIS: Seventeen months later Lehman failed, and Citigroup needed government support.

MARGO: I sat on a credit committee for the other ending. The oil bust, twenty fourteen to sixteen. Oil fell about sixty-five percent.

DEV: And the banks?

MARGO: JPMorgan’s worst case was under three billion, on forty-four billion of energy loans.

TIM: So it came down to who held the loans. Who holds CoreWeave’s?

MARGO: Bondholders, and banks with secured loans. JPMorgan’s analysts call the banks’ exposure “very well secured.” They also rate the stock overweight.

ZURI: And customers still sign, because neoclouds are fast. Chamath Palihapitiya, who sells powered land to them, says hyperscalers “take 3 months to do an NDA.”

DEV: Microsoft’s one of those customers. Short-term contracts, Gil Luria said.

HOLLIS: And here’s how neocloud trouble can reach Microsoft. Amy Hood: “third-party capacity deliveries we expected in Q1 were delayed.”

MARGO: And Azure still grew forty-eight percent.

TIM: So far it reaches Microsoft as late chips, and the money stays with CoreWeave’s lenders.

DEV: And the rent’s still climbing.

TIM: So far, the market’s sorting them. No contagion. Same as Monday’s show: CoreWeave’s customers are good for it; its own borrowing is the risk.

HOLLIS: Unless the Fed moves everyone’s price at once.

TIM: Here’s my take. No Fed put.

ZURI: Babe, explain it like I’m one of your agents.

TIM: The idea the Fed cuts rates to catch a falling market. What it’s offering the AI trade is a hike held back, not a cut.

HOLLIS: Held back by two votes. Lisa Cook, a Fed governor, and Beth Hammack, who runs the Cleveland Fed, wanted the hike Wednesday.

MARGO: And the next meeting’s December ninth, with two votes for a hike already on the record.

TIM: So it’s not held back. It’s put off, maybe to December.

DEV: Alex Altmann again: there was an “almost universal view” that “AI was immune to higher interest rates. Evidently is not true.”

ZURI: Kevin Warsh, the Fed chair, Wednesday: “Interest rates get in the cracks.”

MARGO: And on tighter conditions: “That tells me that policy is working.”

TIM: That’s the opposite of a put.

HOLLIS: Jim Bullard, the former St. Louis Fed president, on Surveillance: the committee “would probably welcome some tempering in the AI boom.”

ZURI: Did he hedge it? Hollis-level?

HOLLIS: He opened with “I don’t know. I don’t think we know.”

HOLLIS: Respect.

DEV: So where did the Fed put start?

HOLLIS: Nineteen eighty-seven. The day after the crash, Alan Greenspan said the Fed stood ready “to serve as a source of liquidity.”

ZURI: And it’s been a thing since?

HOLLIS: It grew in nineteen ninety-eight. Long-Term Capital Management, a hedge fund with two Nobel winners as partners, nearly collapsed after Russia defaulted.

MARGO: The Fed cut three times in seven weeks. Six days after the last cut, the S&P five hundred closed at a record.

TIM: So that one’s a put.

HOLLIS: The Fed said it cut “to cushion” the economy from tighter financial conditions. Easing because markets tightened.

MARGO: And it said a lower rate was consistent with “keeping inflation low.” This week’s statement says inflation “remains elevated.”

TIM: So a nineteen ninety-eight cushion needs low inflation. And two thousand one?

HOLLIS: Yesterday we did two thousand one’s tech spending. Here are its stocks.

MARGO: Eleven cuts, almost five points. The Nasdaq still lost more than half from the day the cutting started.

ZURI: So where does AI even show up in inflation? The power bill?

DEV: A little. Electricity is up a bit over four percent a year. But gasoline drove over half of September’s rise.

ZURI: So the data centers aren’t the inflation.

DEV: Not through the power bill. Through spending. Kevin Warsh says capex is at “a historically high share of GDP.”

ZURI: And who’d feel an unwind first? The richest tenth hold eighty-eight percent of stocks and funds held directly, by the Fed’s count.

HOLLIS: Directly. Most people’s stocks sit in retirement plans, which that count leaves out.

ZURI: Fine. But the top tenth by income does about half the spending, by Moody’s estimate from twenty twenty-four data. Luxury feels it first.

MARGO: The flip side. Nick Timiraos, the Journal’s chief economics correspondent: an AI boom “that’s more rate sensitive means each increase does more work.”

DEV: Meaning?

MARGO: Maybe fewer hikes, because borrowers like CoreWeave feel each one more.

TIM: Okay. No put. A hike put off, and a chair who calls the cooling “policy working.”

HOLLIS: And your cushion for the economy, afterward?

TIM: The ninety-eight cushion came with low inflation. If one comes, it comes later than I thought.

MARGO: Which leaves the cheapest money in the room with Microsoft.

TIM: So who’s paying the A.I. bill? I came in grading it on who pays from cash and who has to borrow. Everybody’s borrowing, Microsoft included: it pays from cash and borrows anyway. What I didn’t see coming in: the divide isn’t who borrows. It’s what the money costs. What I can’t answer yet is what Microsoft finally paid. Margo?

MARGO: The final pricing on its bonds, once it’s public.

TIM: Then that’s my number. I’ll be reading the footnote.

Sources

Sources (clips are from each show’s own podcast feed) Run of show

Microsoft paid for its quarter in cash, then borrowed $25 billion. Why?

Microsoft offered buyouts to 27% of its staff. Is that AI, or pandemic overhiring?

Anthropic earned $2.08 per compute dollar, up from 40 cents. Is AI a real business?

CoreWeave’s bonds trade at 15%. A warning of what?

The Fed held, and Warsh tied rates to the AI trade. Is there a Fed put under AI?

Close

  • None beyond the segments above.
  1. 261102 2 Nov 18:42

    The economy's numbers are right. They answer a different question

    Jobless claims are near their lowest since the 1960s, and a lot of people still feel broke. So are we watching the wrong numbers? I took five of this week's headline numbers to my team and asked which number they'd watch instead. New college grads aren't vanishing from the job market, but the head start they had every year from 1990 to 2018 flipped in 2019, and Cisco's 27-year-old came in the side door. Credit did move before stocks, but the bond market's fear gauge is near its calmest tenth since 1973. Netflix's growth is its price rise, and viewers are cutting the other apps, not Netflix. Switching jobs still pays, half what it did in 2022; it's the person who stays whose raise is only about even with prices. And token prices are falling far faster than cloud storage ever did, but storage never went to zero. It went too cheap to watch.

  2. 261029 29 Oct 17:15

    Uber put a per-person cap on AI coding. That's how a boom matures

    My thesis today: AI is maturing, and a soft landing means it keeps maturing in a controlled manner, the growth curve bending gently instead of collapsing. So the team went looking for the controls. Companies are capping AI token budgets, but they're caps, not cuts, and the budgets arrived much further into the rollout than I thought. An agent could find a saver over $300 a year on every $10,000; for banks that's a margin squeeze, not a run, and the real barrier is the login. Google's new privacy policy lets connected Gmail shape ads, and its help page says summaries of your mail can train its models, if you switch it on. Anthropic's founders hold the same size vote-to-ownership gap as Meta's and Alphabet's, built very differently. And Google's back near the frontier, though its own list price doubles December 1.

  3. 261027 27 Oct 17:37

    AI demand still looks infinite. 60,000 buyout offers don't change it.

    Microsoft offered about 60,000 people a buyout while its cloud grew 43%, and I don't think that's the first sign of AI demand cooling. My thesis, picking up from yesterday: until demand for AI inference cools, I'm not worried about the valuations, the revenue or the borrowing, and right now demand looks infinite. The pod walked me through it: business spending on the gear AI runs on is up a third, software job postings are up 21% in a year, and bots reading the web and agents like Meta's Muse shopping for us are more demand, not less. Amazon's block of Muse is a fight over whose agent does the shopping. The job market is boring because nobody's moving, and boring is steady. And inflating the debt away works at most once: after the war, inflation did about a fifth of the work.

  4. 261026 26 Oct 18:52

    Who's good for it?

    I brought five reads on who pays for the AI build-out, and the pod set me straight, payer by payer. CoreWeave's customers are good for it; its own borrowing, at about fifteen percent, is the risk. The old bitcoin miners aren't equal, but what sets their price is whose name is behind the rent, not size: Nscale has over a hundred billion in contracts and under three billion live. Meta's AI pays in ads, not yet for the build, and last time it took a 77% stock drop to make it stop. Anthropic won the office first, but its lead only held on coding. And Elon Musk's stack still borrows like everybody else. Demand's still climbing, so I'm not braking yet, but both my triggers are starting to tick.

  5. 261023 23 Oct 17:28

    Permits, coders and a cowboy town

    Who gets the AI boom: Amazon's Nvidia racks waiting in warehouses, AI writing code while firms keep hiring coders, and Prineville, Oregon, the town that said yes to Meta's and Apple's data centers. The crew corrects all three of my yeses on air. The chips are a jam the buyers are making worse, my tractor theory is half proven, and Prineville's yes bought a tax base, not its comeback.