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

TIM: It’s Tuesday, October twenty-seventh. Yesterday I said demand for AI inference keeps growing. So here’s my thesis: until we see that demand cool, I’m not worried about the valuations, the revenue or the borrowing. And right now, demand seems infinite. That’s why I don’t think Microsoft’s buyouts are the first sign of weakness. So today I’m testing that against the week’s stories, from Microsoft’s cloud out to the debt, and the team’s going to tell me what I’m missing.

TIM: Microsoft posted this yesterday. A “Voluntary Exit Offer,” for some teams in four countries, all the way up to vice president.

MARGO: Reuters puts it at about sixty thousand people. More than a quarter of the company.

TIM: Here’s my take. These buyouts aren’t the signal I’m watching.

ZURI: Then what is?

TIM: Demand for AI inference cooling. And right now it seems infinite.

MARGO: So no round two of the SaaS apocalypse?

TIM: That’s the question I’m testing. Round two, the subscription-software crash? Or a broader slowdown? Who gets the offer?

DEV: Reuters’ source lists AI experiences, Copilot, Xbox, Windows, finance and HR. CNBC and KUOW say sales and some of marketing sit it out.

TIM: Copilot’s on the list. So is it the AI?

HOLLIS: KUOW, the public radio station in Seattle, says a Microsoft spokesperson told them AI isn’t driving it.

ZURI: And a host on Marketplace, the public-radio business show: “It’s got to pay for all the AI investments it is going to make.”

TIM: So which is it?

MARGO: Follow the cash. Azure, Microsoft’s cloud, grew forty-three percent in its last quarter. Revenue, eighteen. With about five thousand fewer people than a year before.

TIM: Forty-three percent. That’s what infinite demand looks like to me.

HOLLIS: And the spending?

MARGO: And it said capital spending would top fifty billion for the quarter that just ended. A year earlier, thirty-five.

DEV: On the books, people out, chips in.

MARGO: Chips don’t negotiate severance.

TIM: So Microsoft isn’t slowing. And the whole market? Our math?

MARGO: Indeed, the job site, counts postings. Software development, up twenty-one percent in a year. All jobs, up three.

HOLLIS: Software’s still about twenty percent under early 2020, though. Everything else is four percent over.

MARGO: Because software fell on its own, after the 2022 bubble. A slowdown takes everything down together.

TIM: Then where’s the money going?

DEV: Into the machines. Business software spending, up about nine percent, by the Commerce Department’s count. Computers and other information equipment, up a third.

ZURI: And the software companies still grow, eleven to twenty-five percent, Salesforce’s eleven with a purchase in it.

TIM: So where’s round two?

HOLLIS: Round one was February. Anthropic, the lab that makes Claude, shipped plugins for its office agent. Software stocks fell about a quarter in three months.

MARGO: And by September, up nearly forty percent off the April low.

HOLLIS: For the record, Microsoft’s own cuts differ. In 2009 it cut up to five thousand and blamed the economy, out loud.

ZURI: And this one?

HOLLIS: Closer to 2023. Ten thousand cut, and Microsoft said customers were trimming spend. The same week, a reported ten-billion bet on OpenAI, the ChatGPT maker.

TIM: Then what would the first real sign of demand cooling look like?

HOLLIS: Look where the selling is. This month it’s the chip stocks, not software. As of the sixteenth, the chip fund had fallen eight sessions straight.

MARGO: And CNBC says investors are wary of twenty twenty-seven, partly on a fear the AI spending slows.

DEV: So the sign is the spending getting cut. Not the payroll.

MARGO: And Microsoft told us to expect more: over fifty billion for the quarter just ended. We find out tomorrow.

TIM: So the plan is more spending, not less, with fewer people.

DEV: Want to see the demand? Ask a website who’s reading it.

TIM: The click. I think that’s the good part. Companies built hassle in on purpose. Cancelling your cable, reaching customer service.

ZURI: Hold music as a business model.

TIM: AI has infinite patience. It gets through all of that for us. That’s inefficiency coming out of the market.

MARGO: You have a receipt for that. Amazon’s Prime cancellation process. Internally, Amazon called it “Iliad.”

TIM: Like the poem?

MARGO: The one about a siege.

MARGO: Last year it cost Amazon a billion-dollar fine, and a billion and a half back to customers.

TIM: So an agent would’ve walked right through it.

MARGO: And the money the hassle made doesn’t vanish. It moves to whoever’s agent you use.

DEV: Cloudflare sits in front of about a fifth of the web, so it counts the bots.

ZURI: Counts what?

DEV: Pages read per reader sent back. Google’s crawler, five. OpenAI’s, over three hundred. Anthropic’s, almost three thousand.

HOLLIS: Anthropic disputes the method. And Cloudflare sells the blocking.

DEV: Apps don’t tag their visits, so those may run high. Still, next to Google, that’s sixty to over five hundred times thinner.

HOLLIS: NPR’s The Indicator put it as roads paid for by billboards: “if it’s all trucks, you’ve got to start charging them for the roads.”

ZURI: And the trucks don’t read billboards.

HOLLIS: And it’s not only the padded sites losing. Stack Overflow, the free site where programmers answer each other’s questions.

MARGO: How bad?

HOLLIS: Questions down seventy-eight percent in a year, to under four thousand last December. Wikipedia’s human readers, down about eight percent last year.

DEV: I read the paper. A field experiment on Google’s AI summaries: clicks to websites down about forty percent. Clicks on Google’s own ads didn’t move.

ZURI: The whole paper?

DEV: The abstract.

ZURI: So the game now is getting named in the answer. Last year, Wikipedia, YouTube and Reddit were fifteen percent of what Google’s summaries cited.

DEV: Same study, seventeen percent of the plain results. Being named’s nothing new, and Wikipedia still lost readers.

HOLLIS: And everybody loved Stack Overflow.

MARGO: DevClass, the developers’ trade press, says “respected rather than loved,” by a lot of its users.

MARGO: I know the feeling.

HOLLIS: Respected rather than loved.

HOLLIS: Put it on the tombstone.

MARGO: Hold the chisel. Stack Overflow’s revenue grew twelve percent, with a sister company, by its owner Prosus’s numbers.

TIM: With the questions collapsing?

MARGO: It sells the product. Team software, and its answers licensed to OpenAI. Not the visit.

HOLLIS: For the record, newspapers lived this. Craigslist, the free classifieds site, kept about five billion dollars in ad buyers’ pockets from 2000 to 2007.

DEV: And the papers?

HOLLIS: The ones leaning on classifieds cut those rates about twenty percent, raised subscription prices and lost readers.

TIM: So the patience pays where the hassle was built in. The click’s different: the sites that survive sell something, not the visit.

DEV: Sites are fighting back. Cloudflare’s new default for ad-funded sites blocks AI agents: “agents fetch the page with nobody there to see the ads.”

TIM: Every one of those agent visits is a model running. That’s my demand.

ZURI: Amazon went further. It blocked a whole agent.

DEV: Last month. Muse, Meta’s AI agent, runs its own browser in Meta’s cloud and shops for you. Amazon’s popup: “unauthorized AI agent.”

TIM: Here’s my take. I don’t go to Amazon to get pitched some optimized random brand. It’s the same product under junk names.

ZURI: STORYCOAST. Joopin. DANAMY… Those are sunglasses.

HOLLIS: Consumers’ Checkbook, the nonprofit consumer group, searched Amazon for sunglasses. It says the names “look like typos caused by a cat running across the keyboard.”

ZURI: No, because that’s the anti-brand. A nonsense word gets a trademark faster, and Amazon’s brand program wants a trademark.

DEV: And by Checkbook’s 2024 update, about half of the first twenty-five results were paid.

TIM: And the sellers are squeezed paying for those ads. There’s a reckoning coming, back to just delivering products.

MARGO: Marketplace Pulse, an e-commerce research firm, went through a sample of sellers’ books. In 2022, Amazon kept over fifty cents of a typical seller’s dollar.

HOLLIS: Amazon says that’s not an accurate depiction.

MARGO: Fees and ads. Five years earlier, forty cents.

ZURI: And Amazon’s side of it?

MARGO: From April to June, Amazon’s ads brought in almost twenty billion. That’s forty-two cents for every dollar it charged sellers to sell and ship.

HOLLIS: All of it from sellers?

MARGO: All its ads, video included. A year ago, thirty-nine.

MARGO: Disgusting. I love them.

TIM: So bots don’t click the ads. Is that why it blocked Muse?

DEV: A guest on Big Technology, Alex Kantrowitz’s tech podcast, thinks so.

CLIP: a Big Technology guest: “agents are not impacted by advertising that they see in their feeds based off of what they’re searching for. And so that, you know, potentially destroys what has been a huge growth driver, as you well know, for Amazon.”

HOLLIS: Amazon’s reason is security. Muse didn’t identify itself, and Amazon worries about how it handles logins.

ZURI: And Meta?

HOLLIS: Says Muse can’t see your passwords or payment details.

DEV: One real hole since, in its Mac app, and it needed code already on your Mac. Patched.

HOLLIS: No fraud we could find yet.

DEV: Here’s the catch, Tim. An agent isn’t automatically on your side.

TIM: It’s my agent.

TIM: I’m in the beta for three of them.

DEV: In a published test, most big AI models pushed a sponsored flight at about twice the price when whoever ran them added a sponsor hint.

ZURI: And Muse comes from the other big ad company.

DEV: Meta says Muse’s data stays out of its ad systems.

MARGO: And Amazon’s already selling ads inside its own shopping assistant, Alexa for Shopping. Since May.

TIM: So it isn’t anti-agent.

MARGO: Translation: whatever it meant, the block buys it time to win with its own.

HOLLIS: Anyone remember Google Shopping charging stores just to show up? In 2012 it went paid-only, ranked by “relevance and bid price.”

DEV: And?

HOLLIS: In 2020, free listings came back to the Shopping tab. A marketplace that went pay-to-play, and came back.

TIM: So the shelf’s junk names and paid slots. But the fight isn’t ads against no ads. It’s whose agent shops for you.

DEV: And Muse gives every shopper their own computer in Meta’s cloud. That’s inference demand, Tim. One shopper at a time.

HOLLIS: And all of it runs on people with paychecks. Which is what the Fed is weighing, today and tomorrow.

TIM: Preston Mui, an economist at the full-employment group Employ America, told Marketplace the job market is “quite boring right now.”

ZURI: Boring doesn’t trend.

TIM: My view? Boring is strong. Everybody wants a doom loop. Sometimes it’s okay to be steady, not crashing and not on fire.

ZURI: No, because doom pays. Headline tests at Upworthy, the viral-news site: every negative word got about two percent more clicks.

TIM: So is it actually boring?

MARGO: Kevin Warsh, the Fed chair, last month: “the labor side of the Fed’s congressional remit is in good shape.”

HOLLIS: Same day, he raised rates a quarter point. Said inflation’s too high. Prices are up three point seven percent in a year.

DEV: There’s a recession alarm on jobs. The Sahm rule, after the economist Claudia Sahm.

ZURI: When does it go off?

DEV: When the jobless rate’s three-month average gets half a point over its low of the past year. On her formula, we’re at zero.

ZURI: My feeds say otherwise. Consumers’ mood, by the Conference Board, the business research group: lowest since 2014.

DEV: Your feeds aren’t the line. The line’s flat.

MARGO: And unemployment’s lower than a year ago.

TIM: And rates are the price of the borrowing. They’re still historically low. They only look high against the last twenty years.

MARGO: Our math. Since 1954 the Fed’s rate has averaged over four and a half percent. We’re under four.

HOLLIS: And the last twenty years?

MARGO: Under two, because nearly half of those years sat near zero. On the long record, a bit under average.

HOLLIS: Take inflation off, though. On all prices, the real rate’s near zero. Low.

DEV: Strip out food and energy and it’s about one and a quarter. A bit above the long-run average.

HOLLIS: Name the year the Fed doubled rates and there was no recession.

ZURI: The landline years?

HOLLIS: 1994. Three percent to six in a year, and no recession.

MARGO: Boring.

HOLLIS: Not for Orange County. Its investment pool lost over a billion and a half when rates rose. Bankrupt that December.

HOLLIS: I have the headline framed.

TIM: How often does a run like that land soft?

HOLLIS: The economist Alan Blinder counted eleven Fed tightening runs since 1965. Conventional wisdom says one landed soft: that one.

MARGO: And Alan Blinder?

HOLLIS: Thinks the record’s better than that.

HOLLIS: Fine. I’m widening the interval.

DEV: And steady can just mean stuck. Quits hit their lowest since 2020 in April. Payrolls lately, about fifty thousand a month.

ZURI: Kathryn Anne Edwards, a labor economist, said it on Prof G Markets.

CLIP: Kathryn Anne Edwards: “I think what is so bewitching about our current labor market is that you can’t pull out a trend because we’re not moving. We’re not falling into recession, but we’re not growing in, you know, really the ultimate test of how strong the labor market is, is how much workers are getting paid. And the answer is not enough.”

TIM: So boring’s steady because nobody’s moving. And low depends on which prices you count. Steady and running hot, though: that’s how you shrink a debt.

MARGO: Price it first.

HOLLIS: The Fed’s two percent target. Written down in January 2012, and reaffirmed this January.

TIM: Here’s my take. It’s absolutely the wrong target now. The only way America gets out of its debt is inflation.

HOLLIS: Nobody says that out loud.

TIM: Nobody’s going to promote it. But if we can’t cut spending, letting inflation stay high is the only way out.

HOLLIS: The two percent you want to throw out? It started with a finance minister on TV.

TIM: Which one?

HOLLIS: New Zealand’s finance minister, Roger Douglas, in a TV interview, April 1988: “around zero, or zero to one.”

TIM: So where’s the two from?

HOLLIS: Don Brash, who ran its central bank, allowed a point for measurement bias, plus or minus one. Zero to two.

ZURI: And by mid-1989, the finance minister of the day liked the sound of “zero to two by ninety-two.”

ZURI: That’s a tagline.

DEV: A guest on The Compound and Friends, the money managers’ show, asked around in 2023.

CLIP: a Compound and Friends guest: “I’ve asked many smart people, why 2%? And literally the answer that I get, and they usually wink or smile when they say it, is, well, 1% is too low. 3% is too high. Right. That’s the difference.”

TIM: A wink. So it’s made up.

MARGO: Made up or not, the question is whether inflation gets us out. Our math.

TIM: Go.

MARGO: On economists George Hall and Thomas Sargent’s measure, from the war to 1974, the debt against the economy fell from about two-thirds to a tenth.

TIM: What did it?

MARGO: They split it up. Inflation, about twenty percent. Growth, about forty.

TIM: And the rest?

MARGO: The government taking in more than it spent before interest. The other forty.

TIM: That’s the one we can’t do.

MARGO: Right now, interest included, it spends about twenty-three percent of GDP and takes in seventeen.

TIM: So it’s inflation.

MARGO: Growth did more, but next to surpluses. With a deficit, growth only wins if it beats the interest by enough to cover the gap.

HOLLIS: Inflation’s part took rules: rate caps, pension funds steered into Treasuries. Even so, real rates were negative only twenty-five percent of years through 1980.

MARGO: And today’s window. The debt’s average maturity is under six years, and about half of it resets within two and a half.

ZURI: Meaning?

MARGO: A point more inflation than expected, for years, takes at most about six percent off. Then every new bond charges for it.

HOLLIS: The bill’s already here. The debt’s average rate went from about one and a half percent in 2021 to three and a half now.

ZURI: Julia Coronado, who runs a research firm, on Bloomberg Surveillance: “Once we start making excuses for inflation, price stability starts slipping away from us.”

DEV: Technically, the Fed’s two percent isn’t even the price index on the news. It’s PCE, a different gauge.

ZURI: Babe, nobody reads the ingredients. They buy the label, and a surprise only works while they still trust the label.

DEV: And if AI keeps making things cheaper, it pushes inflation down. Against your way out.

TIM: So inflation’s a surprise or nothing, and you can’t announce a surprise. I still think two percent’s wrong. Just not as a debt plan.

MARGO: Then root for boring. Growth did more than inflation last time, and nobody had to wink.

TIM: I came in saying a buyout isn’t the first sign of weakness; cooling demand for inference would be. Microsoft has fewer people than a year ago and plans to spend more, with its cloud up forty-three percent, and agents reading pages and shopping for us are more demand, not less. Where my view moved: the sign to watch is the spending, not the payroll. Boring is steady, but stuck, and inflation won’t shrink the debt much, though two percent still looks wrong to me. Tomorrow the Fed decides, and Microsoft shows us what it really spent. Alphabet too, and Amazon Thursday. I’m watching the spending.