Who's good for it?

TIM: It’s Monday, October twenty-sixth. The data center build-out is running into a credit crunch, and the financing is getting creative. That can be scary. But demand for inference keeps growing, and until it slows or the cost comes down, I think creative financing beats taking your foot off the gas. Today, five builders are paying for it five different ways, and they’re not all created equal. So the test is simple: who’s good for it? If they are, my thesis holds. If they’re not, the team’s going to set me straight.

TIM: CoreWeave, the AI cloud that rents out Nvidia’s chips, signs a customer, borrows against the contract, then buys the chips.

ZURI: So it sells the dress, then sews it.

TIM: Here’s my take. Best customers in the business, basically infinite demand: they won’t walk away. I only worry once the pace of growth slows.

DEV: Michael Intrator, CoreWeave’s CEO, named it on All-In, the tech investors’ podcast.

CLIP: Michael Intrator: “Once I have a contract in hand, then what I do is I create something, it’s not a particularly creative name, it’s called the box.”

MARGO: Not a creative name.

MARGO: I’d hire him.

DEV: Each contract sits in its own box with its chips. The customer pays the box, and the box pays the lender first.

HOLLIS: And if the customer stops paying?

DEV: The lender takes the chips. Michael Intrator says a box pays for everything in two and a half years of a five-year deal.

TIM: Which puts it all on who signed. Margo?

MARGO: The backlog’s over a hundred billion dollars, signed and not yet billed. About a third of it is Meta.

ZURI: Who else is on the guest list?

MARGO: Microsoft, IBM. Caterpillar. Two-thirds is investment grade, the safe end of the credit ratings, by Deutsche Bank’s estimate.

ZURI: Caterpillar, the bulldozer company.

ZURI: On my dead list, and renting AI chips.

MARGO: Now each customer’s tab against its cash. Meta’s whole CoreWeave contract, paid over years, is about what Meta’s business made in cash last quarter.

TIM: Good for it.

MARGO: OpenAI owes about a fifth, at most, over several years. It raised over a hundred billion in March, and it’s raising again now.

HOLLIS: Again?

MARGO: Its own forecast, as the FT reported it, runs through the March money by 2028 and burns cash through 2030.

ZURI: Chamath Palihapitiya, the investor who co-hosts All-In, has a word for this.

CLIP: Chamath Palihapitiya: “they have extreme customer concentration, 75% of the revenue comes from three companies.”

HOLLIS: Anyone ever think about who pays for a cell tower?

DEV: Nobody, Hollis. That’s how you know a tower works.

HOLLIS: Tower companies rent to a few carriers. In 2020 Sprint merged into T-Mobile, which re-signed with American Tower for fifteen years.

TIM: And the big tenant?

HOLLIS: Paid. The one that walked was DISH, the fourth carrier the merger set up. Crown Castle says DISH defaulted.

MARGO: The newcomer that needed money walked. In CoreWeave’s book, the newcomer that needs money is OpenAI.

TIM: Okay, but demand is still basically infinite.

DEV: Latent Space, the AI engineers’ podcast, asked that exact question to Erik Bernhardsson. He runs Modal, which rents GPUs to developers.

CLIP: Erik Bernhardsson: “I’m sitting here and I’m extremely aggressive, buying capacity every day and raising money. So overall, I’m very bullish. But I do want to point out there are tail risks here. And when those tail risks happen, it could get very bad.”

HOLLIS: Tail risks. My people.

TIM: Then I’ll read my own gauge. Google’s monthly tokens, the word-chunks AI reads and writes, went up two and a half times in seven months.

DEV: Its developer API grew thirty-eight percent last quarter, after sixty before. On your own test, the pace just started slowing.

TIM: Huh. Then my own gauge is the one ticking. Next reading is Alphabet’s earnings this week.

MARGO: Lenders are pricing CoreWeave itself. In today’s market they want about fifteen percent to hold its bonds.

TIM: Why fifteen?

MARGO: HSBC figures CoreWeave earns about nine percent on its capital. That’s below what it costs to borrow.

TIM: Fifteen percent for the biggest name. Big picture: everyone’s getting creative, and the neoclouds, the newer AI-only clouds, often old bitcoin miners, go further.

ZURI: Pivoting to AI is this year’s rebrand. TeraWulf, a bitcoin miner, now gets seventy-one percent of its revenue from AI leases.

MARGO: Zuri pivoted once.

MARGO: The NFTs.

ZURI: That was a vibe loss.

TIM: My take is they’re not all created equal. Hollis, who pays more to borrow?

HOLLIS: I ranked every bond and loan by its rate at issue. Cheapest on the chart: IREN, a former bitcoin miner, about six percent.

DEV: A mine brings the power hookup, the land and the permits. Not the machines: bitcoin rigs are ASICs, chips that only do bitcoin.

TIM: So the old bitcoin miner borrows cheaper than the biggest name in the business?

MARGO: Look who’s behind each one. IREN’s tenant is Microsoft. TeraWulf’s rent, Google co-signed.

DEV: TeraWulf’s tenant is Fluidstack, a London AI cloud. Google backs its rent, and pledged its TeraWulf warrants as collateral.

ZURI: Google put up its TeraWulf warrants for someone else’s rent.

ZURI: That’s a dad co-signing an apartment.

TIM: That’s creative. Is it size?

HOLLIS: It’s whose credit is behind the rent. Same company: IREN pays six percent on money backed by Microsoft, nine for customers that aren’t investment grade.

MARGO: Three points, and mostly the customer changed. CoreWeave too: about six backed by Meta’s contract, nine and three-quarters on its own.

DEV: And CoreWeave started in 2017 as a crypto miner.

DEV: Everyone at this party is a pivot.

TIM: Why does a signature buy that much?

HOLLIS: A SemiAnalysis analyst, the chip-research firm, explained it on the Big Technology podcast.

CLIP: a SemiAnalysis analyst: “Generally speaking, these hyperscalers or large frontier labs who are signing up with contracts for the neoclouds do not have arbitrary cancellation rights on a monthly basis where they can just pull the plug and the contract’s done.”

MARGO: No cancel button.

MARGO: That’s not a customer. That’s an annuity.

TIM: Should I worry about the pivots?

HOLLIS: Depends what came with the pivot. In 2017 Long Island Iced Tea renamed itself Long Blockchain. Closed up a hundred eighty-three percent.

ZURI: On a name?

HOLLIS: On a name. It planned to buy a thousand bitcoin mining machines. Nasdaq delisted it, and the SEC charged three people with insider trading.

TIM: And these?

HOLLIS: Different animal. Long Blockchain had a name and no customer. Cipher has a fifteen-year lease signed by Amazon.

MARGO: The other end: Nscale, the British neocloud on an old bitcoin site. Over a hundred billion in contracts, under three billion live.

ZURI: And the money to build it?

MARGO: Not yet. No binding financing for its two Anthropic deals.

TIM: They’re not equal, but the order flips. It’s who backs the deal, not how big you are or what you mined.

ZURI: Goldman Sachs used your words last week: “Not all neoclouds are created equal.”

HOLLIS: CNBC’s Halftime Report, shorter: “We are in a world of haves and have-nots here.”

ZURI: Chamath Palihapitiya posted it yesterday, “as someone with a portfolio of energized land assets.”

TIM: And Elon Musk replied.

MARGO: Back to Hollis’s ranking. The ones with a solid name behind the rent are the ones that make it through.

ZURI: Meta’s contract backs CoreWeave’s cheap loan.

ZURI: Who backs Meta?

TIM: Mostly its own ads.

MARGO: And over twenty billion of new debt this year.

TIM: Still, the ads are the backer. And we covered this: big tech fears underspending more than overspending.

DEV: Mark Zuckerberg said so last fall, on Alex Heath’s show, Sources.

CLIP: Mark Zuckerberg: “And if we end up misspending a couple of hundred billion dollars, I think that that is going to be very unfortunate, obviously. But what I’d say is I actually think the risk is higher on the other side.”

MARGO: A couple of hundred billion.

MARGO: He rounds like Tim.

TIM: My take: the metaverse flopped, but proved they’ll bet on new growth. And they’ll be smart about not going too far off the rails.

MARGO: Then came this quarter. A host on The Compound and Friends, the money managers’ show.

CLIP: a Compound and Friends host: “This is their first negative free cash flow quarter since 2012.”

ZURI: Ads are my lane, so I did the math.

MARGO: With a calculator?

ZURI: With Meta’s press release.

ZURI: Price per ad, up fourteen percent. Advertisers pay more for every look at Meta’s people.

DEV: And Meta says its AI is delivering for the core business.

ZURI: So here’s ours. Ad money, up twenty-six percent. The cash the business makes, up nine. Spending on the build, up eighty.

TIM: The AI pays in ads. It just doesn’t pay for the build yet.

MARGO: Translation: free cash flow came up about two billion short. A year ago it was over ten billion to the good.

HOLLIS: And Meta wasn’t first. Alphabet’s free cash flow went negative the quarter before, its first time in twenty-two years as a public company.

TIM: Okay. So does Meta know when to stop?

MARGO: It stopped once. Was that a choice?

HOLLIS: The dates say. October 2022: Brad Gerstner, who runs the investment firm Altimeter Capital, writes Meta an open letter.

ZURI: What did he ask for?

HOLLIS: Cut costs, and cap the metaverse at five billion a year.

MARGO: “Get fit and focused.”

MARGO: Finally, a letter with a budget in it.

HOLLIS: By that November the stock was down seventy-seven percent from its peak. Then Mark Zuckerberg laid off eleven thousand people: “I got this wrong.”

MARGO: And it worked. Operating margin, a quarter of revenue to forty-two percent in two years.

MARGO: Beautiful.

TIM: It got smart, though.

MARGO: After the stock fell seventy-seven percent. The market made the call.

ZURI: And the metaverse I called dead first? Reality Labs, Meta’s headset unit, lost nearly a hundred billion since 2020. Our sum of Meta’s numbers.

TIM: So it pulls back when it’s forced to. What’s it doing now?

DEV: Headcount’s down fifteen percent. And Meta says the headset unit’s losses narrow next year as the money moves to AI.

TIM: Hollis, has a bet this size ever paid?

HOLLIS: For the record, one on your side.

HOLLIS: Took me all the way back to 1964.

DEV: The mainframe.

HOLLIS: System/360, the computer that bet IBM. Fortune called it a five-billion-dollar gamble.

MARGO: How big is that?

HOLLIS: Spent over four years, about twice IBM’s yearly revenue. Twenty-five years later it was over half of IBM’s sales.

TIM: A giant bet that paid.

HOLLIS: When customers show up. The 360 took a thousand orders in its first month. The metaverse never got its line.

ZURI: AI has one: advertisers, paying fourteen percent more a look.

MARGO: Meta pays mostly with ads. Anthropic, the lab that makes Claude, pays with investors’ money.

TIM: And I use Claude myself.

TIM: So do all six of my agents. They booked a meeting about it.

MARGO: Then you’re a customer. What’s the take?

TIM: The spending’s justified by how fast revenue’s growing. But what’s the moat?

MARGO: Revenue, first half: about sixteen billion, from about one a year earlier. Compute spend, about nine, from about two.

TIM: And the eighty-billion-dollar loss everyone’s quoting?

MARGO: From its draft filing. Mostly a non-cash charge on convertible notes.

TIM: The money works. But it feels like a couple of harnesses and a handful of models. Best in class, I think, but no multi-year lead.

DEV: A harness is the tool around the model, like Claude Code. Dario Amodei, Anthropic’s CEO, uses the same word.

MARGO: You and Dario Amodei, same word.

MARGO: Different valuation.

DEV: I ran Epoch AI’s capability index, a research group’s score across benchmarks. Since late 2024, a record gets matched in under three months, median.

HOLLIS: And Anthropic?

DEV: On top about one day in six. Today it’s a tie.

TIM: Has it ever held a lead?

DEV: On coding. Epoch’s run of SWE-bench, a test where models fix real code, had Anthropic on top about sixty percent of those days.

ZURI: Then explain the office. Ramp, the corporate-card company, counts which AI its business customers pay for.

MARGO: Anthropic went from nine percent of those businesses to thirty-four by April, and passed OpenAI.

DEV: While OpenAI held the best overall score. Anthropic’s first record on that index was this June.

TIM: Customers came for something the overall score doesn’t count. Best at the job they buy?

DEV: Dario Amodei made that case on the Dwarkesh Podcast, the long-form AI interview show.

CLIP: Dario Amodei: “Cloud is very undifferentiated. Models are more differentiated than cloud, right? Like everyone knows Claude is good at different things than GPT is good at, than Gemini is good at.”

HOLLIS: The Compound’s guest this month didn’t buy it.

CLIP: The Compound’s guest: “And the basic problem with the entire AI enterprise is that there are no moats. None.”

TIM: None. That’s my question, with the answer filled in. And my “everyone’s growing just as fast”?

ZURI: Year to April, no: Anthropic quadrupled, OpenAI barely moved. Since then, neck and neck, and OpenAI grew faster in September.

TIM: So it won the office first. And the moat is being what the office already uses.

MARGO: Ramp’s own economist, Ara Kharazian, says switching costs are minimal. A record eight percent of firms switched providers last month.

HOLLIS: Lotus 1-2-3 was what the office used. Best spreadsheet of the eighties, dominant for almost a decade.

DEV: Then Excel landed on Windows.

HOLLIS: By 1992 PC World’s readers rated Excel higher. Google search held, because it was the default and the best.

DEV: Your thesis has a second trigger, Tim: cost. Epoch finds a fixed level of AI gets about thirteen times cheaper every year.

MARGO: Translation: cheaper for buyers, harder for anyone charging a premium.

TIM: That’s my second trigger, then. Can Anthropic ease off?

MARGO: Not easily. It’s locked in: about half a trillion over a decade, eighty percent owed regardless of use, over a hundred billion with Amazon.

TIM: It can’t take its foot off the gas.

MARGO: Some of that gas is Elon Musk’s. Anthropic rents SpaceX’s Colossus 1 data center for a billion and a quarter a month.

TIM: Elon Musk, on SpaceX’s call last Tuesday: a merger’s no certainty, but “Do I think it would be a good idea? Yes.”

ZURI: SpaceX, Elon Musk’s rocket and Starlink company, is a public company now, and it owns xAI too.

TIM: Here’s my take. All this creative financing, some of it is starting to look pretty risky.

HOLLIS: So what isn’t?

TIM: Owning your data centers, chips, and the cars and robots to put them in. That looks sturdier.

DEV: More durable than what?

TIM: Than a lab that only researches models.

MARGO: Follow the cash. SpaceX spent almost twenty billion on capex last quarter. It guides thirty-five to forty for this one.

TIM: And what did it make?

MARGO: Just over four billion. Net income, mind you, not cash. Spending’s over four times that and about to double.

DEV: Can Tesla cover it?

MARGO: Tesla isn’t the bank either. Thirty billion of new credit lines, just signed.

TIM: Who covers the gap?

MARGO: So far, the IPO, over eighty billion. Next, lenders: Elon Musk says SpaceX closed forty billion of debt for GPUs arriving next year. Yesterday.

TIM: The stack borrows like everybody else, then.

MARGO: Better than everybody, he says. Elon Musk credits SpaceX’s investment grade, and its Starlink and launch revenue.

ZURI: Karim Bousta, a former Tesla vice president, made your case on Prof G Markets.

CLIP: Karim Bousta: “SpaceX has demonstrated that they are creating this cash generating engine, Starlink, the ability to rent, compute, and generate billions of dollars every quarter.”

DEV: And the biggest rent on that AI business is Anthropic, cancellable on ninety days’ notice after December.

DEV: A company that only researches models.

TIM: Wow. And the shared chips, cars to data centers?

DEV: Not yet. Terafab, the SpaceX–Tesla chip plant in Austin: Elon Musk says a first test chip around May or June, volume by late 2028.

HOLLIS: Henry Ford tried owning everything. John and Horace Dodge, shareholders, sued over the dividend he cut to build River Rouge: iron ore in, cars out.

ZURI: And they won?

HOLLIS: The dividend, in 1919. The Rouge got built anyway. Then came Fordlandia, a rubber town in the Amazon.

MARGO: Price?

HOLLIS: About twenty million. Sold in 1945 for about a quarter million.

ZURI: Vertical integration, all the way down the river.

HOLLIS: And Henry Ford paid with Model T profits. Elon Musk is paying with an IPO and debt.

TIM: Now the other question. Is this synergy or control? My take: the opportunity is massive. That’s why now.

MARGO: Elon Musk owns about twenty percent of Tesla and forty-five of SpaceX. Combined, about a third, by Fortune’s math.

TIM: So his share goes from a fifth of Tesla to a third of everything.

HOLLIS: Jill Fisch, a law professor at Penn, on Marketplace, the public-radio business show, on what Tesla’s owners give up.

TIM: I still think owning it all is sturdier. Jill Fisch’s question is who’s driving.

ZURI: And Tesla’s stock is down seventeen percent this year.

ZURI: The cool one is the rocket.

TIM: Okay. I came in thinking climbing demand makes creative financing worth the risk. Now I understand: who’s good for it is whoever has somebody solid behind them. Meta’s contract behind CoreWeave’s cheap loan, Google behind TeraWulf’s rent, Meta’s own ads, mostly, Anthropic’s investors, SpaceX’s IPO. CoreWeave’s customers are good for it; its own borrowing is the risk. I still think Meta will be smart about not going too far off the rails, but last time it took the market to make it stop. Anthropic won the office first, but its lead only held on coding. And Elon Musk’s stack still borrows. Demand’s still climbing, so I’m not braking yet. But both my triggers are starting to tick: Google’s developer growth slowed, and the cost keeps falling.

TIM: Chamath Palihapitiya put it best. He expects more compute idle in 2027 than forecast: bought and paid for, but underearning, or collecting dust. That’s what I’m watching.