They bought a place in the queue

Anthropic reportedly paid $3.4 billion for someone else's spot in Caterpillar's order book, a deal that became public in the same week the price of a frontier lab started coming down.

The most revealing AI deal of the weekend did not involve a model, a chip, or a single line of code. It involved a set of heavy gas generators that a small Texas company ordered from Caterpillar about two years ago and then could not afford to finish buying. On Sunday the Financial Times reported that Anthropic had taken those orders over, roughly 1.7 gigawatts of generating equipment for about $3.4 billion, along with the engineering and construction work needed to install it. The paper called it the largest known single purchase of power-generation equipment by an AI lab. Anthropic and the seller, OnePWR Solutions of San Antonio, both declined to comment, Reuters said it could not immediately verify the report, and nobody has said where the machines will end up. So hold it at the weight it deserves, which is one well-sourced report from one person familiar with the matter. Even at that weight, it tells you more about where the AI build-out actually stands than most of what the labs published last week.

What Anthropic bought, mostly, was time. The FT's account is that OnePWR placed the orders, needed financing to complete them, and went looking for a customer. Anthropic first planned to be that customer, buying electricity from OnePWR. At some point in late September it decided it would rather own the place in line. It did not buy the company, and none of OnePWR's staff moved across. It bought the commitments themselves, and the know-how to bolt them down. Piper Sandler, which advised OnePWR, had already logged a $3.4 billion sale of the firm's power-generation platform to an unnamed AI lab in September, which is the kind of quiet corroboration that makes the FT's sourcing easier to sit with. One account on X, Joachim at @nikoraus, put the whole thing in four plain sentences: OnePWR "ordered 1.7GW of Caterpillar generators 2 years ago and couldn't finance it. Anthropic bought its spot in line. That's how scarce power is now." A trader posting as @brent_e_trader came at it from the other end of the transaction, and was just as plain: "Caterpillar will be paid now."

The price of a spot in line

The arithmetic is simple enough to do on the back of a napkin, and it is worth doing. Three point four billion dollars spread over 1.7 gigawatts comes to about $2,000 a kilowatt, before fuel, before connection, before anything has spun. That is a striking number, even allowing that it covers installation work as well as the machines, and it only makes sense once you remember what the alternative is. Korea's Financial News, writing up the deal on Monday, laid out the queue Anthropic was skipping. Caterpillar's order backlog has reached a record of about $78 billion, up 96 percent on a year earlier. Revenue in its power-generation business rose 30 percent to about $3.4 billion, by that paper's figures. Some of its large engines are booked out to 2029, and turbines to 2030, and the company is spending $725 million to expand a plant in Indiana. Orders already sitting on a production schedule can arrive inside two years. A fresh order placed today might not. When the waiting list runs to the end of the decade, a two-year-old order is worth far more than its sticker price, and someone who placed one early and then ran short of cash is holding something a frontier lab will pay handsomely for.

The same paper made one caution that most of the excited posts skipped, and I think it matters. One point seven gigawatts is the combined nameplate capacity of the generators. It is not a measure of power Anthropic has secured and switched on, and it says nothing about where, when, or how much of it will run at once. Generators need gas, permits, sites, and people to install them. The deal gets Anthropic past one bottleneck, the factory queue, and leaves several others standing.

Why a lab is shopping for engines

Step back and the deal stops looking strange. Anthropic's amended IPO filing, published last week, listed roughly $895 billion in commitments, almost all of them for computing power and infrastructure, stretching out to 2045. Reuters had already reported in September that the company expected to spend more than $500 billion on AI infrastructure over the next ten years. A company that has signed up for that much compute cannot afford to discover, three years from now, that the buildings are finished and the grid connection is not. Utilities move at the pace of regulators and transmission lines. Equipment makers move at the pace of their order books. A lab moves at the pace of its next training run, and when those three clocks disagree, the lab either waits or starts buying the slow parts itself.

Anthropic is not the first to make that choice. xAI ran gas turbines at its Memphis site to get its cluster running quickly and is now facing a lawsuit over them. Alphabet agreed to buy the energy developer Intersect for $4.75 billion. The pattern across the industry is that the frontier labs, which spent most of the past three years renting compute from cloud companies, are reaching further and further down the stack, past the chips and the buildings, toward the electricity itself. The FT framed this one as the biggest single equipment purchase yet. It almost certainly will not be the last, and the more interesting question is whether the next one is announced or simply turns up, months later, in a filing.

The week the price came down

What makes the timing so pointed is everything else that happened around it. The deal was signed in late September, when the mood in AI markets was close to its peak. It became public on the first Sunday after a week in which that mood broke.

Start with OpenAI. Bloomberg reported on Friday that the company had closed the first tranche of a new funding round, about $11 billion, at a valuation of $1.25 trillion, with the money wired by the end of the week. Its chief financial officer, Sarah Friar, confirmed the valuation on television and said the business grew 70 percent quarter on quarter in the third quarter and that ChatGPT now has 1.2 billion weekly users. SoftBank and Abu Dhabi's MGX took part. The round could reach about $20 billion over the coming weeks. It had originally been pitched, by most accounts, as $30 billion or more at around $1.4 trillion. That is still an enormous amount of money at an enormous price. It is also the first time in a long while that the headline number on an OpenAI round has come in under the ask, and the market noticed the gap at least as much as the size.

Then look at Anthropic's own listing, which is the thing the generator deal will now be read against. Until last month the company was reportedly aiming at a valuation as high as $2 trillion in its IPO; it was privately valued at $965 billion as recently as May. On Thursday and Friday the numbers started to slide. Bloomberg reported that prospective buyers were discussing a range of $1.2 trillion to $1.3 trillion, with some weighing a figure as low as $1.1 trillion. CNBC quoted an investor calling even $1.5 trillion "a stretch." Reuters said investors were pushing for a bigger discount, were waiting on the October revenue figure that Anthropic left blank in its updated S-1, and were openly asking whether the roadshow could still start this week. The FT reported over the weekend that Anthropic still wants to list before Thanksgiving on 26 November, raising $80 billion to $100 billion at somewhere between $1.25 trillion and $1.75 trillion. On the prediction markets Polymarket and Kalshi, the odds of a listing before 1 December fell from around 80 percent to near 50 percent, according to Yahoo Finance. Gil Luria of D.A. Davidson summed up the mood on the sell side with unusual bluntness: "The market really needs this IPO to go well and the bankers on the deal surely understand that. The best way to set Anthropic up for success is to price the deal well."

The reason for the slide is not mysterious. A punishing post-election selloff hit AI and data-centre stocks after the midterms, with fears that a new Democratic majority in Congress would slow the build-out, and with voters in dozens of places backing local restrictions on data centres. Nvidia had its fifth losing week in a row. And the money anxiety is reaching the demand side too. Bloomberg reported, via Reuters, that Apple has cut some engineers' daily AI spending allowance, the budget that pays for tokens across ChatGPT, Claude, and Gemini, from $300 to $200, and is considering $100, with some non-engineering groups already at $100. None of that means demand is collapsing. It means that for the first time this year, the people paying for AI are acting as though the bill matters.

Hard assets in a soft week

Put the generator deal back into that week and it reads differently from the way it would have read in September. A lab heading into the biggest listing in history, with its valuation being marked down in real time, has just committed $3.4 billion to the least glamorous asset imaginable. You can tell that story two ways, and I think both are true at once.

The first is that it is a sign of conviction. Anthropic is telling the market, without saying a word, that it believes the demand for Claude is real enough to justify owning the power plant, not just renting the server. Equipment like this does not lose its value in a selloff the way a forward revenue multiple does. If the AI trade cools for a year, a two-year head start on Caterpillar's queue is still worth something, to Anthropic or to whoever it sells the slot on to. In a week when the price of the company is being argued over, the price of the engines is not.

The second is that it is a sign of how heavy the business has become. A frontier lab used to be a research team with a cloud contract. Now it is a research team with a cloud contract, $895 billion in long-dated commitments, a utility-scale procurement operation, and an installation crew it had to buy along with the kit. Every one of those things is a fixed cost that has to be paid whether the next model lands well or badly, and every one of them is a reason public investors are asking harder questions about price. The generators are a hedge against the grid. They are also another line on the list of things that make Anthropic look less like a software company and more like an industrial one, valued, for now, like a software company.

It is worth noticing, too, how differently the two biggest labs are talking about the physical build-out in public. OpenAI spent the week after the midterms leaning into reassurance. Friar said the company would keep expanding its US data centres and keep building relationships with the communities that host them, after OpenAI pledged last week to stop asking public officials to sign confidentiality agreements during site talks, part of a broader community compact it published on Thursday. Anthropic's version of the same move came out through a newspaper, about a deal signed six weeks earlier, with no comment and no location. Neither approach is wrong. But a lab that is about to ask the public markets for up to $100 billion is going to be asked where its power comes from, and "a person familiar with the matter" will not be a satisfying answer on a roadshow.

What to watch

The useful questions now are concrete ones. Does the deal show up in Anthropic's next S-1 amendment, and does it come with a location? Does the roadshow actually start this week, and does the October revenue line get filled in before it does? Does OpenAI's round get to its $20 billion, and at what price does the last tranche close? And does anyone else in the industry go looking for the next stranded order in Caterpillar's book, now that the going rate for one has been made public?

The other shadow over the week sits with the company that supplies almost everyone. Nvidia's board has opened an independent investigation into Jensen Huang's ties to a customer executive whom US officials believe passed information to Beijing, and three House committee chairs want him to brief Congress on 18 November, the day after earnings. The board says it has no reason to believe any secrets were compromised. It is a separate story with its own facts, and I am not going to stretch it to fit this one. But it does mean that the three things a frontier lab needs, the money, the chips, and the power, all had their weakest-looking week of the autumn at the same time, and the only one that got more certain was the one you can drop on a concrete pad.

That is the part I keep coming back to. For two years the story of AI has been told in parameters, benchmarks, and valuations, all of which are numbers that can be revised over a weekend. The number that mattered on Sunday was a place in a factory's queue, bought from a company that ran out of money before its turn came up. Anthropic will spend the next three weeks trying to persuade investors what it is worth. Whatever price they settle on, it has already paid $3.4 billion to make sure the lights come on.