Last Week Ignite: 5.31.2026
The Week AI Started Borrowing
On the morning of May 28, most of the venture world read one number and stopped there. Anthropic, the company behind the Claude assistant, had raised sixty-five billion dollars. The price tag attached to the raise was nine hundred sixty-five billion. That figure is what people in the business call a post-money valuation, the worth of the entire company the instant after the new cash lands, the new cash included. Close to a trillion dollars for a company that did not exist five years ago.
That number traveled fast, and it earned the trip. It is the largest single equity round anyone has ever pinned to an artificial intelligence lab. If you wanted proof that the money still believes, there it was in nine digits.
I spent about twenty minutes on it, then moved on to the item that mattered more. It was smaller, it was quieter, and it ran underneath the headline like a current under a calm surface. Two of the biggest investment firms in the world, Apollo and Blackstone, were lining up roughly thirty-six billion dollars in loans. Not to buy stock in Anthropic. To buy computer chips for it.
Here is the shape of the deal, in plain terms. Apollo and Blackstone run what is called private credit, which means they lend money directly to companies instead of letting a bank stand in the middle. The thirty-six billion would buy a large supply of Google’s specialized AI chips, the processors Google designs in-house to run models like Claude. Anthropic would then lease that hardware, the way an airline leases planes rather than buying the whole fleet outright. And Broadcom, the company that helps design those same chips, agreed to stand behind about thirty-one billion of the loans. If the chips end up worth less than the lenders are betting, Broadcom absorbs part of the loss.
Read those two announcements side by side and the week tells a different story than the headline did.
Why the quiet number matters more
There is a real difference between raising money by selling a piece of your company and raising it by borrowing. When you sell equity, a slice of ownership, you hand over part of the upside and you owe nothing back. If the bet goes wrong, your investors lose with you. Patient money. The most patient money in the world, in fact, has been what built the AI labs so far.
Borrowed money is a different animal. You owe it back whether the bet works or not. You owe it on a schedule. And what you owe moves with interest rates, so the cost of the whole thing depends on conditions you do not control. For most of the AI boom, the enormous bills for chips and data centers were paid with equity, the patient kind. This week marked the point where the bills got big enough that even Anthropic’s near-trillion-dollar valuation could not cover them with stock alone, so the buildout started leaning on debt.
That changes the risk in a way the sticker price hides. A company financed entirely by equity can survive a bad year by simply having a bad year. A company carrying tens of billions in lease obligations has payments to make in the bad year too. The AI buildout just acquired a clock.
The Broadcom piece is the part I keep turning over. A chip designer promising that chips will hold their value is a company underwriting demand for its own product. That works beautifully while demand is roaring, and demand is roaring right now. It becomes a problem the moment it stops, because the same firm is exposed twice, once as the seller and once as the guarantor. None of this is reckless on its face. It is the sort of arrangement that looks obviously fine for years and then looks obviously foolish in about a week. Experienced investors have watched that movie in other industries, and the ending depends entirely on whether the demand was real or borrowed.
The timing is the tell
The same morning the debt deal surfaced, the government published its main inflation report, and it did not cooperate.
The Federal Reserve watches one inflation gauge more closely than any other. It goes by an ugly acronym, PCE, and the version that strips out volatile food and energy prices came in at 3.3 percent for the year through April. The Fed wants that figure near 2 percent. It has now been stuck well above target long enough that the word “transitory” has quietly retired. At the same time, the government revised its estimate of early-year economic growth down to 1.6 percent, slower than first thought. And the share of income people save fell to 2.6 percent, which means households are dipping into savings to keep spending at the level they are used to.
Put the two stories together and the tension is hard to miss. The AI industry started borrowing heavily to build, right as borrowing stayed expensive and the broader economy began to soften. For most of the past two years, a lot of venture math quietly assumed that cheaper money was coming, that rates would fall and make every aggressive plan look smart in hindsight. After this week, any founder still leaning on that assumption is leaning on a chair that got removed from the room. The rate-cut rescue is not arriving in 2026.
The rich got richer, on schedule
Step back from the debt deal and the rest of the week rhymed with it. Money kept piling into a small number of proven names while everyone else waited in line.
Anthropic’s raise looks slightly less colossal once you read the fine print. Roughly fifteen billion of the sixty-five was money its big cloud partners had already promised, so the genuinely new equity is smaller than the headline implies. Still enormous. Just not quite the figure that traveled.
Cognition, the company behind an AI software engineer called Devin, raised more than a billion dollars at a valuation of twenty-six billion, more than double where it stood eight months earlier. To see why that number raised eyebrows even in a giddy market, you need one concept. Investors often price a company as a multiple of its revenue, so a company earning a hundred dollars a year and valued at a thousand is trading at ten times revenue. Cognition was valued at roughly fifty-three times its revenue. The company reported it now runs at a pace of about 492 million dollars a year, up from 37 million a year earlier, which is a genuinely wild rate of growth. Fifty-three times is still a price that assumes the growth continues without a stumble. The detail I found most telling sat off to the side. Cognition says its own AI now writes more than 90 percent of the company’s code. The product builds the company that sells the product.
At the other end of the market, Meta did something that squeezes from below. It started selling AI subscriptions for $7.99 a month, with a premium tier at $19.99, testing first in Singapore, Guatemala, and Bolivia. The incumbents charge around twenty dollars for the same general idea. Meta is undercutting them by more than half, and it is doing it on top of an app empire with something close to a billion people already using its AI. So the week pressed startups from two directions at once. Giant rounds pulled talent and capital toward the top, and giant distribution pushed the price of basic AI toward the floor. The comfortable middle, a startup charging twenty dollars a month for a clever general-purpose assistant, got thinner from both ends.
A different track: the machines kept getting smarter
Everything above is the money story. There is a separate story running alongside it, and it is worth keeping the two apart, because they move for different reasons and reward different things.
The capability story had two moments this week.
The first was a new version of Claude, Opus 4.8, which Anthropic released on the same crowded May 28. The interesting claim was not that it scored well on coding tests, though the company says it did. The interesting claim was that this version is also its best-behaved, meaning less likely to do something its makers did not intend. For years the working assumption inside AI was that making a model safer made it a little dumber, that you traded capability for control. Anthropic is now claiming the opposite happened, that the strongest model is also the most reliable one. If that holds up under outside testing, and the right move is to treat any company’s claims about its own product as unproven until others reproduce them, it flips a long-standing tradeoff. Reliability stops being a tax on intelligence and starts being a feature you can sell, especially to banks, law firms, and anyone whose real fear was not that the machine was slow but that it would confidently make something up.
The second moment was stranger and, to me, more important. A team of four mathematicians published a paper disproving a long-standing conjecture in number theory, a problem about whether certain sets of numbers can stay small under both addition and multiplication at once. What makes it remarkable is how they did it. They borrowed a technique that an AI system had produced just one week earlier while cracking a different unrelated problem. The humans read what the machine did, recognized the method could travel, and carried it into their own field in days. An idea discovered by a machine seeded real human research almost immediately.
That is the development I would underline twice. We have spent a couple of years asking whether AI can do impressive things on its own. This was something else, a machine and a room full of humans trading techniques back and forth fast enough that the boundary between who discovered what started to blur. If that loop keeps tightening, the old comfort that frontier knowledge diffuses slowly, that a clever insight gives you years of advantage before the world catches up, gets shorter every cycle.
How this changes the way I read a valuation
The lasting lesson of the week is not any single deal. It is a habit of reading.
When a company is financed entirely with equity, the headline valuation tells you most of what you need to know about how the market prices it. When a company is financed partly with debt, the headline tells you less, sometimes much less, because the real obligations are sitting in a financing structure that the equity price ignores. Anthropic’s near-trillion-dollar valuation does not include the tens of billions in chip leases stacked behind it. Anyone trying to judge what the company is truly (net) worth has to add that weight back in, the same way you would not value a house by its purchase price while ignoring the mortgage.
That habit will matter well beyond this one company, because the debt deal is almost certainly a template, not an exception. The sums required to build frontier AI have outgrown what equity markets will hand over, and private credit is stepping into the gap. From here forward, reading an AI valuation means asking what is borrowed, who guaranteed it, and what happens to the payment schedule if demand cools or rates stay high.
If you want the short version, here are the questions I am carrying into next week:
When a company brags about a valuation, how much of the financing behind it is debt, and who is on the hook if the bet sours?
If basic AI keeps getting cheaper at the consumer level, what does a startup own that a giant with a billion users cannot simply copy at a lower price?
Does the business survive if the cost of computing stays flat or rises, rather than falling the way everyone has been quietly assuming?
The boom is not slowing down. What changed this week is the kind of fuel it is running on. For two years the AI buildout was financed by people who could afford to be wrong. It is starting to be financed by people who cannot, on a schedule that does not care whether the future arrives on time. That is a different machine than the one we have been watching, and it is worth knowing which one you are looking at before you decide what it is worth.