Last Week Ignite - 5.24.26
Customers, Landlords, Rivals
Somewhere deep in the SpaceX filing, past the launch-cadence tables and the Starlink subscriber math, there is a line that makes you put the coffee down.
Anthropic has agreed to pay roughly $1.25 billion a month for computing power. That is about fifteen billion dollars a year, and around forty-five billion over the life of the contract. The computers in question are Colossus, the giant supercomputer built by xAI. xAI is Elon Musk’s AI company, and it now sits inside SpaceX, which is why an arrangement between two AI labs showed up in a rocket company’s paperwork for going public.
Sit with the shape of that for a second. Anthropic and xAI build competing AI models. They are rivals in the most direct sense. And Anthropic is now paying its rival’s parent company fifteen billion dollars a year to keep the lights on. The money goes out the front door of one frontier lab and in the side door of another. We broke down the three separate businesses hiding inside SpaceX in the last article, Three Companies in a Rocket’s Clothing, so I will not relitigate the filing here. The new fact this week is the compute contract, and the contract tells you something the rest of the filing does not. The AI economy has folded in on itself. The biggest companies are now each other’s customers, each other’s landlords, and each other’s competitors, often all three at once.
That is the week. Let me show you how the rest of it rhymes.
Three price tags in five days
A quick piece of vocabulary, because the next part leans on it. When a company is still private, its value is set by whatever investors agree to pay in its latest funding round. That number is called the valuation. An S-1 is the long document a company files with regulators before it sells shares to the public for the first time. And a secondary is when you buy existing shares from an early shareholder who wants out, rather than buying new shares from the company itself. Our team does a fair amount of that last one, which is why these numbers are not academic for us.
Inside one five-day stretch, the two most valuable private AI companies in the world both got repriced.
Anthropic is reportedly closing a thirty-billion-dollar funding round that values it at around nine hundred billion dollars before the new money goes in. Bloomberg put the close as soon as the week of May 26, with Sequoia, Dragoneer, Altimeter, and Greenoaks each writing checks near two billion dollars. If that holds, it is Anthropic’s second thirty-billion-dollar round in about fourteen weeks, and it roughly doubles the company’s price tag in a single quarter. Anthropic has not confirmed the terms, so treat the exact figure as reported rather than carved in stone. The justification, according to reporting Anthropic has not denied, is revenue growing faster than almost anyone modeled and a first profitable quarter on the horizon.
Two days earlier, OpenAI quietly filed to go public. The filing is confidential for now, prepared with Goldman Sachs and Morgan Stanley, aimed at a debut this fall that could value the company north of a trillion dollars. The most revealing detail is not the number. It is that OpenAI’s own finance chief has reportedly told people internally that the company is not ready. The filing is being driven by the competitive clock and by the lawsuit with Musk finally clearing on May 18, not by a tidy set of books. When a company files before it feels ready, the calendar is being set by a rival, not by the business.
And SpaceX, whose filing started this whole thread, is now targeting a price for late June.
So three of the largest assets in private technology all moved within a week. For anyone holding or buying secondary shares in these names, that matters in a specific way. The market is about to know far more about Anthropic and OpenAI in early June than it has at any point before. A real funding round and a real regulatory filing both force real numbers into the open. The quiet window, where a buyer and seller can disagree about what a share is worth because neither has fresh data, is closing fast. We are watching it close rather than rushing to act inside it.
The constraint moved, and then it moved again
For most of the past year, the story people told about Anthropic was simple. Great models, not enough chips. The binding limit was compute, and the assumption baked into a lot of investor thinking was that Anthropic would stay starved of it into 2027.
Three things happened this week that break that assumption.
First, the SpaceX contract, which hands Anthropic forty-five billion dollars of compute over the next three years. Second, Anthropic’s compute chief confirmed publicly that the next big hardware buildout runs through June. Third, reporting surfaced that Anthropic is in early talks to rent servers running Microsoft’s own AI chip, called Maia 200. None of these is a small adjustment. Together they say that the wall Anthropic kept running into is getting torn down with money.
Here is the part worth slowing on. Anthropic now pulls compute from Nvidia chips, from Amazon’s homegrown chips, from Google’s chips, from the xAI supercomputer, and possibly soon from Microsoft’s. Five different supply lines, locked down by one customer. For a founder pitching a startup whose whole plan is to become the cheap inference layer, that picture should be sobering. Inference, by the way, is just the cost of running a trained model to answer a question, as opposed to the much larger one-time cost of training it. If five of the deepest-pocketed players on earth have all signed supply deals with the same buyer, the lane for a small independent chip or capacity startup is narrowing toward a slit.
The same week, the most public talent move of the year landed on the same side of the board. Andrej Karpathy, a founding member of OpenAI and one of the most recognizable engineers in the field, joined Anthropic to work on pre-training, the foundational stage of building a model. He is reporting to Nick Joseph. People follow constraints. When the best engineers move toward the lab that just relaxed its biggest limit, that is the market telling you where it thinks the next advantage gets built.
A separate track: the week a model did original math
Everything above is about money and capacity. This next part is about capability, and it deserves to be kept separate, because the two get muddled constantly and they are not the same thing. A company can raise thirty billion dollars and still not be able to do anything genuinely new. This week produced something genuinely new.
On May 20, OpenAI announced that one of its general-purpose reasoning models disproved a math conjecture that had been open since 1946. The problem comes from Paul Erdős, one of the most prolific mathematicians of the twentieth century, and it concerns how many pairs of points in a plane can sit exactly one unit apart. Erdős first posed it in a 1946 paper and attached a five-hundred-dollar prize. It sat unsolved for eighty years.
Two details make this more than a press release.
The model was not a specialized math machine. It was a general reasoning system, the same kind of tool people use for ordinary work, pointed at a hard problem. And the result was checked by humans who had every reason to be hostile. A group of named mathematicians, including W. T. Gowers and Thomas Bloom, posted a companion paper confirming the work. Bloom is the relevant name. Back in October, when OpenAI made an earlier and softer math claim, Bloom publicly called it a dramatic misrepresentation. He runs a registry of Erdős problems and does not hand out credit. Seven months later he co-signed the verification. The skeptic became the witness.
The honest version of the story includes the humans’ own caveat. The mathematicians noted that with the right experts assembled in advance, a person could probably have found this counterexample within a month. So the machine did not do something no human could ever do. It did something hard, original, and real, faster and more cheaply than the usual path, and it did it in a domain where human judgment was supposed to hold a comfortable lead.
That same lead slipped in a second place this week. A public forecasting benchmark reported that a frontier model matched the performance of human superforecasters for the first time. Superforecasters are the small group of people who are unusually good at putting calibrated odds on uncertain future events, and their edge has been one of the more durable human advantages on record. A model matching them, with no special scaffolding, is a quiet but pointed signal. I would label this one vendor-attributed until it is reproduced independently, since the benchmark named the winning model rather than the lab confirming it. The pattern still rhymes with the math result.
Put the two together and the band of cognitive work that only humans can do well got narrower in a single week. That has consequences for any business whose product is expert judgment about uncertain things. Private credit underwriting. Geopolitical risk advice. Insurance pricing for rare disasters. The moat there was always the scarce human who calls it right. That moat is getting shallower, and the market has not repriced the people who sell access to it.
The ground underneath the story
While the labs reprice and the models reach, the physical bill keeps growing.
Nvidia reported its quarter on May 20, and it was enormous. About eighty-two billion dollars in revenue, up eighty-five percent from a year earlier, with the data-center piece up more than ninety percent. The company guided to roughly ninety-one billion for the current quarter, and that figure assumes nothing from China. Jensen Huang told the call that Nvidia’s China data-center business had gone to zero, against more than four billion a year ago, and that the company has largely conceded that market to Huawei. The capex curve, meaning the money companies are pouring into chips and buildings, is still climbing even with one of the biggest markets in the world stripped out of the numbers.
You can see where that spending lands by watching the fights over electricity. In the last issue we noted a Nevada utility deciding to prioritize data-center load over forty-nine thousand residents. This week the trailing indicator showed up. A city in Missouri voted to effectively ban large-scale data centers. Power has become the real bottleneck, and that makes the boring infrastructure around it interesting. Grid software, demand response, advanced cooling, on-site generation. These are physical-AI investments where the buyer is committed capital rather than a discretionary software budget, which is a much sturdier place to sell.
And the labor side moved too. Intuit cut seventeen percent of its staff, around three thousand people, the same week it raised its guidance. Meta began cutting about eight thousand, ten percent of its workforce, and redirected thousands of those still employed toward AI teams. Zuckerberg’s memo to staff called AI the most consequential technology of their lifetimes. Both companies are converting headcount into compute budget while signing multi-year deals with the same labs that are repricing above. The story is consistent all the way down. Money and people are flowing toward inference, and away from rows in the org chart.
Two governments, two different bets
The rules caught up to none of this, which is itself the news.
The White House had a draft order ready that would have asked AI companies to give the government a look at their most powerful models before release. It was shelved on May 21 after late lobbying from a familiar set of voices, including David Sacks, Elon Musk, and Mark Zuckerberg. So at the federal level, the ceiling on how these models get deployed is back to the companies policing themselves for the rest of the year. That quietly removes a risk that was hanging over the OpenAI and Anthropic public-offering timelines.
The same day, California’s governor signed the first state order of its kind, directing state agencies to study how AI is displacing workers and to propose updates to the law that governs mass layoffs. It also asks for a public dashboard tracking AI’s effect on jobs by sector. Once a dashboard like that exists, every California company that cuts staff will eventually have to explain itself against it. For our founders headquartered in the state, that means AI-specific layoff rules are coming, probably by 2027. The federal government stepped back and the state stepped in, inside twenty-four hours.
There is a connected thread worth flagging on the platform side. Microsoft spent its last major conference turning AI agents into a governed enterprise layer, with identity, security, and admin controls wrapped around them. That is a platform trying to make autonomous software something a company can register, monitor, and shut off, the same way it manages employees and laptops. When a platform builds the control plane for a whole category, it also decides how much room startups get to live in that category. The Maia chip talks with Anthropic and the agent-governance push are the same instinct showing up in two places. Microsoft wants to sit underneath the AI economy at the chip, the model choice, and the management layer all at once.
What we are doing about it
Strip away the noise and a few things got more attractive this week and a few got less.
More attractive, to us, is anything that helps a company control what it spends running these models. Software that routes requests to the cheapest capable model, caps spend per task, and shows you where the money goes. One Chinese lab reported an agent running on its own for thirty-five hours straight this week, and whether or not that exact number reproduces, the direction is clear. As autonomous runs get longer, knowing your cost per task stops being a nice dashboard and starts being survival. Also more attractive is the boring physical infrastructure around power, and the services firms that go install AI inside regulated industries the big labs will not reach for years.
Less attractive is anything sitting directly under a giant. AI bookkeeping and tax tools at our check size now compete with an Intuit that has more compute per remaining employee. Math-specialized AI labs look weaker after a general model out-reasoned the specialists on their home turf. Independent chip startups face that five-supplier wall. And the second-tier model labs, the ones whose value is set by comparison to Anthropic and OpenAI, are getting squeezed by how fast the leaders are raising.
For the founders we talk to this week, a few questions are worth carrying into the room:
If Anthropic’s compute limit just loosened by forty-five billion dollars, what part of your edge survives once your competitors stop being throttled by rate limits?
Has any independent group reproduced the benchmark numbers in your deck since you first published them?
If OpenAI’s public offering compresses the whole sector’s pricing this fall, what does that do to your own raise?
On the secondary book, the names to watch are the same three that moved. Anthropic, where a nine-hundred-billion-dollar round resets the floor and the smart move is to price before the print, not after. OpenAI, where the quiet window shuts the moment the filing becomes public. And SpaceX, which we have already covered, now waiting on a late-June price.
The thread running through all of it is the circle we started with. A lab pays its rival for compute. A rival’s parent owns an option on a startup the first lab powers. A platform sells the chips, picks the models, and manages the agents. The money keeps moving in a loop, and the loop is getting tighter. For a small fund writing early checks, the lesson is not to stand in the middle of that loop and hope. It is to find the few places the giants cannot easily reach, and to remember that a model doing original math this week is a better signal about where the world is heading than any valuation printed alongside it.
See you next week.