The Rocket Company Owns the Meter

00:00:04 On Friday, a regulatory filing went up that rearranged who has leverage in artificial intelligence, and it did it with a single number. Google has agreed to pay SpaceX nine hundred and twenty million dollars a month for computing capacity. Not for a quarter — every month, from October of this year through the middle of 2029.

00:00:23 The filing spells out what Google gets for the money: access to roughly 110,000 Nvidia graphics processing units, plus the processors, memory, and supporting hardware sitting in SpaceX's data centers. Google's cloud arm told CNBC the deal is bridge capacity, meant to cover demand for its Gemini Enterprise product that ran, in their words, even higher than expected.

00:00:45 Sit with that for a second. Google is one of the largest owners of computing hardware on Earth. It is spending somewhere between a hundred and eighty and a hundred and ninety billion dollars on capital this year. And it is renting Nvidia chips from a rocket company.

00:01:01 Last month, Anthropic signed a deal to take all of SpaceX's capacity at its Colossus 1 data center in Memphis. Put the two together and, by one count from the Tesla-watcher Sawyer Merritt, Anthropic and Google are now paying SpaceX a combined two-point-one-seven billion dollars a month.

00:01:18 That's a revenue run rate of about twenty-six billion dollars a year, from a line of business most people didn't know SpaceX was in. Here is how SpaceX ended up here. In February, it merged with Elon Musk's AI company, xAI, in a deal that valued the combined entity at one-and-a-quarter trillion dollars.

00:01:37 SpaceX built a cluster of data centers in and around Memphis to train Grok, Musk's chatbot. Grok hasn't made much of a dent against OpenAI, Anthropic, or Google. So now the data centers built for Grok are being rented out to the very companies Grok was supposed to beat.

00:01:53 SpaceX's own filing puts it dryly: the compute infrastructure gives them, quote, substantial flexibility in how we allocate and monetize capacity. Translation — if you can't win the model race, become the landlord the racers pay rent to. And the timing isn't subtle.

00:02:09 SpaceX is going public next week, aiming for a valuation north of one-and-three-quarter trillion dollars. Musk needs the AI story to look like a business, not a money pit, and the numbers underneath are rough. SpaceX said capital spending in the first quarter alone hit ten-point-one billion dollars, more than double a year earlier, with seven-point-seven billion of that committed to AI.

00:02:32 The AI segment itself? An operating loss of two-and-a-half billion dollars in the quarter, on just eight hundred and eighteen million in revenue. So these compute deals are doing double duty. They bring in cash, and they tell IPO investors that even Google will line up to pay.

00:02:49 One reply under Merritt's post caught the absurdity better than I could: imagine explaining to someone in 1995 that companies would one day spend nearly a billion dollars a month renting computers from a rocket company. We are there. And it concentrates something.

00:03:05 The neoclouds — CoreWeave, Nebius, the companies whose whole pitch is renting out graphics processing units — saw their stocks get hit Friday in a broader tech selloff. SpaceX just walked into their market with a trillion-dollar balance sheet and the most famous name in the industry.

00:03:22 When the people who own the launch pads also own the compute, the rent flows to a very small number of doors.

00:03:28 There is a counter-document to the SpaceX IPO, and it came out the same day. Nathan Calvin, a former OpenAI staffer now working with a group called the Midas Project, has been asking one question for two weeks: would SpaceX's prospectus show that the company is prepared to manage the risks of frontier AI?

00:03:46 The prospectus is public now, and his answer is blunt. Quote: the answer is a clear no. Walk through his case, because it's specific. SpaceX tells investors it relies on AI for the vast majority of its addressable market. It says it intends to keep scaling Grok toward — their words — multiple trillions of parameters and a step change in intelligence.

00:04:07 And then, across two hundred and seventy-seven pages, Calvin says it says almost nothing about the biological-weapon, cyberattack, and loss-of-control risks that a long line of AI researchers treat as the central hazards of the technology. His sharpest line: SpaceX is asking investors to fund a technology its own founder called far more dangerous than nukes.

00:04:28 Musk has said that, publicly, for years. The document raising money on that technology doesn't engage with the danger. Now, you could dismiss this as advocacy. Calvin runs a watchdog group; finding gaps is the job. So weigh the corroboration. When Calvin's group put out an earlier report, xAI updated its public safety page in response — and Calvin says the new page was inaccurate and itself AI-generated, gesturing at industry-standard safety practices without actually adopting them.

00:04:56 Wired covered the initial report. And the harms here aren't hypothetical projections. SpaceX's combined AI business is facing multiple lawsuits and government probes in the United States and abroad after Grok let users generate and share non-consensual sexual images — deepfake porn — by altering photos of adults and children.

00:05:15 That already happened. That is in the record. What I keep returning to is the asymmetry of who carries this. SpaceX gets to raise record money on the upside of scaling Grok. The downside risks Calvin names — bioweapons, cyberattacks, a model that does something nobody intended — those land on the public, not on the cap table.

00:05:34 An IPO prospectus is supposed to be the one document where a company has a legal duty to lay out what could go wrong. When the risk section is silent on the risks the founder himself called civilizational, that silence is itself information. It tells you the company would rather investors not think too hard about it days before the bell rings.

00:05:55 While SpaceX was turning into a landlord, a second story broke that points at a different kind of owner. CNBC confirmed Friday that Sam Altman and the White House have been in ongoing talks about the United States government taking an equity stake in OpenAI. Not regulating it.

00:06:11 Owning a piece of it. The idea is older than it sounds. A source told CNBC that Altman first floated it to the Trump administration back in 2025, and the conversations have run for more than a year. The mechanism on the table is unusual: OpenAI could donate equity to the government to seed something it calls a Public Wealth Fund — an idea from the company's own April policy proposal.

00:06:34 Their pitch: the fund holds diversified long-term assets and lets ordinary citizens share in the upside of AI growth, maybe even receiving returns directly. Trump addressed it himself, talking to reporters on Air Force One. His words: there are concepts where pieces could be given to the American public, where the American public essentially becomes a partner.

00:06:55 This isn't a one-off. Trump signed an executive order in February calling for a federal sovereign wealth fund, and the administration has already taken stakes in Intel, in IBM, and in quantum and critical-minerals companies during his second term. So a government equity position in a frontier AI lab is consistent with a pattern — the state buying its way onto the cap tables of the companies it deems essential.

00:07:20 And the politics are scrambling in interesting ways. Senator Bernie Sanders met with Altman on Wednesday and told CNBC they discussed exactly this, the sovereign wealth fund concept. When a democratic socialist from Vermont and a Trump White House are circling the same idea from opposite directions, something structural is shifting.

00:07:40 I want to name what is appealing and what is dangerous here, because both are live. The appealing version: AI may concentrate enormous wealth in a handful of companies, and a public stake is one way the gains don't pool entirely with private investors and Abu Dhabi's sovereign fund, which co-led OpenAI's last record round.

00:07:59 The dangerous version: once the government owns a piece of OpenAI, it has a direct financial interest in OpenAI winning. The regulator becomes a shareholder. Every safety rule, every antitrust question, every export decision now runs through a government that profits when the company's valuation climbs.

00:08:17 OpenAI is worth more than eight hundred and fifty billion dollars and heading toward its own IPO. The question nobody has answered is who, exactly, is supposed to keep that company honest once the referee owns equity in the team.

00:08:30 I promised yesterday I would track the fight between Anthropic and the Pentagon — the question of whether a lab can refuse to let its model be used for every lawful military purpose. Friday gave us the government's answer, and it came in the form of a signature.

00:08:46 President Trump signed a National Security Presidential Memorandum on AI in the national security enterprise. Michael Kratsios, the administration's science and technology director, laid out what it does. Three pieces. First, it accelerates AI adoption across the defense world from multiple vendors, specifically to avoid single points of failure.

00:09:06 Second, it updates the Defense Department's guidance on autonomy in weapons systems — Kratsios said, to keep pace with the frontier. And third, the piece that answers yesterday's question directly: it ensures, in his words, that no entity can disable or degrade an AI system our warfighters depend on without prior approval.

00:09:25 Read that last clause again, because it is the government drawing a line under the exact dispute we covered. A lab like Anthropic builds a model. The military builds it into something it depends on. And now the federal position is that the lab can't pull the plug, throttle the service, or refuse a use without the government's sign-off first.

00:09:45 The off button that Anthropic wanted to keep — the right to say no to certain uses — this memorandum is reaching for it. The state is asserting that once its warfighters rely on your model, control of that dependency is no longer entirely yours. The multiple-vendor language matters too.

00:10:02 If the Pentagon spreads its AI across several suppliers, no single lab has the leverage to hold out, because there's always another model ready to take the contract. That is deliberate. It turns the labs from irreplaceable partners into interchangeable ones. One thoughtful reply under Kratsios's post asked the right question — the memorandum says citizens deserve to know AI is handled responsibly, but classified military AI means classified decisions, so how do you get public accountability without public visibility?

00:10:33 Nobody has a clean answer. And this comes the same week OpenAI confirmed it will comply with Trump's separate executive order — the one asking AI companies to let the government assess their models for up to thirty days before release. Altman called it getting the balance right.

00:10:50 So inside one week: the government wants pre-release access to the models, equity in the companies, and command over whether military AI systems can be switched off. Three different levers, all pulling the frontier labs closer to the state. The voluntary frontier we talked about on Tuesday is getting a great deal less voluntary.

00:11:09 All of that is the view from the top — trillion-dollar deals, White House equity, presidential memoranda. The view from the ground is moving in the opposite direction, and it is moving fast. On Friday, the New York State legislature passed a one-year moratorium on new large data centers.

00:11:26 If Governor Kathy Hochul signs it, it would be the first statewide ban of its kind in the country. The details tell you what this is really about. The bill defines a large data center as anything with a peak demand of at least twenty megawatts, and it directs the state's environmental agency to produce a report on how much electricity, water, and land these facilities consume, and how much pollution they create.

00:11:50 It also forces any company planning one to hold and fund a public hearing at least three months before it can get approval. That public-hearing requirement is deliberate — lawmakers want the fights out in the open, in front of the neighbors, before the concrete gets poured.

00:12:06 New York's grid operator says it is currently reviewing twenty-four data center proposals totaling more than nine thousand megawatts. A single proposed project in Albany, at a hundred and eighty megawatts, has already drawn organized resident opposition. New York isn't alone, and this is where it gets political.

00:12:25 The same day, Illinois Governor JB Pritzker announced he is pausing the processing of every application for the state's data-center tax-break program starting July 1. He asked lawmakers to suspend the incentives for two years so they could study the impact; they didn't act, so he is using executive control over application processing to do what he can.

00:12:46 He put it sharply: data centers should, quote, pay their fair share, and communities should know ahead of time how much water and electricity a facility will consume. Pritzker is running for a third term and is widely seen as eyeing the White House in 2028. He isn't picking this fight by accident.

00:13:03 An NBC News poll this year found that nationally, Americans believe the dangers of AI outweigh its benefits. This is the same current that ran through Seattle's moratorium, which we covered yesterday, and through Maine's attempt last year. The pattern is bipartisan and it is bottom-up.

00:13:20 The federal government is racing to feed the buildout — more compute, more vendors, more capacity. State and local officials are discovering that the buildout runs on their power grids, their water tables, and their constituents' utility bills, and that slowing it down is, increasingly, a vote-winner.

00:13:38 The binding constraint on AI was never going to be a clever model. It was always going to be a substation, a water permit, and a room full of angry neighbors. That fight is now playing out in two of the largest states in the country at once.

00:13:52 Step outside the United States for a moment, because the contest over who controls computing power isn't only a domestic one. A research team that includes Huawei said this week it successfully used Huawei's own Ascend 910C chips to complete post-training for the DeepSeek V4 Pro model.

00:14:09 If you've been tracking the export-control story, that sentence carries a lot of weight. Here is why it matters. For years, the assumption behind American chip sanctions was that China could use domestic silicon for inference — running models that already exist — but not for the heavy, finicky work of training and refining frontier models.

00:14:29 That work needed Nvidia. Post-training is the step where a model learns to follow instructions, obey safety rules, and handle specific tasks, and doing it on Chinese hardware was supposed to be out of reach. The Ascend 910C is built on the Chinese foundry SMIC's seven-nanometer process, using Huawei's in-house architecture.

00:14:49 DeepSeek V4 Pro was trained on thirty-three trillion tokens. The team validated its parallelism scheme on both Nvidia chips and Huawei's, which means this isn't yet a clean break from American hardware — but it is a real step toward training advanced models without it.

00:15:05 The stakes are about leverage. Export controls are America's main tool for slowing China's AI program, and the entire theory depends on Chinese chips staying a generation or two behind. Every credible demonstration that Huawei silicon can do frontier-grade work chips away at that theory.

00:15:22 Jensen Huang of Nvidia has been warning for a while that cutting China off from American chips doesn't stop Chinese AI — it just hands the domestic market to Huawei and accelerates exactly this kind of self-reliance. Friday's result is evidence for his argument.

00:15:37 And it rhymes with what Europe is doing. The same day, the European Commission published its Communication on European Tech Sovereignty, paired with an EU open-source strategy — the policy package we touched on earlier this week, now formally out. Different motive, same instinct: don't depend on someone else's chips, someone else's models, someone else's cloud.

00:15:59 The Chinese version is forced by sanctions. The European version is chosen out of anxiety about depending on American platforms. But both point at a world fragmenting into compute blocs, each trying to own the full stack so that nobody else holds the off switch over their AI.

00:16:15 That's the deeper consequence of a technology this central to national power — every serious government eventually decides it can't afford to rent.

00:16:24 Underneath all the maneuvering over who owns the compute and who sits on the cap table, the technology kept doing concrete things this week — some of them remarkable. Two are worth your attention, because they show both the upside and the strangeness of where this is going.

00:16:40 The first is from the University of Cambridge. Researchers there announced the first vaccine whose key component — the antigen, the part that teaches your immune system what to attack — was designed entirely by an AI system, and then carried into human trials. The approach is different from a normal vaccine.

00:16:58 Instead of targeting one virus, the AI analyzed the genetic codes of many known coronaviruses and designed a single super-antigen meant to protect across the whole family — every known variant of the COVID virus and related animal coronaviruses with pandemic potential.

00:17:14 The first safety trial ran in thirty-nine volunteers. The vaccine was safe, with what the team called a modest but encouraging immune response, published in the Journal of Infection. A larger Phase 2 study is planned. This is early, and a modest response in thirty-nine people isn't a finished product.

00:17:32 But it is the first time an AI-designed antigen has gone into human arms, and the strategy — design once, protect against a family of future threats — is the kind of thing that could change how we prepare for the next pandemic. The second is stranger, and it happened in crypto.

00:17:49 The privacy coin Zcash dropped about forty-eight percent after Anthropic's Claude model, scanning the code, found a critical vulnerability that would have allowed unlimited minting of the currency — an infinite-money bug. It had sat in the code, unnoticed, for four years, and was patched on June 1st.

00:18:07 Here's the twist: the exploit never actually happened. Nobody minted infinite coins. The market crashed the price anyway, the moment everyone learned the flaw had existed. As one trader put it, a bug can exist for four years and nobody cares; the moment everyone knows about it, the price cares a lot.

00:18:24 And another reply put it best: imagine paying millions to top-tier human security firms for audits over four years, only for an AI model to casually scan the code and find the infinite-money glitch they all missed. That's the capability ledger this week, away from the boardrooms.

00:18:41 An AI designing a vaccine antigen that works in a human body. An AI finding a four-year-old hole that professional auditors walked past. Neither of those is a press release about valuations or equity stakes. They are the actual thing the institutions are fighting to control — and a reminder that the capability is racing ahead of every structure being built to govern it.

00:19:03 Pull the day together and a single pattern shows up in every story. Power in AI is consolidating around whoever controls the physical and financial chokepoints — and this week, those chokepoints got a lot more visible. SpaceX turned its surplus data centers into a tollbooth, and now Google and Anthropic pay it more than two billion dollars a month to pass through.

00:19:23 Washington moved to take equity in the labs and signed a memorandum asserting that it, not the companies, decides when a military AI system can be switched off. New York and Illinois reached for the one lever they actually hold — the power grid and the permit office — to slow the whole thing down.

00:19:39 China demonstrated it can train advanced models on its own silicon, and Europe published its plan to stop depending on anyone else's. Every one of those is a move about ownership: of the compute, of the cap table, of the grid, of the supply chain. The models themselves were almost an afterthought.

00:19:55 What I'm watching next is concrete. SpaceX goes public next week, and that prospectus — the one with two hundred and seventy-seven pages and a silence where the risk section should be — gets tested by the market for the first time. And those government-equity talks with OpenAI either produce real terms or stay a concept Trump muses about on Air Force One.

00:20:14 The first tells us how much the public will pay for an AI story with the danger left out. The second tells us whether the United States is about to become a shareholder in the company it is supposed to regulate. Two answers, both coming soon, both circling the same question: when this technology finally pays off, who's standing at the meter collecting.

00:20:33 That's what I'll chase tomorrow. I'm Jonas.