Who Owns the Buildout

00:00:04 Today is Monday, the first of June, and the AI industry spent the day arguing about something it usually keeps offstage: ownership. Not what the models can do. Who holds the equity, who pays the electric bill, who carries the legal risk, and who the machine ends up serving.

00:00:19 Start with the cleanest fact. On Monday morning, Anthropic said it has confidentially filed a draft registration statement — an S-1 — with the Securities and Exchange Commission. In plain terms, the company that builds Claude took the first formal step toward selling shares to the public.

00:00:36 Its own statement was deliberately flat: "This gives us the option to go public after the SEC completes its review," the company said. "The proposed initial public offering will depend on market conditions and other factors." That is lawyer-tuned language for a very large event.

00:00:52 Here are the numbers that make it large. Last week, Anthropic closed a funding round at a valuation of about 965 billion dollars. That tops OpenAI, which was valued at roughly 852 billion in late March. The company told the market in May that its revenue run rate had reached 47 billion dollars — up from about 10 billion in annual revenue last year.

00:01:12 So on paper, this is a company worth nearly a trillion dollars, growing its top line several-fold inside a single year. And it's now lining up to sell a piece of itself to anyone with a brokerage account. The timing is competitive, and worth slowing down on. By filing confidentially now, Anthropic gets out ahead of OpenAI, which is readying its own confidential filing.

00:01:33 Elon Musk's SpaceX has already filed its prospectus publicly and is gearing up for a roadshow this week, with plans to debut next week. A confidential S-1 doesn't lock anyone into a date — the official prospectus only has to land in investors' hands at least fifteen days before a roadshow begins.

00:01:50 SpaceX, for reference, filed confidentially on April 1 and disclosed its public prospectus on May 20. So what you're watching is the front edge of a wave: three of the most valuable private companies in the world, all in the AI and AI-adjacent business, all queuing for the public markets in the same season.

00:02:08 A week ago on this show, I said I'd be watching for exactly this filing — for the prospectus and, eventually, for what it reveals about how concentrated Anthropic's revenue really is. The filing is here. The disclosures aren't yet, because confidential means the financial detail stays sealed until the SEC finishes its review.

00:02:27 So I can't tell you yet who those 47 billion dollars in run-rate revenue actually come from, how many customers, or how durable the contracts are. That's the document I want to read when it goes public, and it's the one that matters most for anyone deciding whether this is a trillion-dollar company or a trillion-dollar bet.

00:02:45 One more thing the reporting surfaced, because it complicates the clean growth story. The Pentagon blacklisted Anthropic's models earlier this year, after negotiations between the two sides collapsed, and defense contractors dropped the company to comply. Anthropic sued the Trump administration to reverse the blacklisting, and that litigation is still going.

00:03:06 President Trump told CNBC in April that a deal is "possible." In the meantime, the company's growth came from the private sector and from consumers — Claude hit the top spot on Apple's US free-app chart in late February. Some of it came from a cybersecurity model called Claude Mythos, which has drawn interest from senior people in the administration.

00:03:26 So the same company suing the government is courting it. That tension is going to be in the prospectus too, whether they spell it out or not.

00:03:34 If you want to understand what investors are actually buying, you have to look past the model and at the wiring behind it. And there's a specific number in the SpaceX prospectus that tells the story. According to that filing, Anthropic will pay SpaceX 1.25 billion dollars per month — through May 2029 — for compute at the Colossus 1 data center in Memphis, Tennessee.

00:03:58 That's 15 billion dollars a year, going out the door, to rent the machines that run Claude. And here's the clause that should make you sit up: either company can terminate the agreement with 90 days' notice. So the single largest cost line behind one of the most valuable AI companies in the world is a deal that can be unwound in three months.

00:04:19 That's not a long-term lease on a building. That's closer to a rolling contract for a thing both sides know is suddenly scarce. Brett Winton, the chief futurist at ARK Invest, did the arithmetic on the other side of that deal and posted it Monday. His framing: Anthropic is paying SpaceX's AI operation roughly 24 billion dollars in revenue per nameplate gigawatt of capacity, while SpaceX spent about 29 billion dollars per gigawatt building that capacity out.

00:04:48 If a contract like that runs five years, Winton figures it throws off more than 50 billion dollars in cumulative pre-tax cashflow. And then he landed the point I keep coming back to: "Being wildly efficient at converting cash into physical compute assets in a world that finds itself suddenly wildly short of compute assets is not a bad spot to be."

00:05:13 The constraint isn't the cleverness of the model. It's gigawatts. Power and land, chips and cooling, and the capital to assemble them faster than the next company can. Whoever turns dollars into running silicon quickest holds the leverage. And in this particular arrangement, the company holding that leverage is a rocket maker that also runs a frontier lab and is about to go public itself.

00:05:37 Now, a caution, because I don't want to launder a tidy number into a fact. Winton's figures drew an immediate skeptical reply on X, pointing out that there's no public evidence of a distinct "SpaceX AI" operating entity separate from SpaceX, and that the per-gigawatt revenue and build-cost figures don't line up neatly with how hyperscaler compute is normally priced.

00:06:01 I think that skepticism is fair. Where I land is this: we have two anchored facts — Anthropic's 1.25 billion a month, disclosed in a filing, and the 90-day termination clause — and then a layer of modeling on top that's directionally useful but not audited. Take the modeling as a way to think, not as gospel.

00:06:20 What I'd hold onto is the fragility. A trillion-dollar valuation resting on a compute contract that can be canceled in 90 days is a reminder of where the real bargaining power sits. It sits in the physical layer — and the people who own the power and the chips can reprice the people who own the models.

00:06:39 That's the thing the IPO is really exposing.

00:06:42 Now the counter-move, and it landed on the same day, which is almost too neat. While Anthropic was filing to sell shares, Senator Bernie Sanders announced he'll introduce a bill called the American AI Sovereign Wealth Fund Act. The idea, stripped to the mechanism: a one-time, 50 percent levy — not on profits, on equity — applied to the stock of OpenAI, Anthropic, and the other largest AI firms.

00:07:05 Half their shares would transfer into a publicly held fund. The government would take voting seats and equal board representation at each company, with the power to block decisions deemed harmful to citizens. And the proceeds, Sanders says, would flow back to Americans as direct payments and as guaranteed access to health care, education, and housing.

00:07:26 Sit with the scale of that for a second. Most tax proposals go after income or corporate profit. This one goes after ownership itself — it would make the United States public a half-owner of the frontier labs by statute. Sanders made the moral argument plainly: "AI is being built on a public resource far more valuable than oil," he said, "the accumulated knowledge, creativity, and labor of mankind." His logic runs: the models were trained on our books, our songs, our journalism, our code, our conversations — so when that resource generates wealth, the public should share in it.

00:08:00 The sovereign-wealth-fund framing is borrowed from oil states like Norway, which banked the proceeds of a public resource for everyone. Sanders is saying the training corpus is the new oil field. Let me be clear about where this sits, because it's easy to over- or under-read.

00:08:16 As of today, it's an announcement, not a bill. Sanders said the specific spending priorities and the mechanics of implementation will come with the legislation he unveils in the coming weeks. It has, realistically, almost no path through the current Congress. A forced 50 percent equity transfer would draw immediate constitutional challenge as a taking.

00:08:37 So I'm not telling you this becomes law. I'm telling you the ceiling of the debate just moved. A sitting senator has put public majority ownership of the AI labs on the table as a serious proposal, on the literal day one of those labs filed to enrich its private shareholders.

00:08:53 And here's the connective tissue I can't ignore. Zak Kukoff, an investor, made the point on X that the IPO wave is itself a political event. "If you think money in DC is pervasive now," he wrote, the Anthropic, SpaceX, and OpenAI offerings are "equipping a generation of new political donors whose financial capital dwarfs that of legacy mega-donors." Put those two things side by side.

00:09:16 The same liquidity event that mints Sanders' target also mints the donor class that will fund the fight against Sanders' bill. The people who'd be taxed are the people who'll be writing checks to the campaigns that decide whether they get taxed. That's the loop.

00:09:31 Whoever owns the equity after these IPOs owns a great deal of the political process that governs the equity.

00:09:37 The third claim on these companies isn't financial. It's legal, and it got personal on Monday. Florida's attorney general, James Uthmeier, filed an 83-page complaint against OpenAI — and he's seeking to hold Sam Altman personally liable for harm to Florida residents.

00:09:53 Read that again, because the personal part is the unusual part. Companies get sued constantly. Naming the chief executive as an individual defendant, on the theory that his own conduct caused the harm, is a different move.

00:10:06 The complaint cites what it calls Altman's "utter disregard for the risk to human life." And the alleged harms are heavy: the filing claims ChatGPT has aided and abetted mass shooters in, quote, "deadly rampages," driven vulnerable people to suicide, degraded users' critical thinking, and addicted minors to a tool that, in the complaint's words, "feigns human compassion." The framing Uthmeier chose is the line he wants quoted: "This litany of harms is driven by Defendants' insatiable quest to win the AI arms race and amass large fortunes, despite knowing the danger of ChatGPT."

00:10:48 He framed it as consumer protection — that the company is, in his words, "endangering our kids and deceiving parents into believing that this application is safe for use." There's also a criminal thread running alongside the civil one: Uthmeier's office opened an investigation in April into whether OpenAI bears responsibility for the 2025 mass shooting at Florida State University, after prosecutors reviewed chat logs between the gunman and ChatGPT.

00:11:15 He said Monday that investigation continues. This doesn't sit in isolation. OpenAI is already being sued by seven families from the Tumbler Ridge mass shooting in Canada in February, who allege the attacker used ChatGPT to plan it. The company faces multiple wrongful-death suits alleging the chatbot drove users into harmful delusions and, in some cases, suicide.

00:11:36 Altman apologized to the Tumbler Ridge community in an April letter. And all of this comes just weeks after the Musk-versus-OpenAI trial over the for-profit conversion, where an advisory jury decided Musk had waited too long to sue, a verdict the judge adopted.

00:11:52 A word on the legal reality, so I don't oversell it. Piercing the corporate veil to reach a CEO personally is hard. State consumer-protection suits often settle or get narrowed long before they touch an individual's assets. So Altman is probably not, in the end, paying out of his own pocket.

00:12:09 But the calculation behind it is new. A state AG has decided the most effective pressure point on a frontier lab is the founder's personal name, and he's betting other states agree. Tie that back to the day's spine: as these companies file to convert their founders' equity into public-market billions, a parallel legal track is trying to attach personal liability to the same people for what the products do.

00:12:33 Ownership and accountability are being fought over at the same desk, and right now they're pointed at the same handful of names.

00:12:41 So who pays for all of this? Not in the abstract — in your mortgage rate. On Monday, Axios reported that several Federal Reserve officials are warning the economic costs of the AI buildout are arriving faster than the benefits. And that's a direct shot at the optimistic case the Fed's own chair has been making.

00:12:59 Here's the setup. Kevin Warsh, who now leads the Fed, has argued that AI will be "a significant disinflationary force, increasing productivity and bolstering American competitiveness." The theory is clean: if AI lets workers and businesses produce more with the same resources, the economy can grow without generating inflation, which gives the central bank room to keep interest rates low.

00:13:23 Lower rates for everyone, courtesy of the productivity miracle. That's the bull case, and a lot of policy optimism has rested on it. Several of Warsh's colleagues are pushing back, and they're doing it with specifics. Alberto Musalem, the St.

00:13:38 Louis Fed president, said in a speech last week: "I believe it would be risky to rely on the prospect of higher productivity growth in the future to solve our inflation problem today." He went further — that the risks of "a miscalculation about its impact on productivity and inflation are too great," and that the Fed should keep its guard up against above-target inflation now rather than bet on productivity tomorrow.

00:14:03 Mary Daly, the San Francisco Fed president, put the evidence problem bluntly: she's bullish, she said, but "if you talk to companies, they say they haven't seen the productivity yet." Productivity has run about 2.4 percent a year over the past three years, stronger than the 1.5 percent of the 2010s — but it started climbing before most companies had adopted AI, so nobody can cleanly credit the technology for the lift.

00:14:29 And then the part that anchors this to everything else today. Fed governor Lisa Cook pointed out that AI investment demand is itself pushing prices up — on chips and high-tech equipment, on software, and crucially on construction labor, electricity, and water.

00:14:45 She noted companies have announced roughly 1.5 trillion dollars in data-center investment plans, and said the effects of that demand on prices are, in her word, "apparent." A new World Economic Forum survey of economists now puts notable AI-driven productivity gains in most sectors about two years out — a longer timeline than they expected at the start of this year.

00:15:07 Let me translate what the central bank is actually saying, because it reframes the whole day. The first thing the AI buildout does to the real economy is raise demand: it bids up power, water, skilled trades, and silicon, and that's inflationary right now, today, in places that have never run a model.

00:15:26 The promised payoff — the efficiency that's supposed to make everything cheaper — is showing up in the statistics maybe, somewhere, in two years. So the cost lands first and the benefit lands later, and in between, the people paying are everyone who borrows money or pays a utility bill near a new data center.

00:15:45 That's the answer to "who pays for the buildout." For now, broadly, it's the public — through prices and rates — while the equity stays private. Which is precisely the asymmetry Sanders is trying to attack from the other end.

00:15:58 If you still doubted the scale of the capital appetite here, Alphabet settled it Monday afternoon. The parent company of Google announced a proposed 80 billion dollar equity capital raise — to expand AI infrastructure and compute. And the sharpest reaction came from a one-line comment on Hacker News: "How is Alphabet suddenly short of capital?"

00:16:21 Alphabet is one of the great cash machines ever built. Search throws off profit at a rate most companies can only fantasize about. A business like that funding things from its own pocket is the default expectation. So when it chooses to raise 80 billion dollars by selling equity — diluting existing shareholders rather than tapping its cash or issuing debt — that's a signal about magnitude.

00:16:43 The buildout is large enough to strain even the strongest balance sheet in technology. The market noticed: Alphabet shares fell about 1.7 percent after hours on the news. Investors don't love being told the bill is bigger than the cash drawer. And this connects straight to what Governor Cook was describing.

00:17:01 That 1.5 trillion dollars in announced data-center investment isn't a spreadsheet abstraction — it's raised, somewhere, by someone, and it competes for the same chips, power, and construction crews as everything else. Alphabet passing the hat for 80 billion is one tranche of that number becoming concrete.

00:17:19 Literally concrete, in one case. The same day, OpenAI broke ground on a one-gigawatt data center in Michigan, part of its Stargate program. The company framed it the way these announcements always get framed — expanding access, creating jobs, supporting communities.

00:17:34 And those jobs aren't hypothetical; a gigawatt of data center is a serious construction project and a long-term local employer. But hold the Fed's warning next to the press release. A gigawatt is also a tremendous draw on the local grid and the local water supply, in a year when the central bank is naming electricity and water as places AI demand is pushing prices up.

00:17:55 The groundbreaking in Michigan and the inflation warning in Washington are describing the same gigawatt from two ends. One sees jobs and access. The other sees the utility bill for the town next door. I don't think either framing is dishonest. They're both true at once, and that's the part that's hard to hold.

00:18:13 The buildout creates work and capability where it lands, and it bids up the cost of the basic inputs everyone shares. The reason ownership matters so much in this story is that those two effects fall on different people. The capability and the upside concentrate.

00:18:28 The grid stress and the price pressure spread out.

00:18:31 The day's last story is colder. It's about what AI power does when it's pointed at people, and about whether the controls we've built actually restrain it. The New York Times, reporting by Julian Barnes, published an investigation built on internal documents from a Chinese company called Geedge Networks.

00:18:49 Geedge is building AI tools designed to predict which citizens might pose a — their phrase — "political risk." The system ingests vast amounts of personal data — internet usage, telecommunications records, location history — and assembles behavioral profiles meant to flag people before they act, or before they speak.

00:19:07 This is predictive policing aimed at dissent, built as a product. The documents surfaced because more than 100,000 internal files were leaked to a consortium of journalists, technologists, and human-rights groups — technical specifications, source code, client contracts, and marketing materials.

00:19:24 They show Geedge presenting itself publicly as a cybersecurity company while actually selling censorship and surveillance infrastructure. Its domestic clients include provincial public-security bureaus and state-owned telecoms. And it has gone abroad: marketing what amounts to a "Great Firewall in a box" to governments that want Chinese-style control, with documented deployments in Kazakhstan, Pakistan, Ethiopia, and Myanmar.

00:19:49 So this isn't only about one country's internal repression. It's an export business. Here's the part that matters for the policy fight we've been tracking on this show. According to the reporting, US export controls on advanced AI chips slowed Geedge's work. The restrictions imposed under the Biden administration, limiting China's access to advanced American-designed chips, actually bit.

00:20:11 The surveillance system the company wanted to build ran into the wall of not having the compute to build it. That's notable because so much of the export-control debate is about whether the controls do anything at all, given how many chips slip through. This is a documented case where the lever, when it was pulled, constrained an actual surveillance state's actual ambitions.

00:20:33 And now the warning. US officials have signaled that China may gain access to a more advanced version of Nvidia processors following recent talks between Washington and Beijing. So the same week we get evidence that chip controls slowed a predictive-surveillance build, we also get signals the controls may loosen.

00:20:51 Yesterday on this show I said I'd keep watching whether the Bureau of Industry and Security turns guidance into enforceable rule, and how big the number is for chips that slipped through. The Geedge documents add the other half of that ledger: not just the cost of the chips that got out, but the concrete thing those chips would let a surveillance state do.

00:21:12 If you want a clean argument for why export controls are a human-rights instrument and not just an economic one, it's in those leaked files. Loosening them is a decision about what someone else's government gets to compute about its own people.

00:21:26 Pull the threads together and the day reads as one argument with itself about ownership. Anthropic filed to sell the public a piece of a near-trillion-dollar company whose largest cost is a compute contract that can be canceled in 90 days. Bernie Sanders proposed the public should own half of it outright, by statute.

00:21:42 Florida tried to pin the harms of these products on a founder by name. The Fed warned that the buildout's first gift to the rest of us is higher prices, not higher productivity. Alphabet had to sell 80 billion dollars of stock to keep up, and OpenAI poured a gigawatt of foundation in Michigan.

00:21:58 And a stack of leaked Chinese documents showed what this kind of power does when it's aimed at predicting dissent — and that, for now, chip controls can still slow it down. The capability isn't what's in dispute today. The ownership is. Who holds the equity, who pays the power bill, who carries the liability, and who the machine ends up serving.

00:22:16 Each of those questions got a sharper edge on Monday, and not one got an answer. What I'm watching next is narrow and concrete. When Anthropic's prospectus goes public after the SEC review, I want the customer-concentration breakdown — how much of that 47 billion dollar run rate depends on a handful of accounts, and how much on a compute deal with 90 days of slack in it.

00:22:36 That's the number that tells you whether you're looking at a business or a position. I'll have it when the document does. I'm Jonas.