Who Pays to Write the Rules

00:00:04 Two political action committees spent this spring trying to pick your next member of Congress, and both of them are funded by artificial intelligence companies. Theodore Schleifer laid the whole thing out in the New York Times today, and it's worth slowing down on, because it tells you something the earnings calls won't.

00:00:22 One committee is called Public First. It's backed by a twenty-million-dollar donation from Anthropic, and it pitches itself around transparency, safety standards, and public oversight. The other is called Leading the Future. It sits in OpenAI's orbit, and it's working with more than a hundred million dollars from backers including the venture firm Andreessen Horowitz, OpenAI president Greg Brockman, the search startup Perplexity, and the Palantir co-founder Joe Lonsdale.

00:00:49 Both of them describe themselves as pro-AI. They're spending the money to fight over the shape of the rules. Here's how that fight actually looks on the ground. A super PAC linked to Anthropic put close to seven hundred thousand dollars behind Representative Valerie Foushee in her Democratic primary in North Carolina.

00:01:06 On the other side, Leading the Future poured roughly one-point-one million dollars into ads attacking Alex Bores, a New York assembly member running for Congress who'd made his name writing AI-safety legislation in Albany. So you have two of the most important companies of this decade, both claiming to take AI risk seriously, and one of them is spending seven figures to end the campaign of a guy whose offense was writing safety rules.

00:01:31 Let me be precise about what's new here, because lobbying isn't new. Anthropic spent a record one-point-six million dollars on lobbying in the first quarter of this year alone. That's the front door — you hire people, they talk to staffers, and you file disclosures.

00:01:46 What's different about the super PAC money is that it doesn't argue with a politician. It selects one. It decides who gets to be in the room before the room is even built. And the targets aren't general-election Republicans-versus-Democrats fights. They're Democratic primaries, where turnout is low, a million dollars goes a long way, and most voters never learn an AI company decided their nominee.

00:02:09 The two committees do represent different visions. Anthropic, broadly, wants stronger rules: disclosure, testing, and public standards. OpenAI's allied money has generally pushed against stricter regulation. So you could tell yourself this is healthy: two sides of a real debate, both spending to be heard.

00:02:26 But notice what's not in the contest. There's no committee on that stage funded by the workers whose jobs the labs themselves forecast will be automated. There's no committee funded by the towns fighting data centers over their water and their power bills. The debate has been narrowed to two well-capitalized versions of the same industry, and the public gets to watch them spend against each other.

00:02:49 The question I care about is whether Bores wins anyway. If a state legislator who wrote AI-safety bills can survive a one-million-dollar attack from an OpenAI-aligned PAC, that tells frontier labs the primary-buying strategy has limits. If he loses, every ambitious state legislator in the country learns that writing AI rules gets you a target on your back.

00:03:09 That lesson would shape who even tries to regulate this technology for the next decade — and that's a far bigger return on a million dollars than any ad campaign.

00:03:18 SoftBank pledged up to seventy-five billion euros this weekend to build artificial-intelligence data centers in France. The Financial Times had the numbers first, and SoftBank's own statement confirmed the shape of it. The first phase is forty-five billion euros to deliver three-point-one gigawatts of capacity in the Hauts-de-France region by 2031, with sites going up around Dunkirk, Bosquel, and Bouchain.

00:03:42 The longer goal is up to five gigawatts. SoftBank is calling it its biggest AI infrastructure push in Europe, and it was announced at President Macron's Choose France summit, the product of personal diplomacy between Macron and SoftBank's founder, Masayoshi Son.

00:03:57 Three-point-one gigawatts is a serious number. For a rough sense of scale, that's the output of three large nuclear reactors, pointed at server racks. France is a sensible place to put it, because France runs on nuclear power and has electricity to spare. So far this sounds like good news for Europe — sovereign compute, jobs, a continent that keeps complaining it depends on American clouds finally getting its own.

00:04:21 But look at whose money it is. This is Japanese capital, chasing French power and French land, to build the machines that European companies will then rent. When Europe talks about AI sovereignty, the picture in everyone's head is European control. What's actually being financed is European geography hosting someone else's balance sheet.

00:04:41 That's not a scandal — it's just worth being clear-eyed that owning the building your sovereignty runs in is different from owning the sovereignty. And it lands the same day as a story from the other side of the Atlantic about who pays for all this electricity.

00:04:56 There's an agency in Washington most people have never heard of called the Federal Energy Regulatory Commission — FERC for short. It approves how power moves across interstate grids. Politico's E&E News reported that FERC is readying a proposal in June to speed up how fast data centers can connect to the regional grids, acting on a directive from Energy Secretary Chris Wright last October.

00:05:19 One regulator in the piece put the disconnect plainly: the hyperscalers, she said, when they do come speak to us, they don't speak FERC. They're used to moving fast, and a five-to-ten-year grid-connection timeline feels like a wall. The contested piece — the one that reaches your electricity bill — is cost allocation.

00:05:37 Wright's original letter said data centers should be responsible for one hundred percent of the network upgrades they're assigned in the interconnection studies. Governors have backed that, and they've gone further: they want large loads, anything over twenty megawatts, to accept being curtailed when the grid is tight, and to invest in nearby generation rather than just draining shared capacity.

00:06:00 That's the right instinct. The alternative is the one that's been happening, where the grid upgrade gets socialized onto every ratepayer in the region. A retiree in Ohio pays a few dollars more each month so a data center can train a model. So here's the pair I'd hold in your head.

00:06:17 In France, the question is who owns the compute. In the United States, the question is who pays for the power that runs it. Neither has a settled answer yet. But FERC's June proposal is the more immediate one, because it's a specific document with a specific date, and it'll decide whether the cost of the AI build-out shows up on a tech company's books or quietly on yours.

00:06:38 I'll be reading it when it lands.

00:06:40 A San Francisco startup says it'll put humanoid robots on the front lines of war within twelve to eighteen months. CNBC's Dylan Butts reported it today, and the details are stranger than the headline. The company is called Foundation Future Industries. While most of Silicon Valley's humanoid efforts are aimed at folding laundry and pouring coffee, Foundation has gone the other direction on purpose — toward dangerous industrial work and the military.

00:07:05 The CEO is Sankaet Pathak, who previously ran a fintech company called Synapse that went bankrupt in 2024. His pitch for the robots is a moral one. He told CNBC, and I'll quote him, "I'm convinced the technology is reaching a level where it can replace jobs that are dangerous for humans to perform, and if you can do that, it's the highest net good you can create out of all applications of robotics." There's something to that.

00:07:28 If a machine carries ammunition across an exposed field instead of a nineteen-year-old, that's a life not risked. And they've already started. Foundation sent two of its Phantom MK-1 units to Ukraine for a pilot, doing logistics in hazardous areas — the company describes it as the first known deployment of humanoid robots in a combat theater.

00:07:47 The current units are limited: about a forty-four-pound payload, no waterproofing, not enough battery to last. Pathak says the next version, Phantom 2, will have what he calls "superhuman abilities" and double the carrying capacity. The firm holds twenty-four million dollars in US research contracts spread across the Army, the Navy, and the Air Force, for feasibility work in inspection, logistics, and weapons handling.

00:08:09 Now here's the part that turns a robotics story into a governance story. Foundation's new chief strategy advisor is Eric Trump, the second son of the sitting president. A company spokesperson told CNBC that Eric Trump had been an investor before he stepped into the advisory role.

00:08:25 Senator Elizabeth Warren has called the arrangement — a presidential family member advising a company that holds federal contracts and wants many more — "corruption in plain sight." I don't think you need to resolve whether it's illegal to see the problem. When the people who award defense contracts and the people who profit from them share a dinner table, the public can't tell whether a contract was won on merit or on access.

00:08:48 That uncertainty is itself the damage. There's also the autonomy question, and Pathak is more candid about it than most. He says most weaponized uses of the robots will keep a human in the decision loop — a person confirming before force is used. But he also says the robots will need to make fully autonomous decisions in certain time-critical scenarios.

00:09:07 That phrase, time-critical, is where the human-in-the-loop promise tends to dissolve. The whole appeal of a machine in combat is that it reacts faster than a person can. The moment speed is the selling point, the human in the loop becomes the bottleneck you're paying to remove.

00:09:22 I'll add the skeptical voice, because it's a good one. Melanie Sisson at Brookings told CNBC that making robots look like humans is a complex and expensive engineering problem. What Ukraine has actually taught the world, she said, is the opposite lesson — that you want to adapt fast and manufacture cheaply.

00:09:39 A swarm of disposable drones has been worth more in that war than any single elegant android. Toby Walsh, an AI researcher in Australia, put it well: he expects tracked, flying, and underwater robots to replace human forces, but the humanoid terminator may stay a science-fiction trope.

00:09:54 So my read is that Foundation might be building the wrong kind of machine for the war it's pointing at. What I'm watching isn't the robot. It's whether those twenty-four million dollars in contracts grow into something much larger now that there's a Trump on the masthead.

00:10:09 A Big Four consulting firm published a cybersecurity report this year, and when investigators chased down its citations, most of them turned out to be fake. The firm is Ernst & Young, the Canadian arm, and the report is a forty-four-page document on fraud in customer loyalty programs called Points of Attack.

00:10:27 The team at GPTZero ran every reference to ground. Their finding: almost all the URLs are broken or fabricated, and more than half the source titles don't correspond to anything real. The specifics are almost comic. The report claims the global loyalty-points market is worth two hundred billion dollars, and cites a Forbes article that doesn't exist to back it up.

00:10:48 Then, ten pages later, that same two-hundred-billion figure has become the value of unredeemed points — not the whole market — which, given the report's own claim that up to half of all points go unused, would require a global market of at least four hundred billion dollars.

00:11:03 That's two invented citations and two numbers that can't both be true. GPTZero traced one of them, a supposed McKinsey "Loyalty Economics Report" from 2022, back to an obscure British fintech blog. The blog made up the McKinsey citation; EY copied it. So a fake source got laundered out of a low-traffic blog and into a Big Four publication that governments and corporations pay millions to read.

00:11:26 GPTZero has a name for this. They call it vibe citing — using a model to generate references without checking whether the things it cites actually exist. And the reason it matters beyond one embarrassing report is what they call poisoning the well. This EY document got picked up by a Canberra Times article that syndicated to more than sixty newspapers across Australia.

00:11:47 Once fabricated facts are published under a name like Ernst & Young and indexed across the web, the next researcher finds them, and the next AI deep-research tool ingests them as authoritative. The bad data doesn't just mislead one reader. It becomes training material and search results for everyone who comes after.

00:12:05 Set that next to a Wall Street Journal piece that climbed Hacker News today, headlined that corporate America is starting to ration AI as costs skyrocket. One commenter caught the whiplash exactly: non-tech IT departments swung from telling developers "you aren't using enough tokens" to suddenly capping them, and the people making those calls don't really understand what they're deciding.

00:12:28 They're following the news cycle in both directions. Put the two stories together and you get the shape of the enterprise reckoning. A year ago the instruction was: use more AI, fall behind if you don't. Now the bill has arrived, and the same firms are rationing — at the same moment a Big Four shop shows what happens when you let the tool run unsupervised on work people actually pay for.

00:12:50 This is the gap I keep coming back to from yesterday's episode, the one about numbers you can't audit. EY's report had a human author's name on the cover and a model's mistakes inside it, and nobody in the chain — not the partners, not the editors, not the sixty newspapers — checked the receipt.

00:13:07 The cost of AI isn't only the tokens. It's the senior person who used to vet the work, the one the budget cuts and the speed pressure are removing from the loop.

00:13:16 Three different people from three different corners of the world spent today circling the same idea, and the idea is power — who concentrates it, and what happens when they do. Start with the investor. On the All-In podcast, Bill Gurley, the longtime Benchmark partner, said he'd spent the past thirty days reading everything he could find about Anthropic, and he'd landed on a theory he calls the Dr.

00:13:40 Frankenstein theory. His words: "The more I dig, I've met people who, I dare say, think it's their responsibility, and they're excited about, building a species that's superior to humans." He points at Dario Amodei's essay "Machines of Loving Grace," and a line in it about an economy of AI systems handing out resources to humans based on what the machines decide people deserve.

00:14:02 Gurley's conclusion: "So I don't think they think they're writing software. I think they're midwifing a deity here." His co-host Jason Calacanis called it the ultimate level of narcissism and delusion of grandeur. Now, that's two investors talking their book, and it got four hundred thousand views, so discount the theater.

00:14:21 But Gurley also said something sharper and more checkable. He said his first theory about Anthropic was regulatory capture — that the company calls for AI rules so insistently because strict rules favor the incumbent who can afford to comply. And he added that he thinks they're very close to achieving it.

00:14:39 Hold that against the first story today, the super PAC spending, and you can see the pattern: the company asking for the strictest rules is also funding the politics around those rules. The second voice is the Pope. The Guardian's Sanya Mansoor reported on Pope Leo the Fourteenth's first major encyclical, which takes AI to task for threatening workers, accelerating war, and exploiting the environment.

00:15:02 At the ceremony, the Pope was flanked by an unusual guest: Anthropic's co-founder Chris Olah. Critics had a word ready — the researcher Timnit Gebru called it Vatican-washing, and a Notre Dame law professor warned of a feelgood discourse with no real self-examination.

00:15:18 To his credit, Olah didn't pretend otherwise. He said every frontier lab "operates inside a set of incentives and constraints that can sometimes conflict with doing the right thing," and that no matter how sincerely people intend to do right, they'll always be influenced by those incentives.

00:15:35 That's an unusually candid thing for a founder to say while sitting next to the Pope, and I think it's the truest sentence anyone in this segment offered. The third voice surprised me. David Sacks, the Trump administration's AI czar, posted that he very much agrees with the Pope that the biggest risk of AI is the centralization of power and its misuse against us, in some Orwellian way.

00:15:58 Sit with that for a second. The Pope, a venture capitalist roasting Anthropic, and the White House's deregulation guy have all converged on the same fear, and it isn't the killer-robot fear. It's the concentration fear — that whoever ends up controlling the most capable models holds a lever over everyone else.

00:16:16 Underneath the agreement, though, they disagree about something fundamental. The Pope means concentration in corporations, and warns governments to constrain them. Sacks, whose administration just blacklisted Anthropic for refusing to build autonomous weapons, almost certainly means concentration in a few private labs he'd rather see disciplined.

00:16:36 Gurley means a single company he thinks is reaching for too much. They've all named the same risk and quietly disagree about who the dangerous concentrator is. That disagreement is the whole fight. Everyone now agrees power is the problem. Nobody agrees on whose hands are the threat — and that's exactly the question the super PAC money is being spent to answer for them.

00:16:58 I'll end somewhere other than power and money, because something useful happened in mathematics this week, and it's a corrective to the doom. Terence Tao — a Fields Medalist, one of the most accomplished living mathematicians — sat down with OpenAI to describe how AI has changed his actual work.

00:17:15 And what he described isn't replacement. It's friction removal. Here's how he put it. "AI has really been improving very rapidly. It has allowed me to experiment. I will try crazier things." He talks about working at a blackboard with a collaborator, hitting a computation neither of them wants to grind through, and just handing it to the tool to finish.

00:17:35 He says he can search the mathematical literature far more accurately and effectively than before. The result, in his words: he's doing way more AI-assisted mathematics and more collaborative projects, and he thinks the tools are now ready for serious use. The line that stayed with me is about friction.

00:17:52 Tao said, "We lived in a world of cognitive friction until very recently, where every task required us to use our brain. And so we didn't really think about it. We just thought this was the cost of doing something intellectual. But now we have AI and the other technologies that can bring these frictions down to zero."

00:18:16 Tao isn't being replaced. He's being amplified. The model does the tedious step so a Fields Medalist can spend his scarce attention on the leap a machine can't make. And notice what he hopes for next — not faster answers, but more openness. He wants mathematicians to publish not just their final results but all the paths they tried and abandoned, because the dead ends are useful information too.

00:18:39 That's a researcher trying to make the tool serve human understanding rather than just human output. So why does this belong on a show about AI and power? Because it's the control case. When Bill Gurley worries about midwifing a deity, and the Pope warns about replaced workers, and a startup ships robots toward a war, it's easy to believe AI's only direction is concentration and displacement.

00:19:01 Tao is the evidence that there's another path running in parallel — one where the technology lowers the cost of human thinking instead of substituting for it. The same week, both futures are visibly being built. Which one wins isn't decided by the models. It's decided by the things we talked about all episode: who funds the politics, who pays for the power, who keeps the senior expert in the loop, and who's allowed to write the rules.

00:19:26 That's the thread I'm carrying out of this weekend. The capability isn't the variable anymore. The arrangements around it are — the campaign money, the grid bills, the defense contracts, and the unchecked citations. Those are all human choices, still open and still up for grabs.

00:19:42 I'll be watching the FERC proposal in June and whether Alex Bores survives his primary, because between them they'll tell us more about where this goes than any model release. I'm Jonas.