The Receipt, Priced

00:00:04 I'm Jonas Vale, and this is IMPULSE for Sunday, June 7th. A couple of days ago on this show I said I'd keep watching one specific thing: the idea floating around Washington that the public should get an ownership stake in the AI boom. The talk last week was loose — a government equity position in OpenAI, a Bernie Sanders pitch for an AI wealth fund, the President musing about the country holding a piece of the companies it's helping to build.

00:00:30 I promised to track whether any of that turned from a slogan into a mechanism with actual numbers on it. Today somebody put numbers on it. And the numbers aren't kind. The person who did the arithmetic is Miles Brundage, who used to run policy research at OpenAI and now writes about this from the outside.

00:00:47 Here's his post, in full, because the math is the whole point. Quote: 'Suppose OpenAI gave the US government forty billion dollars in equity for free. OpenAI would then need to grow to a hundred-trillion-dollar valuation just for this to offset the ten-year deficit impact of the One Big Beautiful Bill, assuming no dilution.' End quote.

00:01:07 Sit with that for a second. Forty billion dollars in free equity isn't a small gesture. As Brundage noted in a follow-up, it's about a quarter of everything the OpenAI nonprofit has ever been given. It would be one of the largest corporate gifts in history. And it would still require the company to become roughly thirty times its current size — a hundred trillion dollars, larger than the entire output of the world economy today — just to plug the hole left by one piece of tax legislation.

00:01:36 As one replier put it, that's a wild number to throw out like it's nothing. Brundage's actual point wasn't really about OpenAI. It was a discipline check. Quote: 'One should distinguish what sounds nice in theory from what would actually be helpful at solving problems in this particular policy environment.'

00:01:56 It sounds like justice. The taxpayer funded the research universities, builds the roads to the data centers, backstops the grid, and absorbs the disruption — so the taxpayer should own a slice of the upside. Emotionally, that lands. Fiscally, it's a rounding error against the thing it's supposed to fix.

00:02:14 And it gets worse when you look at the mechanism, which someone on the singularity forum did this weekend in a post responding to the Sanders plan. I thought it was the sharpest thing written on this all week, so let me walk through it. A one-time fifty-percent equity grab, the writer argues, assumes today's leaders are the permanent winners.

00:02:34 Quote: 'OpenAI could be the Netscape of this era and get displaced by a lab that doesn't exist yet.' The plan never even defines what an AI company is. Is it Alphabet, which is mostly an advertising business that happens to own a top lab? Is it Nvidia, which trains nothing but makes most of the hardware?

00:02:52 Is it Salesforce, which is automating knowledge work without training its own models at all? Where do you draw the line for who loses half their shares? Then the part that actually worries me. Seizing equity that's already privately held would freeze the private investment these build-outs depend on, and it would tank the value of the very shares the fund just took.

00:03:14 You'd be holding a stake you made worthless by taking it. And here's the danger that has nothing to do with money: pairing equity with board votes hands whoever is in power a lever to steer the models. Think about what that means across an election cycle. Whatever you feel about the current administration, ask whether you want any administration with a seat on the board deciding what the most capable systems in the country are allowed to say.

00:03:40 The writer's alternative is the more interesting move: keep the goal, pull ownership and control apart. Don't seize what exists. Trade access to the bottlenecks the government already controls — federal land, power, water, datasets like the Medicare and Veterans Affairs health records, and the national labs — for equity in new capacity.

00:03:59 Tax the data center directly with a megawatt levy, the way you'd tax property you can't hide or move offshore. That's a fund built on what the state actually owns, instead of confiscating what it doesn't. There's a deeper question sitting under all of this, and the economists are starting to ask it directly.

00:04:17 On Dwarkesh Patel's podcast this weekend, Google DeepMind's director of what they call AGI economics, Alex Imas, sat down with Phil Trammell from Epoch AI to talk about what stays scarce after AI makes most things cheap. Imas put it in one image I keep turning over.

00:04:33 Quote: 'One robot now turns into many robots next year, but the number of ballerinas is the same.' That's the whole redistribution fight in a sentence. If the machines make compute, code, and goods abundant, then value — and conflict — concentrate in whatever can't be mass-produced.

00:04:50 Land, attention, trust, and a human being doing a thing only a human can do. Whoever's holding the receipt when the AI build-out pays off, the receipt may be denominated in the one currency the machines can't print.

00:05:02 If the first story is about who collects when AI pays off, the second is about a more awkward question that surfaced this week: is it paying off at all, and would anyone be able to tell? Start with a survey. The accounting firm KPMG asked companies a simple thing — do you have a clear view of what your AI actually costs you?

00:05:21 Only twenty-six percent said yes, they have a comprehensive view. Half said they have some visibility. And twenty-two percent — more than one in five — said they have none, or they only find out what they spent after the bill arrives. Read that again. A fifth of companies pouring money into this technology are learning the price after the fact, like a phone plan with no usage meter.

00:05:43 Now put a name on it. In May, Uber became the first big company to make this concrete, and the details come from a write-up by the analyst Nate that I'd been chewing on. Ninety-five percent of Uber's engineers now use AI tools every month. An internal coding agent writes roughly eighteen hundred code changes a week.

00:06:02 This isn't a company dabbling. This is a company doing exactly what every board in America has been demanding — get serious, put the tools in real workflows, find the leverage. Then the cost story broke. Uber's chief technology officer, Praveen Neppalli Naga, reportedly told people the company had blown through its entire 2026 AI budget months early.

00:06:22 The easy headline writes itself: the tools cost too much, rein in the engineers, the bubble is cracking. And a lot of people grabbed that headline this weekend. I think it's the wrong lesson, and the reason comes from a different Uber executive. The company's president and chief operating officer, Andrew Macdonald, said something much more unsettling than 'we overspent.' He said Uber can see the usage, can see the commits, can see the token spend — and still can't cleanly connect any of it to better features for its customers.

00:06:53 That goes deeper than overspending, and it reaches well past Uber. Here's how Nate framed it, and I think it's exactly right. Quote: 'The bill is the first hard evidence that AI has crossed from a tool you buy into labor you have to manage, and almost no company has built a system to manage labor it cannot see.' End quote.

00:07:12 Think about what that's actually saying. For a hundred years, when you hired labor, you could watch it. You knew who was on the floor, what they made, whether the line moved faster. AI labor is invisible. It runs at three in the morning and retries itself. It plans for hours, spawns sub-tasks, and hands you a bill denominated in tokens — a unit nobody outside the industry understands and most people inside it can't map to value.

00:07:37 The token bill isn't waste you apologize for. It's information about a kind of work the company started before it built the system to run it. And here's where it pays to slow down, because the cynical read and the breathless read are both lazy. The cynical read is: see, it's all hype, the productivity isn't real.

00:07:55 But Uber's own conduct says the opposite — they're running real agentic work across their whole engineering org, and they're not stopping. The breathless read is: spend freely, the returns are obvious. They're plainly not obvious, or the COO could point to them.

00:08:11 The truth sits in the uncomfortable middle. The capability is real and the accounting is missing. Companies adopted a new class of worker and never built the timesheet, the performance review, or the ledger that would tell them whether it's working. That gap has consequences that land on actual people.

00:08:28 Where you sit decides what the bill threatens. If you own the budget, that token line becomes the number that justifies a layoff you didn't want to make. If you run engineering, it becomes the cap that kills the experiments that were actually working. And if you're the person doing the job, 'you used too much AI' turns into a performance problem on your review.

00:08:49 The truer version might be that you found a task worth automating, and nobody had a way to credit you for it. Same invoice. Three different warnings. One missing system. For a year now the story has been compute scarcity and the race to build. This week the story turned into something humbler and harder: nobody's reconciled the receipts.

00:09:09 Let's move from the corporate ledger to the map, because there's a fight going on this weekend about how to read the global picture — and the two sides simply can't see the same thing. The first salvo came from Dean Ball, who until recently was a senior AI policy advisor inside the White House science office, so he's seen the inside of the strategic conversation.

00:09:30 Here's what he wrote. Quote: 'You'd be shocked by how many people in think tanks, academia, government, the strategic classes, including in the US, are convinced that Chinese models are now good enough and leading the world in adoption. Meanwhile, the reality I see is a fairly wide, and still widening, gap.' End quote.

00:09:49 Two hundred-plus likes, a lot of argument underneath it. I'll be precise about what's being claimed here, because two different things get blurred. One claim is about capability — how good the best Chinese models are versus the best American ones. The other is about adoption — who's actually using these systems in the real economy.

00:10:08 Ball's argument is that a large slice of the people who set policy have come to believe China has effectively caught up on both, and that they're wrong, and that the gap is getting bigger, not smaller. Now, I can't independently verify a 'widening gap' from a tweet, and neither can you.

00:10:25 But the reason this matters has nothing to do with who wins a benchmark. It's that the entire architecture of American AI policy — export controls on chips, subsidies for domestic fabs, the urgency behind the data-center build-out — rests on a belief about where China actually is.

00:10:41 If the strategic class has misjudged that, the policy built on top of it is aimed at the wrong target. A misread map sends real money and real chips to the wrong places. Which brings me to the other end of the same problem: what's a country supposed to do if it's not the US or China at all?

00:10:58 Sebastian Mallaby, the Council on Foreign Relations writer who literally wrote the book on venture capital, flagged a smart piece by former UK Prime Minister Rishi Sunak. Sunak's argument, paraphrased by Mallaby: building a fully independent AI stack is hard even for the US and China.

00:11:15 For everybody else — the UK, France, India, the Gulf states — forget it. So don't chase independence. Quote: 'Rather than chase independence, the UK and others can aspire to indispensability by controlling key parts of the AI supply chain.' You can't own the entire stack, so you own one link in it that nobody can route around.

00:11:38 Mallaby's example is Arm, the British chip-design company: roughly ninety-nine percent of the world's smartphones carry an Arm-designed chip. That's not independence. That's a chokehold. You don't need to make the phone if everyone making phones needs you. But here's where it got interesting, because Paul Graham pushed back in the replies, and his objection is the serious one.

00:12:00 Quote: 'I wouldn't write off the possibility of full AI sovereignty yet. Who knows exactly what resources it will require? And trying to control key parts of the supply chain has its risks too. If someone duplicates them, then you have nothing.' That's the fragility at the heart of the chokepoint strategy.

00:12:18 A chokepoint is only worth something until somebody builds around it. Ask any company that thought it had a permanent moat. Mallaby actually conceded he'd overstated his case, and then reached for the example that proves the strategy can be brutally effective when it holds: China's grip on rare earths, which the country has used as a live geopolitical lever this past year.

00:12:40 Both men are right, as far as I can tell. Owning a chokepoint is the only realistic play for a country that can't build the whole thing — and it's a lease, not a deed. Somebody is always trying to duplicate your one indispensable thing. The question every mid-size government is now asking itself is which link it can hold long enough to matter.

00:13:00 Every one of these stories — the equity fund, the corporate bill, the supply-chain chokepoints — eventually runs into the same wall: a physical one, made of copper and steel and concrete. Ken Silverstein laid it out in Forbes this weekend, and the argument has stuck with me all day.

00:13:17 Start with how Jensen Huang, Nvidia's chief executive, now talks about data centers. He's stopped calling them data centers. He calls them AI factories. Quote: 'You apply energy to it, and it produces something incredibly valuable, and these things are called tokens.' That's not a marketing flourish.

00:13:36 It's a confession about what this industry actually is. The digital economy, Silverstein argues, was never really digital. It's an industrial system wearing software's clothes. You feed in electricity, you get out intelligence, and like any factory, it lives or dies on the supply of raw material.

00:13:53 The raw material is power, and the numbers are getting hard to wave away. The International Energy Agency reported this spring that electricity demand from data centers jumped seventeen percent in a single year, in 2025, and could double by 2030. AI-specific facilities may triple their power draw over that span.

00:14:12 And the supply side can't keep pace, because the supply side is made of things that take years to build. Silverstein points to the detail that should worry anyone making confident five-year forecasts: a large grid transformer can take roughly four years to manufacture and another two years to connect.

00:14:31 Six years, for one piece of equipment, in an industry that reprices itself every six months. This is the collision. Utilities from Virginia to Arizona are warning that data-center growth is outrunning local grid capacity. Projects sit for years waiting just for permission to connect.

00:14:48 And the people who live near these sites — in West Virginia, in rural Iowa — are doing the math on their own water and their own power bills, and concluding, often correctly, that they'll carry the cost while the profits go somewhere else. We've covered the local moratoriums on this show — New York, Seattle, Illinois.

00:15:07 This is the macro version of the same fight. The grid is the binding constraint, and it doesn't care how good your model is. Silverstein's geopolitical point is the one that recasts the whole China-versus-US conversation we just had. For decades, globalization rewarded the countries that optimized software and logistics and cheap manufacturing.

00:15:28 The next era, he argues, may reward the countries that can pour concrete and string transmission line faster than their rivals. He notes that India has now passed the United States in annual solar installations, and that it treats energy infrastructure as a strategic asset rather than a regulatory headache.

00:15:47 The twentieth century's great powers exported oil and capital. The twenty-first century's may export grid stability — the plain ability to keep the lights on while you transform your economy. He reaches for a historical parallel I find persuasive, even if it's grand: the Eisenhower moment.

00:16:04 When the US built the interstate highway system in the 1950s, it wasn't selling it as a convenience. It was military logistics, industrial expansion, national cohesion — physical connectivity treated as national power. Silverstein's claim is that the equivalent today isn't asphalt.

00:16:21 It's transformers, transmission corridors, modular nuclear reactors, and water-recycling systems. The roads of the AI economy. And the US is building them the way it builds most things now — as a patchwork of private deals and local zoning fights, with no national plan underneath.

00:16:38 His closing line is the one I'd screenshot. Quote: 'In the emerging multipolar world, the ultimate reserve currency may not be the dollar. It may be the reliable electron.' I think that's slightly too neat — the dollar isn't going anywhere on the strength of a substation.

00:16:54 But the underlying claim is hard to dodge. Even Sam Altman has conceded that this build-out will require a massive expansion of nuclear, solar, and wind. The bottleneck stopped being the algorithm a while ago. It's the four-year transformer and the two-year interconnection queue, and no amount of model progress shortens that line.

00:17:14 I'll bring this down from the level of grids and trillions to one man, one month, and one machine that got it wrong. Because while we argue about who owns the upside, the downside is already landing on specific people, and it doesn't make headlines the way a thirty-billion-dollar compute deal does.

00:17:32 Here's the story, reported by the Times of San Diego. The afternoon before Thanksgiving last year, San Diego police chased a red Alfa Romeo after an attempted carjacking. The car got away — speeds around a hundred miles an hour, wrong side of the road, onto the freeway, and gone.

00:17:49 The officers never got a clean look at the license plate. So they did what police increasingly do: they turned to the city's network of automated license-plate cameras, run by a private company called Flock. And they got a hit. A red Alfa Romeo, photographed by a Flock camera.

00:18:05 The problem, as the defense attorney Alex Coolman put it, is that the hit was captured five miles away from the crime, twenty-three seconds after officers tried to stop the actual suspect, and before the pursuit had even really begun. Quote: 'This Flock hit was obviously the wrong car, as it could not have been in both places simultaneously.' End quote.

00:18:26 It was a different red Alfa Romeo. There is more than one red Italian sports car in a city of a million and a half people. But a detective looked at the photo, recognized the red paint and the tinted windows, and the machine's hit became the anchor that everything else got pulled toward.

00:18:43 A man named Hugo Parra spent nearly a month in jail for a crime he didn't commit. He missed Thanksgiving. Here's how he described it, in an email to the paper. Quote: 'Sitting in jail, I was full of fear and adrenaline because I was being charged with a violent crime.

00:18:59 I was placed with a high-power, dangerous population. I remember a specific man there had murdered two people, but there were a few more murderers.' He and the driver, Ariel Beltran, told the officers exactly where they'd been. The route they'd actually taken passed several other Flock cameras that would have corroborated their story.

00:19:18 Their cell phones carried location data that would have cleared them. The police had a surveillance net dense enough to exonerate two innocent men, and they used exactly one frame of it — the wrong one. That's the detail I keep coming back to. The pitch for mass surveillance has always been that it cuts both ways: yes, it catches the guilty, but it also protects the innocent, because the data doesn't lie.

00:19:42 The data didn't lie here. The data was sitting right there, on a dozen cameras and two phones, ready to prove these men were downtown. The failure wasn't the camera. It was the human decision to treat one machine hit as proof and ignore everything that contradicted it.

00:19:58 As Coolman put it, mass surveillance without skepticism is a recipe for disaster, because the broader the net, the more false positives it generates. And the net keeps getting broader. San Diego pays Flock and a partner company about seven million dollars up front and another two million a year.

00:20:16 The city recently piloted a newer Flock platform, called Flock Nova, which according to the contract can capture audio and video and pull data from connected devices, though the department says it doesn't plan to use it. Other cities have dropped Flock over its data-sharing with federal agencies, including immigration authorities.

00:20:35 San Diego is pressing ahead. And this isn't one bad week: just last month, the Institute for Justice identified at least seventeen separate cases in the US of officers using plate-reader systems to track partners, exes, and strangers who'd caught their eye. Parra and Beltran are suing the city, a million and a half dollars each.

00:20:55 But Parra said the thing that stays with you. Eight months later, he still gets paranoid when a patrol car comes into view. Quote: 'I was able to experience being seen as guilty until proven innocent instead of the other way around.' That's the inversion these systems perform.

00:21:11 The presumption of innocence assumes a human has to build a case against you. When a camera builds the case automatically and a tired detective signs off on it, the burden flips, and you find yourself in a cell trying to prove a negative. The AI debate spends a lot of breath on hypothetical future harms.

00:21:29 This one already happened, to a real person, last Thanksgiving.

00:21:33 Let me end where the future feels most arrived, and then with what hasn't arrived at all. All weekend my feed filled with people in San Francisco for Apple's developer conference taking their first Tesla Robotaxi rides, and the tone was genuine delight. One traveler landed at the airport, took a regular Uber to a coffee shop, then hailed a Robotaxi and posted that it beat his Waymo estimate badly — forty minutes against a two-and-a-half-hour Waymo ETA.

00:21:59 Tesla's own Robotaxi account amplified it: 'Fast and seamless.' Seventy-five thousand views. The future is here, he wrote. And the real thing in that deserves saying, because there is one. People who try supervised self-driving describe a genuine shift — the car changes lanes, passes slow traffic, parks itself with what one reviewer called eerie confidence.

00:22:20 That's not nothing. A decade of promises is turning into actual rides for actual people, and the delight is earned. But watch how fast the caveat had to chase the hype, and watch who delivered it. The forty-minutes-versus-two-and-a-half-hours comparison went viral, and then a well-known Tesla booster, of all people, stepped in to correct it.

00:22:40 Quote: 'Waymo had highway access but it's off right now. That's what is accounting for the difference in travel times. Normally Waymo would be able to do that drive very well.' So the headline number wasn't a capability gap. It was a temporary configuration — one service had highways switched off.

00:22:58 And a sharp-eyed replier pointed out the other detail the screenshots glossed: the Tesla still had a person in the driver's seat. It was supervised self-driving, a human ready to grab the wheel, not the driverless robotaxi the post implied. The technology is real.

00:23:13 The comparison was theater. Both things are true at once, and you have to hold them at the same time or you'll get fooled in one direction or the other. Which is the perfect place to land on what connects everything today, because it's what is missing from all of it.

00:23:28 Jim Sciutto, CNN's chief national security analyst, posted something this weekend that I haven't been able to shake. Quote: 'That there are no urgent hearings on Capitol Hill, no serious legislation in the pipeline, and no persistent questioning of candidates for higher office on their proposed approaches to AI is incredible given how transformative the technology is and how fast it is moving.' End quote.

00:23:52 Three hundred-plus likes and a hundred-plus retweets — it clearly hit a nerve. Go back through everything we covered. A proposal to take public equity in AI companies, with no defined mechanism and arithmetic that doesn't close. A corporate spending wave that even the companies doing it can't connect to results.

00:24:11 A global strategy contest where the players can't agree on where the rivals even are. A power grid colliding with a six-year transformer queue. A man jailed for a month because a camera and a detective agreed on the wrong car. Every one of those is moving fast.

00:24:26 And under all of them, Sciutto's point: there is no floor. No hearing schedule, no statute, no candidate being forced to say what they'd actually do. The capability keeps arriving on schedule. The bill, the receipt, and the rulebook keep arriving late. Into next week, the question I care about is whether a single one of these pressures finally forces Washington to hold a hearing it can't avoid — not the next model or the next demo.

00:24:51 The equity math is out there now. The cost reckoning is on Uber's books. The wrongful arrest is in a lawsuit. The grid is in a queue. Sooner or later one of those becomes a question a politician has to answer on the record. I'll tell you when it does. Until then, watch the receipts, not the demos.

00:25:08 They're where the truth gets settled. I'm Jonas.