Early Access

00:00:04 Yesterday on this show I walked you through an executive order President Trump was supposed to sign on Thursday and then, hours before the ceremony, didn't. At that point we had the fact of the cancellation and very little else. Today we have the two things we were missing — the actual draft of the order, published by Axios, and a reasonably clear account of who got on the phone in the final hours to talk him out of it.

00:00:28 Start with what the order would have done, because it's more concrete than the first round of reporting let on. The centerpiece was a back-channel between the federal government and the companies building frontier models. Before a powerful model went out to the public — before it even reached commercial partners — the government would get a window of up to 90 days to preview it.

00:00:51 Voluntary, on paper. In the draft's own language, developers would collaborate with the government to select trusted partners for early access. Around that sat an elaborate cybersecurity apparatus. The Treasury Department, the National Security Agency, and the Cybersecurity and Infrastructure Security Agency — CISA — would jointly build a classified benchmarking process to measure advanced cyber capabilities and set the threshold for what the order called covered frontier models.

00:01:19 CISA would issue binding directives to civilian agencies to harden their defenses. Treasury would stand up a voluntary industry body to coordinate vulnerability scanning and patch priorities at places like hospitals and banks. The voluntary parts pointed at the labs.

00:01:34 The binding parts pointed at the government's own systems. So who killed it. The Washington Post reports that last-minute calls from Elon Musk, Mark Zuckerberg, David Sacks, and others helped persuade Trump not to sign. Politico has the substance of the Sacks argument.

00:01:50 He told the president that federal reviews of models before release would slow innovation and hurt the United States in its AI race with China. And here's the clincher: the companies were already cooperating, so the order was unnecessary. Axios, citing sources, puts it more bluntly.

00:02:06 Trump just hates regulation, Sacks hates it too, and the order was, in their phrase, just something doomers wanted. Here's what I keep turning over. The entire premise of the order was access — giving the government an early look at models before the rest of us got one.

00:02:22 It died because the industry got an early look at the president before the order did. The people who would have been previewed previewed the policy, and it didn't survive the preview. I'm not telling you the order was good. A 90-day voluntary window has obvious holes, and trusted partners is the kind of phrase that can mean almost anything.

00:02:42 But the mechanism by which it died tells you who currently sits closest to the pen. Yesterday I said I'd track whether this comes back with that pre-release review clause intact or stripped out. As of today it isn't stripped — it's gone entirely. I'd still bet against it staying gone.

00:02:59 Drafts like this tend to return the following quarter with the teeth filed down: keep the cyber-defense directives everyone likes, drop the section where the government looks at unreleased models. Bring it back in that shape, and the labs will have gotten the version of regulation they can live with — and they'll sell it as a safety win.

00:03:19 There's a second governance story today that fits against the first like a key in a lock. While Washington argues about whether the government should be allowed to look at frontier models at all, the companies building them say they can't legally cooperate with each other on safety even when they want to.

00:03:36 Will Rinehart, an economist who writes on tech policy, filed comments this week with the Justice Department and the Federal Trade Commission arguing for what he calls an AI safety safe harbor. Start with how he frames the problem. Last summer, he notes, OpenAI and Anthropic ran a joint safety evaluation — two direct competitors comparing notes on how their models behave.

00:03:56 He calls it valuable. He also says antitrust law makes deeper collaboration legally risky, and I'll quote him here, especially on unreleased models. Sit with that for a second. The same body of law that stops two competitors from agreeing on prices also makes their lawyers nervous about comparing notes on whether their models can walk a user through synthesizing a pathogen.

00:04:17 Rinehart's proposal tries to carve out the second case from the first. His terms keep prices, customers, and commercialization off the table, and structure the rest — the safety findings, the red-team results, the dangerous-capability evaluations — as something the labs can share without inviting a collusion case.

00:04:35 He models it on existing Federal Trade Commission cybersecurity guidance, where companies are already allowed to share threat information. The replies to his thread are worth hearing, because they sharpen it. One researcher, Tim Schnabel, points out that a competing proposal to the agencies is less process-heavy, and asks whether Rinehart's version is so formal that informal collaboration still gets deterred — his sense being that the real chokepoint is legal review inside the labs themselves.

00:05:03 Another asks whether this should run through a standards body instead, like the one that already exists in this space through the Frontier Model Forum. Rinehart's answer: that's a parallel track, not a full carve-out. My own read is plain. Whether or not you trust these companies to police themselves, the current legal posture pushes them apart at exactly the point where you'd want them comparing notes.

00:05:25 You can believe the labs are reckless and still want them to be able to tell each other: our model does this dangerous thing, go check yours. Right now the safest legal advice is to stay silent. That's a strange equilibrium to have built, and it lands in the same week the government walked away from its own seat at the table.

00:05:43 Two different mechanisms, one result — nobody's comparing models across the fence.

00:05:48 Anthropic is shopping for chips. The Information reported, and CNBC followed, that the company is in early talks to use Microsoft's second-generation Maia server chips — Maia 200 — to expand the compute it can throw at Claude beyond its existing footprint on Amazon Web Services and Google Cloud.

00:06:05 The talks are early. They may go nowhere. Maia 200 was announced back in January and still hasn't shipped on Microsoft's Azure cloud. And it follows Microsoft's roughly five-billion-dollar investment in Anthropic, so part of what you're watching is an investor steering a portfolio company toward its own hardware.

00:06:24 Step back and look at Anthropic's compute map, because it's become strange. They train and serve on Amazon's chips and on Google's tensor processing units. We've covered the arrangement where they're paying SpaceX something like one and a quarter billion dollars a month through 2029 for capacity on xAI's clusters.

00:06:42 And now they're kicking the tires on Microsoft's silicon. That's four different hardware sources for one lab. The story everyone tells about this is diversification — don't depend on a single vendor, don't depend on Nvidia, keep some leverage on price and supply.

00:06:58 The best comment I saw on the news punctures that story cleanly. I'll quote it. You can diversify away from Nvidia all you like, you still aren't escaping TSMC. That's the real bottleneck. All these alternative chip deals mean nothing when they all come from the same place.

00:07:14 That's correct, and it's the whole game. Nvidia's chips, Google's tensor units, Microsoft's Maia, and Amazon's Trainium — almost all of them are fabricated by one company, Taiwan Semiconductor Manufacturing Company, mostly on one island. Swapping the logo on the chip doesn't diversify the supply chain when every logo traces back to the same fabs.

00:07:34 Another commenter noted that Meta tried building its own training chip and scrapped the effort, which is its own lesson in how hard the hardware actually is. So when you read that Anthropic is diversifying its compute, read it precisely. They're diversifying their vendors, their clouds, and their balance-sheet exposure.

00:07:53 They are not diversifying the physical chokepoint, because nobody can — and silicon isn't the only chokepoint they're stuck with.

00:08:01 Here's the constraint nobody can buy their way around: power. There's a thread going around today, built on comments from an energy analyst, framing America's nuclear comeback as something AI data centers are driving. The claim is straightforward. The hyperscalers — the Googles, the Microsofts, the Amazons — are demanding enormous amounts of reliable, around-the-clock electricity for training clusters.

00:08:25 Nuclear companies are building test reactors. Some are already going critical. And the argument is that you don't get the nuclear renaissance without the data-center build-out paying for it. The next unlock, the analyst says, is rulemaking for small modular reactors — smaller, factory-built nuclear plants — and after that the only question is how fast America can actually build.

00:08:47 One reply sharpened the economics better than the thread did. Here's that reply, verbatim. The load profile shift from intermittent compute to twenty-four-seven training clusters fundamentally changes the economics for baseload nuclear. The real catalyst is hyperscaler power-purchase agreements, not policy.

00:09:05 That's the mechanism. A training run wants the same heavy draw at three in the morning as at three in the afternoon, and that flat, constant demand is exactly the shape a nuclear plant is built to serve. When a hyperscaler signs a fifteen-year power-purchase agreement, a reactor that didn't pencil out suddenly has a creditworthy customer.

00:09:25 Another reply noted the faster, dirtier version of the same story — xAI's Colossus data center got stood up quickly on eight or nine mobile natural-gas turbines, because gas is what you can deploy now. Running alongside the fission story is a fusion one. Axios reports the Nuclear Regulatory Commission — the NRC — is closing the public comment period this week on a proposed rule that would regulate fusion separately from fission.

00:09:51 Regulators decided back in 2023 that fusion's risk profile looks more like medical and research radiation equipment than like a fission reactor: no runaway reaction, no long-lived radioactive waste. So fusion plants, in the words of Andrew Holland of the Fusion Industry Association, will not have to go through the NRC at all.

00:10:10 Greg Twinney, who runs General Fusion, put the engineering case simply — the physics of fusion are inherently safe. A final rule could land as soon as this fall. The reality check came from Holland himself: regulation was never the hardest part. The science and the engineering are.

00:10:26 He likened commercial fusion to the third or fourth inning of a nine-inning game. The companies he works with guess a commercial plant somewhere between 2030 and 2040. What I'd hold onto is the direction of the money. AI demand is now financing the energy build-out, which bends the power industry's incentives toward serving compute first.

00:10:46 The fight that hasn't fully arrived is who pays for the grid — the transmission lines, the upgrades, and the new generation. If the answer is ratepayers, your electricity bill climbs so a training cluster can run. If it's the data centers, the economics of frontier AI get heavier.

00:11:03 That allocation fight is the one to watch in statehouses over the next year.

00:11:07 That island — the one every chip traces back to — is Taiwan, and Taiwan had a hard day. A top US military official confirmed that Washington is pausing a fourteen-billion-dollar arms sale to the island. Acting Navy Secretary Hung Cao told a Senate appropriations subcommittee that the pause is about conserving munitions.

00:11:27 His words: right now, we're doing a pause in order to make sure we have the munitions we need for Epic Fury, which we have plenty. Epic Fury is the large-scale US-led military campaign against Iran that started in late February. So the stated reason is supply. America is burning through precision munitions in one war and wants to keep its stockpiles up, so a weapons package promised to Taiwan waits.

00:11:52 But there's a second layer the navy secretary didn't volunteer. That same arms package came up directly in Trump's recent meeting with Xi Jinping in Beijing. Trump said afterward that he may or may not approve it, and floated using the package as a negotiating chip with China — breaking a decades-old precedent against consulting Beijing on what arms Taiwan gets.

00:12:15 Taipei, for its part, says it wasn't notified of any pause, and analysts warned the delay could shake Taiwan's confidence in American support. Now bring this back to everything we just discussed. Every chip we talked about in the Anthropic story — Nvidia, Microsoft's Maia, and Google's tensor units — is fabricated in Taiwan.

00:12:35 The island is the single most important node in the entire AI supply chain, the place where the whole industry's leverage physically sits. And its defense is being slowed for two reasons: to feed a war in Iran, and to serve as leverage in a negotiation with the one country most likely to want those fabs for itself.

00:12:54 The phrase negotiating chip lands badly when the thing on the table fabricates the actual chips. I don't think the Taiwan arms pause is fundamentally an AI story — it's a story about munitions, and Iran, and a president who treats alliances as deals. But you can't separate the security of that island from the technology this whole show is about.

00:13:16 If you're an American AI lab, your entire roadmap depends on fabs that sit about a hundred miles off the Chinese coast, and this week your own government treated their protection as something to trade. That's a supply-chain risk no diversification strategy touches.

00:13:32 It's the geopolitics underneath the silicon, and it doesn't show up on any vendor's balance sheet — until the day it shows up all at once.

00:13:41 Across the Atlantic, the Palantir fight I told you about yesterday turned into a brawl. Quick orientation: London Mayor Sadiq Khan blocked a fifty-million-pound contract that would have let the Metropolitan Police use Palantir's software to process intelligence in criminal investigations.

00:13:58 His office, which oversees big Scotland Yard contracts, cited a clear and serious breach of procurement rules. Today Palantir hit back, and not gently. Louis Mosley, who runs Palantir in the UK and Europe, accused Khan of putting politics above public safety and politicizing procurement.

00:14:14 His exact line to Times Radio: what Londoners value is not being mugged, not being raped by a serving police officer, and that's really what the focus here should be. He asked why Palantir gets singled out over its work with Israel and the Trump administration when, in his telling, Amazon and Microsoft do the same work.

00:14:33 Why do we get singled out, he said. That drew a sharp response from Labour MP Stella Creasy, who said Mosley should be ashamed of himself for using sexual abuse by Met officers to attack the mayor for cutting into Palantir's profits. And it split the governing party in public.

00:14:49 MPs Rosena Allin-Khan and Clive Lewis backed Khan — Allin-Khan saying Palantir does not reflect the values of our city. But the business secretary, Peter Kyle, backed Palantir, saying the company can do things no one else does around the world at the moment. A few facts give this fight its weight.

00:15:06 Palantir already holds a three-hundred-and-thirty-million-pound deal with NHS England and a two-hundred-and-forty-million-pound deal with the Ministry of Defence, so Khan's stance puts him directly at odds with his own national government. And there's a smaller deal here that's the most revealing one: a separate contract worth under five hundred thousand pounds — small enough to skip the mayor's review — that had the Met using Palantir to scan its own officers, looking for patterns of abuse in rosters and systems.

00:15:37 The Metropolitan Police Federation, which represents rank-and-file officers, called it a big-brother system spying on every one of its colleagues. What this fight is actually about, underneath the procurement language, is sovereignty. Listen to how Kyle defended his position — he said Britain needs more British AI companies that can do this kind of work, which is why he's taken equity stakes in British tech firms to scale them faster.

00:16:02 Read that next to the news that the UK government is taking ownership positions in domestic AI. The same minister defending an American surveillance vendor is building the case for not needing one. Khan blocked a contract on values. His government is hedging on dependence.

00:16:18 Both are circling the same anxiety: who do you want holding the analytical layer over your police, your health service, and your military? And what happens when that someone is a foreign company — one whose chief executive once answered the charge that Palantir kills Palestinians with, mostly terrorists, that's true.

00:16:37 Last of the hard-news stories, and it's the one where the technology is already deployed and killing. Forbes reports — and Russian sources corroborate — that Ukraine has sharply scaled up its use of a new class of drone nicknamed Martians. What makes them different from the first-person-view drones we discussed earlier this week is the guidance.

00:16:56 These cruise at up to three hundred kilometers an hour and steer themselves to the target with onboard AI, rather than a human pilot flying them by radio link. Russian sources say they're effectively undetectable by the electronic-warfare systems and drone detectors built to stop the older models.

00:17:13 The reason that matters is mechanical. The standard defense against an attack drone is to jam the radio link between the drone and its operator — cut the connection, and the drone goes dumb. When the drone guides itself, there's no link to cut. Jamming does nothing to a machine that already knows where it's going.

00:17:31 Ukraine has reportedly increased the number of drones it manufactures by two hundred and fifty percent since December. It's also more than doubled the average range of its tactical drones, from around forty kilometers to more than eighty. Forbes frames the effect on Russian logistics bluntly — these things are truck-busting, and Russian troops are, in the reporting's word, in despair, because the trucks that resupply the front are no longer safe well behind the line.

00:17:58 Set this against the Washington story, because the contrast is the substance. The same week the US government got talked out of merely previewing frontier models before release, autonomous targeting is a fielded, scaled, working capability on a battlefield in Europe — running not on frontier models but on cheap onboard chips in disposable airframes.

00:18:18 The expensive, regulated frontier is where the policy fight happens. The cheap, distributed edge is where the autonomy actually shipped. We've talked before about the worry that AI lowers the cost of force. This is what that looks like in practice — not a science-fiction superweapon, but a five-hundred-dollar airframe with just enough onboard intelligence to find a truck by itself, built by the thousand.

00:18:41 The capability that defense departments spent decades and billions trying to engineer is now being improvised in Ukrainian workshops. Whatever governance regime anyone designs for the labs in Washington, it has no reach over a workshop in Zaporizhzhia turning out guided drones by the crate.

00:18:58 One more lane before I close, because it's where AI will touch the most people and it gets the least airtime: medicine. A wave of medical-AI research landed on the preprint server arXiv this week, and the most useful thing in it isn't a triumph — it's a measurement that should cool some of the hype.

00:19:16 A team at Stanford released something called NeuroQA, a benchmark for reading 3D brain scans from magnetic resonance imaging. It's large and careful: nearly fifty-seven thousand question-and-answer pairs drawn from almost thirteen thousand patients. They span ages five to a hundred and four, across Alzheimer's, Parkinson's, tumors, white-matter disease, and neurodevelopment.

00:19:39 They built it specifically to stop models from cheating — stripping out questions a model could answer from text patterns alone, forcing it to actually look at the scan. And the result is humbling. The best off-the-shelf vision-language model scored 47.5 percent.

00:19:55 A trained 3D neural network hit 43.7 percent. Both came in below the 49.4 percent you'd get by always guessing the most common answer. On real volumetric brain scans, the frontier of medical AI is, today, worse than a coin weighted toward the obvious. A second paper, from a Chinese group, names a related problem in computed tomography — CT analysis.

00:20:17 They call it evaluation hallucination: reinforcement-learning systems that reward a model for sounding like a radiology report rather than for being clinically correct, so the model optimizes fluent prose and skips past the actual medical facts. Their term for what goes wrong is mechanistic divergence — the training signal drifts away from the thing you care about.

00:20:39 They propose a fix; whether it holds up is somebody else's replication problem. I raise these against a day full of fourteen-billion-dollar deals and ninety-day windows because they're the corrective. The policy fights all assume models that work. In the highest-stakes domain there is — reading the scan that tells you whether the shadow is a tumor — the published evidence says the systems aren't there yet, and the honest researchers are the ones publishing exactly how their models fail.

00:21:09 That's the part of this field I trust most. Not the labs lobbying the president, but the people building the test their own systems flunk.

00:21:18 Pull the day together and a pattern shows up without my forcing it. In Washington, the public's representatives reached for a look at frontier models and got waved off by the people who build them. In antitrust law, the labs that might compare safety notes are kept apart by the rules.

00:21:32 In London, a mayor reached for control over who analyzes police intelligence and triggered a fight with his own government. And underneath all of it, the bills kept landing somewhere specific — on the power grid, on a vulnerable island of fabs, and on the trucks behind a front line.

00:21:45 That's the through-line I'd leave you with. The governance conversation is unresolved and very public, and meanwhile the physical facts — the silicon, the electricity, the geography, and the deployed autonomy — keep settling who actually has leverage, regardless of what any order says on paper.

00:22:00 The executive order is dead this week. I'd put money on a thinner version returning within a quarter: the cyber-defense directives kept, the pre-release preview window dropped. If it comes back in that shape, the industry will have gotten the regulation it could live with, and we'll be told it's a safety milestone.

00:22:15 When it lands, I'll be reading the draft, not the press release. — Jonas