The Only Network That Works

00:00:04 Let me start with a company that spent today doing two opposite things to the same customer. SpaceX raised the price it charges the Pentagon to use Starlink — in the middle of an active war — and on the same day, the US Space Force handed SpaceX a contract worth two point two nine billion dollars to build a space-based military data network.

00:00:23 One hand collecting, the other squeezing, and both pointed at the same buyer. Let me give you the squeeze first, because Reuters reported it and the numbers are sharp. American kamikaze drones — the Pentagon program goes by the name LUCAS — have been flying in the Iran war guided by Starlink.

00:00:40 SpaceX executives argued that the Pentagon had been paying about five thousand dollars per terminal for the connection while, in SpaceX's telling, using a higher tier of service worth closer to twenty-five thousand. So they wanted to charge more. A lot more. On top of that, the reporting says SpaceX floated charging as much as five hundred million dollars to stand up a direct-to-cell capability, plus a hundred million dollars a month to operate it.

00:01:06 The Pentagon pushed back. The drones only use the link for minutes or hours at a time, the argument went, so the higher tier doesn't reflect what's actually being consumed. Pentagon spokesman Sean Parnell denied the framing outright — he said the claims, quote, are simply not based in reality and do not reflect the close, effective collaboration between our teams.

00:01:26 Elon Musk took the other side, saying the military had made improper use of the Starlink civilian system for military purposes, which in his telling violated the terms of service. Here's the part that matters more than who's right about the tiering. Reuters notes there was no rival whose reach matched Starlink's for what these drones needed.

00:01:46 That's the whole ballgame. When you're the only network that works over the territory where the war is happening, you don't have to win the pricing argument on the merits. You just have to be unavoidable. The Pentagon can have the largest defense budget on earth and still find itself negotiating from weakness, because the alternative to paying SpaceX more isn't a cheaper vendor — it's no service.

00:02:09 And the leverage isn't theoretical. The same reporting says US government revenue for Starlink and its Starshield military variant fell more than twenty-five percent in early 2026 compared with a year earlier. So you've got a supplier whose government business is shrinking, looking for ways to raise prices on the customer who can least afford to walk away.

00:02:29 That isn't a stable arrangement. That's a pressure point. Now set the contract next to it. The two point two nine billion dollars from the Space Force is for a military space data network — exactly the kind of orbital communications backbone that deepens the dependency we just described.

00:02:45 So today's news is the dependency growing on both ends at once: the Pentagon paying SpaceX more for the network it already can't do without, and committing billions more to build the next layer of it. I've been watching this dynamic for a while, and I'll say what I think the lesson is.

00:03:02 We spend a lot of energy debating whether a company's model is the best, whether its rocket is the cheapest, or whether its founder is reckless. None of that is the source of the power here. The power is structural. It's that one private actor controls a piece of physical infrastructure the state needs and can't quickly replace.

00:03:20 Everything downstream — the pricing fights, the terms-of-service threats, and the contracts — flows from that one fact. Keep your eye on who owns what has no substitute. That's who sets the terms.

00:03:31 The second story is about a different kind of indispensable thing: not a satellite network, but the few hundred people who can build a frontier model. And China has decided those people are too valuable to let leave the country. Bloomberg reported, and Nathan Lambert flagged on X, that Chinese government agencies have begun imposing overseas travel restrictions on individuals involved in advanced AI work.

00:03:55 This was previously rumored to apply only to DeepSeek. Now the reporting extends it to top talent at private firms, including Alibaba. Lambert's framing was blunt — China, he wrote, begins restricting travel for top AI talent at key orgs. One reply in his thread nailed it: treating researchers like state assets, it said, tells you how seriously they take the AI race.

00:04:17 That phrase, state assets, is the right one. Export controls have always been about chips and equipment — the physical inputs you can stop at a border. This is the inverse. It's a country treating the human beings who carry the knowledge as a strategic resource that can be confined.

00:04:34 You don't restrict the travel of people you consider replaceable. You restrict the travel of people you've concluded are central to national power. Lambert added a line I keep coming back to: I unfortunately only expect more like this. AI is going to pressure so many power structures in the world.

00:04:51 We're still early. I'd take that seriously, because there's a flip side already showing up in the numbers. While China is busy fencing in its researchers, the contest over open models — the ones anyone can download and run — is shifting. Lambert also noted that Google's Gemma 4 is now outpacing Alibaba's Qwen 3.5 and 3.6 in adoption for the same model sizes, and he called that a big shift in the international balance of influence via open models.

00:05:18 For most of the last two years the story of open weights ran through Chinese labs. If Gemma is pulling ahead on actual downloads and deployments, that's American influence flowing back into the layer of the stack that the whole developing world builds on. Soft power, measured in which model becomes the default.

00:05:36 And China isn't standing still on the hardware all of this runs on. Bloomberg also reported that Qualcomm — an American chip designer — reached a deal with ByteDance to supply millions of custom application-specific chips for AI data centers, to power the AI agents inside ByteDance's Doubao chatbot.

00:05:54 Qualcomm's stock jumped about four and a half percent on the news. Sit with the shape of that for a second. A US company supplying the custom silicon that lets a Chinese tech giant scale its own AI agents at home — at the same moment the Chinese government is locking its researchers inside its borders.

00:06:12 The chip trade and the talent control are running on different logics, and they don't fully agree with each other. What I'd watch over the next few months is whether the travel curbs harden into something formal, and whether they catch people on the way back in — researchers at American firms who travel home to visit family.

00:06:31 One person in Lambert's thread raised exactly that fear, and noted it could become a diplomatic crisis. If a Chinese national working at a US lab can't safely visit their parents without risking being kept, that changes the hiring math for every frontier lab in the world.

00:06:47 Talent has been the one input that moved freely. That era may be closing.

00:06:51 Let me connect two things the US government did to itself, because together they tell a story about timing I find unsettling. We've spent several episodes on Anthropic's Mythos model and Project Glasswing — the security work where the company and its partners say they found more than ten thousand high- or critical-severity vulnerabilities in essential software in about a month.

00:07:14 The capability to find software flaws at machine speed is here, and it cuts both ways: defenders can patch faster, and attackers can hunt faster. So the obvious question is whether the agency responsible for defending American critical infrastructure is being built up to match that moment.

00:07:32 The answer, according to Axios, is the opposite. The Cybersecurity and Infrastructure Security Agency — CISA, the lead civilian cyber agency — has lost roughly a third of its workforce since early 2025, through buyouts and budget cuts. The administration proposed cutting up to seven hundred and seven million dollars in funding and eliminating around seven hundred and sixty-six full-time positions.

00:07:56 There's been a partial walk-back. The acting director announced plans to hire three hundred-plus people back for mission-critical roles. But there's still no permanent director, after the nominee withdrew, and no chief AI officer — that position sits empty after the last one left.

00:08:13 And here's the detail that ties it straight back to our Mythos coverage. Axios reports that CISA did not receive initial access to Anthropic's Mythos model, while other agencies did. Remember, a couple of weeks ago we covered the intelligence community securing classified access to that same model.

00:08:31 So the picture is an intelligence apparatus getting the most capable bug-finding tool in the world, and the civilian agency charged with protecting hospitals, water systems, and power grids being left off the list — and short a third of its people. Senator Gary Peters put it this way: starving CISA of personnel, resources and leadership in this high-stakes environment puts our homeland security and national defense at risk.

00:08:57 One source in the piece described CISA's new position as, quote, at the table, not in the game. Why did this happen? The reporting traces a lot of it back to grievance — President Trump associates CISA with its former director Chris Krebs, who publicly affirmed the security of the 2020 election.

00:09:14 So an agency built to defend infrastructure is being run down over an old political fight, at the exact moment the offense is getting cheaper and faster. That's the timing I can't get past. You don't get to choose when the attacks arrive. Now lay the second move next to it.

00:09:31 Wired's Daniel Boguslaw reported that the Department of Homeland Security, the FBI, and other agencies have introduced a new domestic threat category they're calling anti-tech violent extremism. A document from the New York Intelligence and Counterterrorism Bureau uses the phrase anti-tech violent extremist activity to describe opposition to AI and to data center construction.

00:09:54 This is landing as Americans get more anxious about AI taking jobs, and about what the data center buildout does to their water and their electricity bills. So read the two together. The government is cutting the agency that defends against real, technical attacks on infrastructure — and at the same time it's standing up a surveillance category aimed at the people protesting that infrastructure.

00:10:18 Critics quoted in the coverage drew the obvious line to how fusion centers were turned on Occupy and Black Lives Matter and environmental movements. I don't think every data-center opponent is about to be treated as a terrorist. But naming a category is how a label gets operational.

00:10:35 Once anti-tech violent extremism exists on a form, somebody has to fill it in. I'll be watching who ends up in it.

00:10:42 There was a strange and revealing scene yesterday in Rome, right at the seam between what these companies say AI is and what it's doing to people's working lives. Chris Olah — co-founder of Anthropic and the head of its interpretability research, the work of trying to understand what's actually happening inside these models — spoke at the Vatican.

00:11:01 The occasion was the presentation of Pope Leo the Fourteenth's encyclical on AI, titled Magnifica Humanitas. We've covered the encyclical itself before; the Pope's line that stuck was that technology is never neutral, because it takes on the characteristics of those who devise, finance, regulate, and use it.

00:11:19 What stuck with me is what Olah brought into that room. He said, of the models his own company builds, that we keep finding things that are mysterious. He pointed to evidence of what he called introspection inside the models — signs that they represent something about their own internal states — and he spoke about large-scale labor replacement as a serious near-term prospect.

00:11:40 Set the venue aside and that's a remarkable thing for a builder to say in public: we made this, it's already reshaping work, and we don't fully understand it. Now, I try to hold two thoughts here at once. The first is that Olah is, by reputation, one of the more careful and less hype-prone people in this field.

00:11:57 Interpretability is the corner of the lab where people admit what they can't explain. So when he says mysterious, I don't think he's reaching for drama. The second thought is the one the skeptics raised immediately, and it deserves airtime. On the thread about his remarks, the top replies were sharp.

00:12:14 One person wrote that this alarmist rhetoric is working well for Anthropic for raising money and buying regulatory capture, where they'd be inside the tent instead of outside it. Another just said: this is marketing. They want more funding and for the government to take interest.

00:12:30 I don't think those two readings cancel out. They can both be true. A model can be mysterious to its makers and also extremely useful to dramatize when you're raising money and shaping the rules you'll be governed by. The mystery framing serves a company in Anthropic's position: it justifies why they should be trusted to self-govern, why they should be in the room when regulation gets written, and why the public should defer to the people who, by their own account, don't fully understand what they've built.

00:12:59 Where I'd push, and where the Pope's line actually bites, is on the labor part. Mystery is a story about the model's insides. Labor replacement is a story about someone's paycheck. Those aren't the same altitude. If large-scale labor replacement is close enough to mention at the Vatican, then the question that matters isn't whether the model has introspection — it's who absorbs the cost when the work goes away, and whether the people building these systems carry any of it themselves.

00:13:26 The encyclical puts that squarely on the builders. The builders, so far, are mostly talking about the mystery.

00:13:32 If labor replacement still sounds abstract, here's the physical version of it, and it has a clock attached. A Chinese company called EngineAI shared a look inside its Shenzhen manufacturing base this week, claiming it can produce one humanoid robot every fifteen minutes.

00:13:48 Do that math out and it's about thirty-five thousand humanoid robots a year from a single line — which, as far as anyone tracking this can tell, is the highest production rate any Chinese humanoid robotics company has publicly claimed. And EngineAI says it has another line planned in Zhengzhou rated for ten thousand a year on top of that.

00:14:08 The poster who surfaced it estimated that, between EngineAI and rivals like Unitree, AgiBot, and Leju, China is positioning to output something on the order of a hundred thousand humanoid robots a year. Now, one caution: a production-rate claim isn't the same as a hundred thousand robots doing useful work.

00:14:26 The most upvoted question in the discussion was the obvious one — and I mean that literally, it's the question everyone should ask: where are these robots actually going? What's the market? You can build a line that stamps out a humanoid every fifteen minutes and still not have thirty-five thousand jobs for them to do.

00:14:44 So treat the number as a statement of manufacturing intent, not deployment. China has decided to build the capacity first and find the demand second. That's a bet, and it's a very Chinese industrial bet — the same playbook that gave them solar panels, batteries, and electric vehicles.

00:15:00 Build at scale until the unit cost collapses, then let the cheap supply create its own demand. And there's a demand signal showing up on the other side of the world that fits. Y Combinator highlighted a startup called Twolabs that's building humanoid robots for caregiving, starting in nursing homes — where, as they put it, caregivers are overworked and elderly residents often need more support than the system can provide.

00:15:25 That isn't a gimmick market. The caregiving shortage is demographic and getting worse across the developed world. There aren't enough human aides, the work is hard and underpaid, and the population that needs care keeps growing. So here's the pairing I'd sit with.

00:15:40 On one side, a Chinese factory tooling up to produce humanoids at the rate of a car plant. On the other, a Western startup pointing those machines at the most intimate, most human work we have — caring for the old and the frail. The manufacturing capacity and the social need are converging, and they're converging fastest in exactly the place where the labor market is most strained and the people on the receiving end have the least power to say no.

00:16:06 I don't have a clean verdict on this one. A robot that helps an overworked aide lift a patient safely is a good thing. A nursing home that replaces human contact with a machine because the machine is cheaper is a different thing, and the line between them is going to get blurred by economics.

00:16:23 What I'll be tracking is which version actually ships — augmentation or replacement — because that choice won't be made by the robots. It'll be made by whoever's paying for the care.

00:16:34 A while back I promised to keep watching who ends up paying for the power that all this computing demands — the grid upgrades, the new generation, and the deals that get cut to keep the data centers fed. Today there's a payment of a very particular kind, and it comes out of the Cold War.

00:16:51 The US Department of Energy announced today that it has selected five companies to enter advanced talks about using America's surplus Cold War-era plutonium as reactor fuel. The lead name is Oklo, the advanced-nuclear company; the others are Exodys Energy, SHINE Technologies, Standard Nuclear, and Flibe Energy.

00:17:09 The government plans to make about twenty metric tonnes of plutonium — drawn from dismantled nuclear warheads — available to US power firms. Oklo's stock rose more than five and a half percent on the day, to around sixty-nine and a half dollars a share. The company said it would develop the fuel with a European partner, newcleo.

00:17:28 Let me translate why this is happening now. Advanced reactors need fuel, and standing up a fresh fuel supply chain takes years. The government is sitting on a stockpile of weapons-grade material it would otherwise have to spend billions to dilute and dispose of.

00:17:44 So the plan, in Oklo's own words, is to use existing surplus material as a bridge fuel for advanced reactors to bring more reactors online sooner. The chief executive of newcleo framed it as a way to reduce US nuclear liabilities. In plain terms: turn the bomb cores into electricity, and do it fast, because the demand curve for power — driven hard by AI data centers — is bending up and the grid is the bottleneck.

00:18:08 Now here's where I'd slow down, because this is where the trade-offs live. Plutonium has a half-life of twenty-four thousand years and has to be handled in protective gear at heavily guarded sites. And the twenty tonnes in question isn't a trivial amount of material.

00:18:24 In a letter last September, Senator Ed Markey, along with Representatives Don Beyer and John Garamendi, warned that twenty metric tonnes of weapons-usable plutonium is enough for roughly two thousand nuclear bombs. They argued the plan, quote, raises serious weapons proliferation concerns, makes little economic sense, and may adversely affect the nation's defense posture, and they asked the Energy Department to cancel it.

00:18:49 The department, for its part, didn't immediately respond to a question about how the program would keep the material safe. And there's a conflict-of-interest thread you should know about. The current Energy Secretary, Chris Wright, sat on Oklo's board of directors before joining the cabinet.

00:19:06 Oklo is the lead company in the program his department just announced. I'm not alleging anything improper — board service before government is common, and recusal rules exist. But when the official overseeing a multi-company nuclear fuel program used to serve on the board of the company at the front of the line, that's a fact the public is entitled to weigh.

00:19:27 So tie it back to the through-line. The power demand from AI is now strong enough that the US government is willing to feed weapons plutonium into the civilian grid to get reactors online faster, over the objections of senators worried about proliferation. That's the cost showing up in a form most people never imagined when they heard the phrase data center.

00:19:48 The bill for the buildout is being paid, in part, in plutonium.

00:19:52 Let me end by holding two registers next to each other, because today gave us both, and the gap between them is what I'd ask you to carry out of this. In one register, the future is enormous and it's arriving on a schedule. Demis Hassabis, the head of Google DeepMind, told Axios today that he now thinks artificial general intelligence — machines that match humans across the board — could arrive as soon as 2029, pulled in slightly from his earlier sense of around 2030.

00:20:19 He was candid about why he talks the way he does: this is partly why I use some of the terms I used, he said, which were a little bit provocative. He wants to create urgency, because in his view economists and society are still not taking this seriously enough.

00:20:34 Meanwhile a researcher named Yo Shavit announced he'd left his job at OpenAI to join the OpenAI Foundation, building what they're calling an AI Resilience program — and he wrote that there's a great deal to do before superintelligence, and little time to do it.

00:20:49 So the people closest to the frontier are telling you the big thing is close, and they're choosing provocative words on purpose. In the other register — the one I've spent most of this episode in — the future is mundane, physical, and already here. It looks like a satellite network charging the Pentagon five times more because there's no alternative.

00:21:09 A government deciding its researchers can't leave the country. A cyber-defense agency losing a third of its staff while the tools to attack it get cheaper. A humanoid robot coming off a line every fifteen minutes, and weapons plutonium routed into the power grid to keep the servers running.

00:21:26 Here's where I land on the gap. The 2029 talk may turn out to be right, and I'm not dismissing it. But notice that it asks you to look up and out, at a horizon, while every story I covered today is about leverage being exercised right now, in the present tense, by actors who already have it.

00:21:43 The mystery framing, the countdown to general intelligence, and the resilience programs are all arguments about a future nobody can verify yet. The price of Starlink, the travel ban, the CISA cuts, and the plutonium are settled facts with names and numbers attached.

00:21:58 When someone points at the horizon, it's worth checking what's moving at your feet. I keep returning to that one structural idea from the top of the show: power flows to whoever controls what can't be replaced. The network with no rival. The researcher who can't be substituted.

00:22:14 The model the spies want. The reactor fuel sitting in a vault. Whatever general intelligence turns out to be, the contest over who owns the irreplaceable is happening today, and the bills are already coming due — in dollars, in liberty, and now in plutonium. That's what I'm watching tomorrow: not the date on the horizon, but who's becoming unavoidable before it arrives.

00:22:35 I'm Jonas.