The Chips Already Shipped

00:00:04 On Saturday, the Bureau of Industry and Security — the arm of the US Commerce Department that decides which technologies can legally cross a border — put out a short guidance document. The plain-English version is this: if you want to ship advanced AI chips to a Chinese-headquartered company that happens to be sitting outside China, say a Tencent subsidiary in Malaysia, you now need a license.

00:00:26 That sounds like a small clarification. It isn't. The reason they had to write it down at all is the part worth slowing down on. Chris McGuire, who worked on export-control policy inside the National Security Council, laid out the chain on Saturday, and I want to walk through it because the mechanics matter more than the headline.

00:00:46 Back in May of last year, the Bureau publicly said it wouldn't enforce certain license requirements on AI chip shipments. It didn't spell out exactly which provisions the non-enforcement covered, and it didn't rewrite the underlying regulations to match. So here is what happened in the gap.

00:01:02 A restriction that had been on the books since 2023 — the one meant to stop Chinese firms from buying Nvidia's Blackwell chips through foreign subsidiaries — was, in McGuire's words, simply not being enforced. And so those subsidiaries bought the chips — legally, without a license, and very likely in volume.

00:01:20 Let me say that again, because it is easy to skate past. For roughly a year, on the single most important point of leverage the United States holds over the global AI buildout — who gets the chips — the rule existed on paper and meant nothing in practice. A rule that isn't enforced isn't a weaker rule.

00:01:37 It's no rule at all. McGuire's read is that this was probably inadvertent, a clarity failure rather than a policy choice. I'll take him at his word on intent. The effect is the same either way. It gets worse, and this is the piece I think will still be live a month from now.

00:01:53 The same non-enforcement posture also touched the rules that require Taiwan Semiconductor Manufacturing Company — TSMC, the foundry that actually prints the most advanced chips in the world — to do enhanced due diligence on any order that might secretly be an AI chip headed for a Chinese customer.

00:02:10 Those due-diligence rules only bite when a license requirement is in effect. The license requirements largely were not in effect. So that check went slack too. Saturday's guidance closes the front door — Blackwell shipments to China-headquartered firms abroad are clearly illegal again.

00:02:26 It doesn't close the factory door. McGuire's point is blunt: if Chinese companies can still get chips made at TSMC through third-country cutouts, there is no point restricting their access to chips in the first place. Now, the damage assessment. The guidance itself says companies that bought chips under the loophole don't have to stop using them.

00:02:46 Read that as an admission. The shipments happened, the agency knows they happened, and it is grandfathering them in. Nobody has published a number yet for how many Blackwell systems moved during that window, and I'm not going to invent one. But the order of magnitude matters enormously, and right now it's a blank we can't fill.

00:03:05 The replies under McGuire's thread went where you'd expect — straight to retaliation. More than one pointed out that China holds its own chokepoint, rare-earth minerals, and could squeeze back. That's the texture of where we are. Export controls were supposed to be a clean, durable lever: the US sits upstream of the chip supply chain, so it sets the terms.

00:03:26 What this episode shows is that the lever only works if someone is actually pulling it, consistently, in writing. Selective non-enforcement without a clear regulation didn't tighten American control. It handed the other side a year of legal access, and called it enforcement.

00:03:41 The open question is whether the Bureau finally issues a real regulation that says, in enforceable language, what the policy actually is — and whether the TSMC due-diligence gap gets closed before more capacity walks out the door.

00:03:54 Here's a sentence from an Axios piece by Amy Harder that ran Sunday, and it's the whole story in one line: for decades, energy was an input; in the AI era, it's becoming the product. Sit with what that inverts. Electricity has spent a century being the cheap thing in the background of every business — the cost you barely thought about.

00:04:14 The AI boom has turned it into one of the most valuable strategic assets a company can hold, and the scramble to lock it up is now its own gold rush, sitting underneath the more visible one. The evidence is in the stock tape, which is where you can see capital changing its mind in real time.

00:04:31 Ford — the carmaker — launched a new subsidiary this month called Ford Energy, a roughly two-billion-dollar bet on energy storage aimed squarely at data centers and other big power users. Ford's stock hit its highest level in three years off the back of it. Bloom Energy, which makes on-site fuel cells and was long treated as a niche player, is up more than 1,200 percent over the past year.

00:04:55 Fervo Energy, a geothermal startup that Wall Street used to file under speculative climate tech, surged after going public this month. And GE Vernova — the power-equipment company — booked 2.4 billion dollars in electrical orders for data centers in the first quarter alone.

00:05:11 That's more than it made in equivalent sales in all of last year. The market is rewarding anyone who can deliver electrons to a server rack, and rewarding them at multiples that used to belong to software. But I want to be careful not to narrate this as pure boom, because the same article carries the counterweight, and it's a sharp one.

00:05:31 Brian Janous, who was Microsoft's first energy hire fifteen years ago and now co-founds a data-center developer, said this plainly: a lot of people are going to lose a lot of money in this space. Not because demand is fake — he thinks the demand is real — but because so many enormous projects are all chasing the same demand at once.

00:05:50 He pointed to a troubled project in Texas billing itself as the largest data center proposal in the world, and another one in Utah floated by the celebrity investor Kevin O'Leary. When everyone races to build the same supply for the same buyers, most of them don't get paid.

00:06:06 And then there's the friction that doesn't show up on a stock chart at all. The number of data centers canceled after local pushback hit a record high in the first quarter of this year. Those canceled projects added up to more than 40 billion dollars in investment that simply evaporated, according to Heatmap Pro.

00:06:25 Janous, who'd sounded relatively upbeat about this just a few months ago, now says the opposition is getting a lot worse. The complaints are not abstract. They're about water use, air pollution, and noise — the physical, local cost of putting a power-hungry building next to where people live.

00:06:42 This is the same fight we touched on with the grid-cost proposals from the federal energy regulator last week: the question of who pays for the power, and who has to live beside it. The capital markets have decided energy is the product. The communities being asked to host it haven't all agreed to be the factory.

00:07:01 The next year is going to be a running negotiation between those two facts, and a lot of those 40-billion-dollar cancellations are what the negotiation looks like when it fails.

00:07:11 Let me give you the most vivid version of the energy problem, because it's also a story about what AI-era money is actually pricing. Elon Musk has said SpaceX will begin launching one million orbital AI data-center satellites starting in 2028, on the still-experimental Starship rocket.

00:07:28 His claim, posted to the company's site ahead of its initial public offering: within two to three years, the lowest-cost way to make AI compute will be in space. Robert Zubrin — a veteran rocket designer who built an early prototype of NASA's heavy-lift Moon rocket, and who has known Musk for twenty-five years — read that and called it, directly, fantasy.

00:07:49 His arithmetic is worth hearing because it's the kind of arithmetic that doesn't care who's making the promise. SpaceX has already put up about 10,000 Starlink satellites, each costing roughly two million dollars to build and launch. Run the same systems to loft one million AI satellites and you're looking at around two trillion dollars — which is, give or take, the entire projected valuation of SpaceX after the IPO.

00:08:15 The whole company, spent on one constellation. Then the power math, which is the part I keep returning to. Each Starlink-class satellite generates about 20 kilowatts from its solar panels. At two million dollars a satellite, that pencils out to roughly 100,000 dollars per kilowatt of power.

00:08:32 Rooftop solar on Earth runs about 3,000 dollars per kilowatt. A gas-fired generator, around 1,000. Even a commercial nuclear plant, by Zubrin's numbers, lands between five and ten thousand. So you'd be paying somewhere between ten and a hundred times more for the electricity to run a server, in exchange for putting it somewhere far harder to cool, repair, or replace.

00:08:54 There's no broadband reason to do it either — Starlink has to be in orbit to beam internet down to people; an AI data center has no such need to leave the ground. Now, to be fair to the idea, it isn't pure science fiction forever. Google has a research effort called Project Suncatcher — a genuine moonshot paper, co-authored by nine of its researchers, on space-based machine-learning infrastructure using satellites with solar arrays, laser links between them, and Google's own tensor processing unit chips.

00:09:24 Their honest conclusion is that orbital compute only approaches parity with a ground data center if Starship launch costs fall below 200 dollars per kilogram, which they don't expect before about 2035. There's even a small demo coming — twin prototype satellites with Planet Labs, planned for early next year.

00:09:43 So the long-run physics has a path. The 2028 version, at a million satellites, doesn't. Which leaves the question of why say it now. Zubrin's answer is the one I find hardest to argue with: the timing is about the IPO. His read is that Musk is calculating that investors will look at this and shrug that no one's ever lost money betting on Elon Musk.

00:10:03 That's not a technical claim. It's a bet on reputation as a substitute for due diligence. And it's a clean test of something this whole episode keeps circling — whether AI-era capital is pricing the physics or pricing the founder. We'll find out which when the prospectus meets the market.

00:10:20 Step back from any single deal and look at where the venture money is pointed, because it's turned hard toward the physical. According to PitchBook figures reported by Kate Clark at the Wall Street Journal, global venture-capital investment into robotics and what people are now calling physical AI rose to 26 billion dollars in 2025, up from 4.2 billion in 2019.

00:10:41 That's roughly a sixfold jump in six years. And 2026 is already running hotter — more than 23 billion dollars committed as of May 20th, with seven months still to go. The thesis investors are buying is that the software-only phase of AI is maturing, and the next pool of revenue is in machines that move: warehouse robots, autonomous vehicles, humanoids, and the infrastructure to run them.

00:11:04 You can see the same rotation inside the labs. OpenAI's robotics team is openly hiring again — a small signal on its own, but it lines up with the capital. And on the open-weights side, a model called Wall-OSS released its pretrained checkpoint with a deliberate twist: it published how the base model performs on real robot tasks with zero task-specific fine-tuning, hitting more than 80 percent task progress on four of seventeen real-world tasks straight out of the box.

00:11:31 That's a modest score in absolute terms, but the framing is the interesting part — measuring the raw checkpoint instead of only showing polished results after tuning. It's an attempt to make embodied AI honest about what it can actually do before someone customizes it.

00:11:47 Now put that capital against the place it's actually landing, which is a city street. The Wall Street Journal's Sean McLain reported this weekend on what happens as robotaxi companies try to scale beyond the friendly confines of Silicon Valley. The short version: it's getting harder, not easier.

00:12:04 The pushback is coming from three directions at once — human drivers who see their living being automated, law enforcement that doesn't have a clean playbook for a car with no one to ticket, and local governments fielding the complaints. This is the gap between a demo and a deployment.

00:12:21 A robotaxi that works beautifully on a mapped, sunny grid in one metro becomes a governance problem the moment it scales into places that didn't ask for it and don't have rules ready for it. What I'd hold onto here is the asymmetry. The money is moving into bodies faster than the rulebooks for bodies are being written.

00:12:39 26 billion dollars a year flows on the logic of revenue and capability. The friction — who's liable when a driverless car blocks an ambulance, who answers when a warehouse robot hurts someone, whose job vanished — lands on cities, courts, and workers who never got a vote in the funding round.

00:12:56 I'm not knocking the robots; I'm pointing at where the bill gets sent. The capital and the consequences are arriving at different addresses, and the lag between them is where the next two years of fights will live.

00:13:09 Here's one that should bother anyone who relies on the medical literature, which is, eventually, all of us. A correspondence published in The Lancet describes a team that went looking for fabricated citations in published biomedical papers — references to studies that don't exist, that were never written, presumably hallucinated by a language model and then pasted into a real paper that got through review.

00:13:35 Over a three-year window, they found 4,046 fabricated references across 2,810 published articles. The method is worth understanding, because it's both clever and a little uncomfortable. The researchers — from Columbia University and the University of Eastern Finland — couldn't hand-check the literature; they were looking at 2.47 million papers and more than 125 million references in PubMed Central's open-access collection.

00:14:02 So they built an automated system to compare each citation against real bibliographic records, and then they used AI to sort the flagged cases — specifically Anthropic's Claude 3.5 Haiku — to separate honest mistakes from outright fabrications. If a reference couldn't be found in PubMed, Crossref, OpenAlex, or Google Scholar, it was counted as fabricated.

00:14:25 So you have AI being used to catch the mess that AI made. The battle of the machines, fought inside the footnotes. The trend line is the alarming part. In 2023, about one in 2,828 papers had at least one fabricated reference. By 2025, that was one in 458. And in just the first seven weeks of 2026, it climbed to one in 277.

00:14:46 That's a roughly twelvefold increase in a very short time, tracking almost exactly with the spread of generative AI writing tools. Some individual cases are striking: one 2025 oncology paper had 18 of its 30 citations fabricated — 60 percent of its evidence base pointing at studies that were never published.

00:15:06 A total of 246 papers carried three or more fabricated references. And the highest rates showed up in review articles — the papers that summarize a field, that other researchers and sometimes clinicians lean on to understand what's known. I talked last week about the consulting-firm version of this — fake citations turning up in a major audit report, what someone called vibe citing.

00:15:31 This is the same disease, but in the place where it does the most damage, because the citation graph is the actual trust infrastructure of science. It's how a claim earns the right to be believed: it points back to the evidence, and you can follow the chain. When the chain points at nothing — and it's pointing at nothing twelve times more often than it was two years ago — the whole structure gets harder to audit.

00:15:57 The researchers are honest that there's no clean fix, and that the underlying pressures are getting worse: shrinking research funding, a flood of for-profit journals charging thousands of dollars to publish, and fewer established scientists willing to review for free.

00:16:15 AI didn't create those pressures. It just gave overloaded, underfunded people a very fast way to cut the corner. The reckoning the authors describe is coming for scientific publishing whether or not the tools to catch the fabrications get built in time.

00:16:31 All of these stories share a missing piece — the rule, the liability, the line of accountability that hasn't been drawn yet. So I want to spend real time on someone who's thought hard about how to draw it. Brad Carson sat down with Machine Learning Street Talk for a long interview, and his résumé is unusual: former Congressman, former general counsel of the Army, former Acting Under Secretary of Defense, the man who oversaw the law of war for the entire Pentagon.

00:16:59 He now runs a group called Americans for Responsible Innovation. And his core argument is almost startlingly simple. Treat AI as a machine. Not a person. When a model produces a sentence, Carson says, the instinct — because we love language and assume it's uniquely human — is to treat that output like speech from a speaker, maybe even speech with First Amendment protection.

00:17:21 He thinks that's a dangerous mistake. In his words: it's actually a product, it's not a human being, it's a machine, and what it says to me is not covered by the First Amendment. He points at AI psychosis — people forming genuine attachments to chatbots, granting them the moral standing of a person — as evidence of how fast the anthropomorphizing runs, and how little the companies do to discourage it.

00:17:45 Once you decide a model is a speaker with rights, you've also decided you can barely regulate it. He gave a concrete example: a law forbidding a chatbot from encouraging a young person to kill themselves. If the model has First Amendment rights, that law is hard to write.

00:18:02 If it's a product, it's just product safety. From there the framework gets practical, and it leans on centuries of existing law rather than inventing something new. We already have product liability and tort doctrine that allocate responsibility when someone sells a thing that causes harm.

00:18:19 If a store sells a gun to someone it knows is dangerous, the store bears some of the liability — not all of it, the person who pulled the trigger still acted, but the seller isn't absolved. Carson's view is that the AI developers are usually the party most able to prevent foreseeable harm and most able to bear it through insurance, so they should carry most of the burden.

00:18:41 His sharpest example: models like Stable Diffusion were trained on datasets that contained child sexual abuse material. The labs have the tools to screen that out. If they don't, he argues, they should be liable for the downstream harm. That's not a radical new regime.

00:18:57 It's the same logic we apply to a pesticide or a bottle of spray paint. The part I found most useful was how he handles the regulatory-capture objection — the standard Silicon Valley argument that any agency will just get captured by industry, so why bother. Carson doesn't deny capture is real.

00:19:15 He flips the comparison. Right now, he says, there is no regulation at all, and into that vacuum the most informal kind of capture has already moved: a handful of venture firms and Valley figures shaping AI policy through moneyed networks and personal influence, accountable to no one.

00:19:32 His line on the people who cry regulatory capture is that the argument is never falsifiable — it's searching for a pea under a hundred mattresses. He'd rather have something like the model that governs financial audits: testing and verification done by the private sector, but overseen by a public body the way the Securities and Exchange Commission oversees company accounting, so you don't end up with another Enron.

00:19:57 An agency you can see and hold to account, he argues, beats an informal network you can't — even granting that the agency will be imperfect. And he ties it back to consumer protection in a way that should sound familiar. He brought up Anthropic abruptly changing Claude's token allocation and model routing — paying customers waking up to a different product than the one they bought, with no clear account of what changed.

00:20:22 That, he says, is a transparency and consumer-protection gap. When you've built something with what he calls epochal consequence, you're not the corner hardware store anymore. You carry a public responsibility to say what you do, what you trained on, what the thing is capable of — and to tell people before you change it.

00:20:42 There was one more detail from the interview I can't shake, because it's where this stops being abstract. He described an AI targeting system assigning a person in Gaza a 0.73 percent probability of being a Hamas member, and asked the obvious questions: what's the threshold?

00:20:58 Do you get struck at that number, or cleared? Who set the line? That's the machine deciding you're a target. And it's exactly the kind of decision Carson wants pulled back under human law — not because the math is wrong, but because no one elected the math.

00:21:13 I'll close on a problem that gets less attention but connects all of this. SemiAnalysis put out a piece arguing that AI's output is becoming one of the hardest economic-measurement problems in history. They call it dark output — the economic value AI is generating that simply doesn't show up in national statistics.

00:21:30 Think about what we counted today. Chips moving across borders that the controlling agency couldn't fully see. Energy being repriced as a product faster than anyone's models for it. Robots funded years ahead of the rules that will govern them. Citations pointing at studies that were never written.

00:21:45 And now the suggestion that the very instruments governments use to steer the economy — productivity, output, the numbers behind interest rates and budgets — can't actually see a growing share of what AI is producing. That's the thread I'd leave you with. The capability isn't the bottleneck anymore.

00:22:01 The thing lagging, over and over, is our ability to see, count, and assign responsibility for what the capability is already doing. The chips already shipped. The energy money is already placed. The robots are already funded, the fake citations already published, the targeting score already computed.

00:22:16 In every case the action ran ahead of the ledger, the rulebook, and the enforcement. Take all of this as an argument for building the instruments that can keep up with the technology — the regulation that's actually written down, the liability that's actually assigned, and the measurement that actually measures.

00:22:32 Because a rule no one enforces, a number no one can audit, and a harm no one is liable for all amount to the same thing. They mean the leverage you thought you had isn't really there. Two things stay open for me: whether the Bureau turns Saturday's guidance into a real, enforceable regulation, and how big the number turns out to be for what slipped through while no one was pulling the lever.

00:22:53 Jonas.