The Kill Switch

00:00:04 On Wednesday morning, the European Commission published a set of proposals it calls a technological sovereignty package. The phrase everyone carried away from it was two words: kill switch. And the worry behind those words, stated plainly by Henna Virkkunen, the Commission's vice-president for tech sovereignty, is that a foreign government or a foreign company could one day reach into Europe's cloud computing, its AI systems, or its chip supply, and turn something off.

00:00:31 Not as a figure of speech. As an actual switch that somebody outside Europe gets to flip. The fear has a recent precedent, and the Commission named it directly. Last year, China stopped semiconductor exports and nearly brought the European car industry to a standstill.

00:00:46 That happened. On the other side of the ledger sits the United States, and one law in particular — the 2018 Cloud Act, which lets American federal authorities reach data held by US providers even when that data is sitting on European soil. Virkkunen said that law, quote, was not in line with our rules here.

00:01:04 And she said Europe has to be sure that nobody holds, in her phrase, the kill switch possibility. Then comes the number that frames the entire effort. The Commission says Europe relies on foreign suppliers for more than 80 percent of its digital products and services, its infrastructure, and its intellectual property.

00:01:22 Europe produces roughly 10 percent of the world's semiconductors, and for the most advanced chips — the ones that train and serve frontier models — it is almost entirely dependent on the United States and East Asia. That is the hole this package is trying to climb out of.

00:01:38 So what's actually in it. There's a Cloud and AI Development Act, which would require member states to run a risk assessment on the providers handling sensitive work — defense, criminal justice, and border management. If a provider is judged risky, authorities could be forced to switch to someone else.

00:01:55 The draft suggests US cloud companies operating in Europe might have to prove they can't be compelled to hand EU data to American authorities. There's a Chips Act 2.0 that would let the EU make direct investments in chip production. And there's a plan to triple Europe's data center capacity within five to seven years, with what the Commission calls acceleration zones — fast-tracked permitting for the power-hungry buildings — plus an energy-efficiency rating system bolted on.

00:02:23 The industry reaction was immediate and unhappy. The Computer and Communications Industry Association, whose members include Amazon and Google, called the proposal a, quote, dangerous recipe for progressive market shutdown — meaning, in their telling, that trusted providers get pushed out of slices of the European market.

00:02:41 And the skeptics aren't only the incumbents. Olivier Darmouni, an economist at HEC Paris, doubts Europe can build a leading-edge chip fab on anything like the timescale it would need to catch the US. He thinks the more realistic win is memory chips, where there's a global shortage right now.

00:02:58 He also flagged the part the proposal sidesteps: tripling data center capacity while hitting Europe's climate targets. His line was that very soon it will be in Europe as it is in the US, where people worry the data centers are going to blow up their electricity bills and drag the climate trajectory down with them.

00:03:16 Here's my read. This package is a bet that depending on someone else for your compute is now a national-security exposure, not just a procurement choice. I think that instinct is correct — the China chip freeze and the Cloud Act are both real levers, and Europe got caught flat-footed once already.

00:03:33 What I'm less sure about is whether you can legislate your way out of an 80-percent dependency in half a decade. You can write the risk assessments today. The fab and the grid take years, and the proposals still have to clear member states and the European Parliament.

00:03:49 The gap between passing the law and pouring the concrete is exactly where this either becomes real sovereignty or stays a press release with a good headline.

00:03:58 Yesterday I spent real time on the President's AI executive order — the voluntary thirty-day model-submission idea, with the national-security benchmarking pointed toward the intelligence side of the government. The question I left open was whether that voluntary frame would harden into something binding, and whether any major lab would actually balk at it.

00:04:19 Today, one did, and it's the biggest one. OpenAI published a policy paper that breaks with the executive order on two specific points. First, it wants cyber-risk evaluations of advanced AI systems to be mandatory, not voluntary. Second, it wants those evaluations led by CAISI — that's the Center for AI Standards and Innovation, a civilian body inside the Commerce Department — rather than the National Security Agency, which is where the administration's order leans.

00:04:48 OpenAI is packaging this as, in its own words, a blueprint for democratic governance of frontier AI. Greg Brockman and a policy lead named Jason Wolfe were out promoting it, Wolfe pointing to an expanded role for CAISI and a more careful stance on federal preemption.

00:05:04 And the timing wasn't subtle: Sam Altman was in Washington on Wednesday for a run of meetings with White House officials. The paper came out as he walked the halls. Stop and sit with how unusual this is. A frontier AI company is asking the government to regulate it harder, and asking that a civilian standards office, not the country's largest spy agency, hold the evaluation pen.

00:05:27 Companies don't normally lobby for mandatory rules. So why would this one? I can think of a few explanations, and I won't pretend to know which one dominates. One is competitive: mandatory evaluations are a cost, and costs are easier for the largest, best-funded lab to absorb than for a startup, so a rule that looks like safety can also function as a moat.

00:05:49 A second is procedural preference — a civilian process run by a standards body is more transparent and more predictable than a classified one run by the NSA, and predictability is worth money when you're trying to ship. And the third is the straightforward one: some people inside OpenAI think cyber-risk review should be binding and visible rather than optional and secret.

00:06:11 These aren't mutually exclusive. They can all be true at once. Now set the other half of the day beside it. At the Axios summit in New York, also on Wednesday, IBM's chief executive Arvind Krishna came out and backed the administration's narrowed executive order — the lighter, voluntary version.

00:06:29 So in a single day you had a frontier lab pushing for more binding rules under civilian control, and a blue-chip enterprise vendor endorsing the looser federal touch. That's not the industry speaking with one voice. That's two different business models reading the same order and wanting opposite things from it.

00:06:48 My take: watch who benefits from each posture. OpenAI sells the frontier and would rather the rules of the frontier be written in a forum it can engage as a peer. IBM sells into regulated enterprises that mostly want Washington to stay out of the way. Neither is wrong about its own interest.

00:07:06 But the through-line from yesterday holds — the government asked politely for a look, and now the people being looked at are negotiating, in public, over who gets to do the looking and whether they can say no. That negotiation is the real governance story, and it's happening in policy papers and summit panels rather than in any statute.

00:07:26 Here's a sentence that should puncture some of the AI-buildout euphoria. JP Morgan estimates that more than 60 percent of the US data center capacity planned for completion in 2027 isn't yet under construction. Not behind on furniture and fit-out. Not yet broken ground.

00:07:41 That's reporting from Katherine Blunt at the Wall Street Journal, and it lines up with what the grid analysts have been saying for months: the binding constraint on the AI buildout isn't chips and it isn't money. It's power — the interconnection queues, the transformers and substations, and the electrons that have to actually show up.

00:08:01 You can sign the lease and order the servers, and still wait years for the utility to energize the site. I've been circling this point all week, because it keeps showing up from different directions. Over the weekend the story was an export-control loophole and the question of how many chips slipped through.

00:08:18 Earlier in the week it was local communities canceling data center projects in their own backyards. Now it's the construction timeline itself slipping. The common element isn't a conspiracy or a single villain. It's that the AI buildout has run into the physical economy — the permits, the power lines, the neighbors, and the transformers with eighteen-month lead times — and the physical economy moves at its own pace no matter what the term sheet says.

00:08:44 The Journal frames Google as the company with a strategy to get around the bottleneck, and Google is the bridge to the second number I want to put in front of you. Alphabet is raising roughly 85 billion dollars in equity, and the whole thing was sold as a way to pay for AI infrastructure and compute.

00:09:01 But Cory Weinberg at The Information reports that nearly 40 percent of that raise — call it 30-some billion dollars — is going to cover tax obligations tied to employee equity awards. Not graphics chips. Not power purchase agreements. Taxes on the stock the company hands its people.

00:09:17 Think about what that reframes. When a company tells the market it's raising 85 billion for AI infrastructure, you picture concrete and silicon. And a real chunk of it is the cost of keeping engineers from walking across the street to a competitor who'll pay them more.

00:09:32 The talent war and the compute war are drawing on the same pool of capital, and they're not the same fight. One buys you a building. The other buys you the people who know what to put in it. I find this clarifying, because the dominant story for a year has been capital — the eighty-billion-dollar raises, the trillion-dollar infrastructure projections, the sense that whoever can write the biggest check wins.

00:09:55 And both of today's numbers cut against that. The data center figure says capital doesn't equal capacity, because capacity runs through a power grid that doesn't care how much you raised. The Alphabet figure says the headline number isn't even all going where the headline says.

00:10:11 Forty cents of the AI-infrastructure dollar is paying a payroll tax bill in disguise. None of this means the buildout is fake. Demand for compute is enormous, and the spend is too. But there's a difference between announced capacity and energized capacity, and the gap between them is filling up with caveats — power not yet contracted, sites not yet started, and raises partly spoken for before the first server arrives.

00:10:34 If you want to know how fast AI actually scales over the next two years, I'd stop counting dollars raised and start counting megawatts interconnected. That's the number the grid will enforce whether the market likes it or not. And it's the one that's been slipping while everyone watched the IPO filings.

00:10:52 Around ten in the morning on Wednesday, Tesla's unsupervised robotaxi — no safety driver in the seat — started running across the entire Austin metro geofence. Ashok Elluswamy, who runs Tesla's autonomy and AI work, posted it in a single line: unsupervised robotaxi now available in the greater Austin area.

00:11:10 Elon Musk reposted it, the company's robotaxi account confirmed the whole metro, and the original tweet cleared 670 thousand views by the evening. The same day, Tesla was at the big computer-vision conference, CVPR, presenting how it's building foundation models for robotics — the architecture, the large-scale multimodal training, end-to-end control, and the part that matters here, safety and deployment.

00:11:34 Let me separate what we know from what we're being asked to take on faith. What we know: Tesla expanded a driverless service from a smaller supervised pilot to a wider area without a human minder in the car, in a major American city, in daylight, today. That's a real step.

00:11:51 Driverless operation across a whole metro is physical-world deployment at city scale, and it puts Tesla into the territory Waymo has been grinding through for years — except Tesla's pitch has always been that its camera-and-neural-net approach generalizes faster and cheaper than Waymo's sensor-heavy maps.

00:12:09 What we're asked to take on faith is the safety record, because we don't have one yet. A launch tweet and a 670-thousand-view repost aren't an incident report. We don't have miles-per-intervention data, we don't have a disengagement disclosure, and we don't have an independent regulator's readout of how the first unsupervised days in the full geofence actually went.

00:12:31 Geofenced, daylight, good-weather operation is the easy part of the problem. The hard part is the long tail — the construction zone, the emergency vehicle, and the pedestrian doing something no training set anticipated. I'm not saying it'll fail. I'm saying the announcement and the verification are two different things, and only one of them happened today.

00:12:53 The consequence I keep coming back to is labor and liability, because those are the parts that touch people who don't follow autonomy launches. Austin has ride-hail drivers. A driverless fleet across the whole metro is, eventually, a direct substitution for some of that work — not all at once, but on a curve.

00:13:11 And the liability question gets sharper the moment there's no human in the seat to blame. When a supervised car crashes, the safety driver is the legal shock absorber. Take the driver out, and the crash points straight back at the company, the software, and whatever the regulator decided to require or not require before letting it run.

00:13:31 Texas has been permissive on this. That permissiveness is now being tested at metro scale, in public, with real passengers. So I'd hold both thoughts. This is further than a lot of skeptics expected Tesla to be by now, and it deserves to be counted as progress.

00:13:47 And the meaningful number — the safety data that tells you whether unsupervised at metro scale is actually safe — is exactly the number we don't have on the day of the announcement. I'd want to see ninety days of it before I'd call this settled.

00:14:02 Morgan Stanley is about to let AI agents walk in through a door that, until now, only humans were allowed to use. CNBC's Hugh Son reported on Wednesday that the bank will soon let its corporate clients' autonomous agents pull data and insights directly from its stock-administration platforms — ShareWorks and Equity Edge — bypassing the software interfaces built for people.

00:14:23 The chief product officer of Morgan Stanley at Work, Mark Mitchell, put it about as bluntly as a bank executive ever does. Quote: in a future state, our corporate clients will not be logging into ShareWorks or Equity Edge. Instead, he said, they'll use agentic AI tools on their own desktops, interacting with the platforms in a purely agentic way.

00:14:43 A little context on why this funnel matters, because it's not a side business. Morgan Stanley turned the dull work of administering employee stock plans into a feeder for its wealth-management division — and that division is the largest in the world, at 7.35 trillion dollars in client assets.

00:15:00 Run the stock plan, and you can convert employees into advisory clients as their wealth grows. The firm says it touches almost half the companies in the S&P 500 and eight of the ten biggest unicorn startups. So this isn't a peripheral experiment. It's the bank wiring AI agents into the on-ramp of its richest business.

00:15:19 A handful of clients already have early access, and Mitchell says all 3,400 administration clients should get it by next year. Underneath it is the Model Context Protocol — the open standard that lets AI models plug into data sources — which is becoming the way agents reach into systems that were never built for them.

00:15:37 Two things here are worth more than the announcement itself. The first is labor. Mitchell was explicit that the appeal to fast-growing tech and biotech clients is administering ever more complex stock plans without adding headcount in support roles like human resources.

00:15:53 And internally, he said the same thing — agentic AI lets Morgan Stanley scale customer support, plan administration, and the wealth funnel without adding, his words, thousands and thousands of employees. That's not a hypothetical about future job loss. That's a chief product officer naming the headcount that doesn't get hired as the product's selling point.

00:16:13 The second is the moat. For decades, the entire game was locking users into your proprietary interface — make them log into your front door, keep them there, own the relationship. Morgan Stanley is now betting that the front door barely matters. Mitchell said the companies that survive are the ones with proprietary data and business logic, and that clients not logging into the website, quote, doesn't scare us at all.

00:16:37 I think that's a striking concession from a firm that's spent decades building those interfaces. If the agent becomes the user, the value moves from the screen people looked at to the data and the logic sitting behind it. That repricing is going to ripple well past one bank — every enterprise software company that sold a slick interface is about to find out whether anyone was paying for the interface or for what it was sitting on top of.

00:17:03 While Wall Street was opening its doors to agents, a British regulator was forcing a different door shut. The UK's Competition and Markets Authority ruled on Wednesday that online publishers can now block their content from appearing in Google's AI summaries in UK search results — and, crucially, do it without dropping out of regular Google search.

00:17:23 That caveat is the whole story, so let me make sure it lands. Until now, if a news organization didn't want its articles scraped and condensed into one of Google's AI overviews — the summary that sits at the top of the page so people read it and never click through — its only option was to withdraw from Google search entirely.

00:17:42 Given that Google handles more than 90 percent of general searches in the UK, that's not a real choice. It's a threat dressed as an option. You either let the AI eat your work and your traffic, or you make your journalism invisible. The CMA just severed those two things.

00:17:57 You can now stay in search and say no to the summary. Sarah Cardell, the watchdog's chief executive, called it a, quote, world-first requirement on Google's search services, and said it was crucial that publishers have appropriate bargaining power over how their content is used.

00:18:13 The News Media Association, which represents UK publishers including the Guardian, called it a step toward an economy where premium content is properly respected and fairly compensated. Google, for its part, said it would start testing the control on a subset of UK media sites that same day, with the stated aim of rolling it out globally.

00:18:32 And the CMA tacked on a second requirement: Google has to properly attribute publisher content with clear links inside its AI results. Why this matters beyond Britain. The economic damage publishers have complained about is concrete — AI summaries answer the question on the results page, the click-through never happens, and the ad revenue that funds the reporting dries up.

00:18:53 Up to now, the AI search products held all the leverage, because the alternative to participating was disappearing. A regulator just rebalanced that, and did it in a way Google says it'll export to other markets. Set this next to the OpenAI copyright litigation grinding through a New York federal court, where the same underlying question — can you train on and reproduce other people's work without permission or payment — is being fought as a matter of law rather than competition policy.

00:19:21 Two different legal machines, the same fight: who controls the value of the content that AI systems are built on top of, and who gets paid when a machine reads it back to someone. My read: this is the first binding lever I've seen that gives content owners a genuine opt-out from an AI product without a suicidal cost attached.

00:19:40 It won't restore the old traffic economics — those are probably gone. But it changes the negotiation from begging to bargaining, and that's a real shift in who holds power in the deal.

00:19:50 I'll end on the hardest one, because it moved from theory toward procurement today. David Omand — the former head of GCHQ, the UK's signals-intelligence agency — told the Guardian he has changed his mind about autonomous weapons. More than a decade ago, he chaired a commission that doubted whether autonomous drones could ever distinguish civilians from combatants well enough to comply with international humanitarian law.

00:20:15 Now he argues the opposite: that AI can give a drone what he calls an adaptive moral control layer, a set of parameters a human sets before a mission, which the machine then applies in the heat of combat. The key phrase in his argument is a shift from in the loop to on the loop.

00:20:31 In the loop means a human authorizes each individual strike. On the loop means a human supervises the system and sets its rules, but doesn't sign off on every shot, because — in Omand's framing — warfare has gotten too fast for that. His words: it's a physical and operational inevitability.

00:20:48 With drones, hypersonics, and the speed of modern combat, he says, you simply won't have time for a human to make each call, and a well-designed moral framework might actually produce less collateral damage than a stressed human deciding in seconds. He even argues the machine's reasoning could end up ethically superior to the human's.

00:21:08 This isn't an abstract debate, and that's what makes today's version of it land differently. The US 2027 budget, released in April, sets aside 54 billion dollars for autonomous and remotely operated systems across air, land, and sea, including a program literally called drone dominance.

00:21:25 The UK armed forces minister, Al Carns, said there might be circumstances where machines make their own targeting decisions — you must have the ability, he said, to take the human out of the loop when required. And the commercial AI industry is already inside the kill chain: technology from Palantir and Anthropic has been deployed to shorten the process from selecting a target to launching a strike in the Iran conflict.

00:21:49 The companies whose models you might use to summarize a document are, in another configuration, compressing the time between identifying a human being and killing them. The counter-argument was just as direct, and I want to give it room, because I think it's the stronger one on the merits.

00:22:06 Chris Cole, who runs Drone Wars UK, called Omand's position as nonsensical as it is dangerous. His point: AI merely processes data. It doesn't make judgments. It lacks the ability to distinguish a civilian from a combatant, or to weigh whether a loss of life is proportionate to a military advantage — and proportionality is the exact legal test Omand says the machine can be programmed to pass.

00:22:29 And it's worth noting that even David Petraeus, the former CIA director, has said the US has no military doctrine for autonomous formations. So you have the budget line, the cabinet minister, and the spy chief all leaning toward on the loop, while the legal and ethical machinery meant to govern it doesn't exist yet.

00:22:48 Here's where I land. The thing that's actually decided isn't whether autonomous targeting is wise. It's that the money has already been appropriated, the commercial tooling is already in the field, and a respected former intelligence chief just gave the policy its moral language.

00:23:04 The case against — that a system processing data cannot exercise human judgment about proportionality — is, to me, the more accurate description of where the technology is. But accuracy rarely beats 54 billion dollars and an inevitability argument. Step back across the whole day and you see the same question asked six different ways: in Brussels, in Washington, on a Texas street, inside a bank, in a British regulator's ruling, and over a drone's targeting decision.

00:23:31 Who holds the off button, and can the people it gets pointed at say no? The deadline I've got circled is sixty days out, when the administration's classified AI benchmark and framework come due — the same process I tracked yesterday — because that's where the question of who evaluates these systems, and against what standard, stops being a panel discussion and starts being a rule.

00:23:54 I'll tell you which way it broke. Jonas.