From left, Russian President Vladimir Putin, Chinese President Xi Jinping, and President Donald Trump take part in a trilateral meeting in Shenzhen, China. | Mandel Ngan/AFP via Getty Images
This story appeared in Today, Explained, a daily newsletter that helps you understand the most compelling news and stories of the day. Subscribe here.
President Donald Trump is on his way back from China after a four-day Asia trip that included taking in a surreal cover of “Y.M.C.A.,” heaping praise on a noted kleptocrat, and making earnest — if slightly cringe! — declarations of star-crossed presidential friendship.
But the highlight of the trip was surely Trump’s unprecedented trilateral meeting with Chinese President Xi Jinping and Russian President Vladimir Putin, the first between those countries’ leaders.
The three men sat around a comically large and reflective roundtable to discuss issues like reforming the United Nations, increasing the world’s supply of petroleum, and reducing the risk of nuclear war, which they all thankfully agree would be a bad idea.
A major nuclear treaty between the US and Russia lapsed nine months ago, and China has been a sticking point to negotiating a new one. In fact, all three countries are — in one realm or another — major obstacles to each other’s geopolitical ambitions.
Accordingly, no breakthroughs were reached during their hour-long chat. But they did agree to keep chatting…including about nuclear proliferation.
How, exactly, did we get here? To find out, I reached out to Rose Gottemoeller, the chief negotiator of the New Strategic Arms Reduction Treaty, or New START — that US-Russia nuclear deal that just recently expired.
The original START was signed in 1991, just before the collapse of the Soviet Union. For the first time, the two countries agreed not just to stop growing their nuclear arsenals, but to actively reduce their strategic weapons.
New START slashed their stockpiles even further: to 700 missiles, 800 launchers, and 1,550 deployed warheads, Gottemoeller said. (When you consider that the world had roughly 70,000 nuclear warheads at the peak of the Cold War, it’s a pretty massive accomplishment.)
But New START was negotiated during a brief, shining, Obama-era moment of optimism for nuclear arms control. Then Russia started getting grabby in Ukraine, and that optimism curdled.
The treaty was extended once, in 2021, about a year before Putin began a full-scale invasion. Since then, neither country has been in a particularly diplomatic mood. On the contrary, both Trump and Putin have at times signaled they might be a little nuke-curious.
The good news, of course, is that neither president ever acted on that curiosity. And in fact, the lapse of New START earlier this year hasn’t triggered a buildup. “So far Russia and the United States are sticking with the status quo limits,” Gottemoeller told Vox.
The bad news is that either country could reverse course, and there are complications to negotiating firmer limits in the future. One of the biggies is China, which is rapidly building up its own nuclear stockpile — and thus disincentivizing Russia and the US from cutting theirs back further.
That’s the context for this morning’s trilateral, where Trump spoke explicitly about nuclear arms control before heading into the closed-door meeting. Gottemoeller, for her part, said she was “hopeful” the summit would “be a catalyst for new talks with the Russians [and] also nudge open the door to first talks with the Chinese.”
We’ll have to wait and see about all that. But the three leaders did release a vague post-summit statement that reaffirmed “a nuclear war cannot be won and must never be fought.” That, crucially, is a callback to a statement that US and Russian leaders released in 1985 after a summit that reinvigorated nuclear arms-control talks.
One link for later
➨ Talk to more strangers. That occasionally awkward but ultimately life-affirming small talk helps you feel connected to other people. And according to University of Kansas researcher Jeffrey Hall, it boosts your entire community, too. By offering “dignity to another person…and by stretching yourself outside just the people you know well, you’re actually building a healthier environment for all of us,” Hall says.
Before you go…
Did you know…that almost 15 percent of adults suffer from tinnitus — a persistent, often ringing sound in their heads? It was once believed to be psychosomatic, but new research suggests it’s the result of a very specific type of hearing damage.
Today’s trivia: Which popular therapy method was influenced by Buddhism? (You can find this and other brain puzzles in Vox’s daily crossword. Look for the answer in tomorrow’s edition.)
Yesterday’s trivia: Yesterday we asked you for the name of the US soccer player who scored the game-winning penalty kick at the 1999 Women’s World Cup. That would be Brandi Chastain, who famously — iconically! — fell to her knees and tore her jersey off in celebration.
BRUSSELS — Andy Burnham has gone big on Britain’s relationship with the EU during the first months of his premiership.
By the weekend, he may actually have something to show for it.
The British prime minister is heading to Brussels on Friday for his first U.K.-EU summit. But on the eve of the meeting, what exactly he’ll be coming back with is still up in the air.
POLITICO analyzed leaked draft communiqués and spoke to diplomats and officials on both sides of the talks to piece together how the agreements are shaping up — and what’s still to be decided.
Here’s what you need to know.
Why is this happening?
The U.K. and EU have committed to holding annual summits to build on the relationship they were left with after Brexit.
This is the second such meeting, and is primarily about concluding negotiations on agreements announced at the first gathering back in May 2025.
That summit was attended by then-British Prime Minister Keir Starmer. This week’s meeting was originally due to be held in July, but it was postponed after Starmer stepped down — to give Burnham time to put his stamp on the process.
What’s on the table?
Three new international agreements are set to be announced at the meeting.
The first is a sanitary and phytosanitary deal that would align British food and drink standards with the EU’s. The aim of the agreement is to remove bureaucracy and to eliminate the need for border checks on food exports, smoothing trade.
The two sides have also negotiated a deal to link their emissions trading systems. The main benefit would be to exempt U.K. and EU businesses from having to pay new carbon border taxes levied by both sides. It would also provide more liquidity in carbon trading markets.
It still wasn’t clear on Thursday night when the deal will take effect, though. The longer it takes to introduce, the longer firms would be caught between the two competing tax regimes.
Brussels and London also want to establish a youth experience scheme, which would provide low-cost visas for young people to live and work abroad for a fixed period of time. The U.K. has insisted the system should be subject to a cap on numbers — expected to be around 100,000 after fraught negotiations.
Complications implementing the scheme in some member countries means it isn’t set to take full effect until 2030. This would also mean its full impact on politically-sensitive U.K. migration data won’t register until after the next election, which must be held no later than 2029.
In addition to these three agreements, the two sides will also publish a “joint statement” outlining their plans for other areas of cooperation. This isn’t a binding international agreement, but a statement of intent, covering areas ranging from migration and defense to cultural exchange, a digital partnership and a health security pact.
On tech, British firms are concerned that they could be shut out of parts of the EU public administration market because of the proposed Cloud and AI Development Act (CADA) and its new sovereignty requirements. They’ve lobbied hard for reassuring language in talks, both with the negotiators working on the law and with those drafting the joint statement.
On health, the U.K. and EU are looking to strike a deal to tackle medicine shortages by sharing intelligence on supply disruptions and stockpiles. As Brussels steps up efforts to manage medicine shortages across the bloc, Britain risks being left out of the increasingly joined up system. Closer cooperation could help both sides “avoid procurement battles” for limited stocks and “tit-for-tat export bans, as happened during the pandemic,” said Mark Dayan, Brexit lead at the Nuffield Trust.
One area the statement will definitely cover is the question of tuition fees. EU countries want European students studying at British universities to qualify for lower domestic fees, as they did before Brexit.
The U.K. is open to the idea but wants, in turn, assurances its businesses won’t be hit by new EU Made in Europe rules. Brussels says it can’t give those assurances yet, so the statement is likely to include a compromise to delay talks by up to six months. Only then will we find out how much lower tuition fees might get.
Another key question is whether the implementation of the agri-food, emissions trading and youth mobility deals will be contingent on a successful outcome for tuition fee talks — which some member countries have been pushing for. This will be determined by the wording of the final statement, but the latest version does not make this link explicit.
What does Burnham want to talk about?
Since Labour Party conference in September, Burnham has talked up building a closer relationship with the EU — ditching some of the caution that characterized his predecessor’s approach to European relations.
He may try to use the meeting as an opportunity to sound out EU leaders on where things could go next.
But grumblings from Burnham’s own Cabinet about his policy shift on Europe could mean he keeps a lot of that under his hat. He’s expected to set out options for a new look-relationship in mid-December.
In the meantime, the government wants a political win from the deals to be agreed on Friday. In the run-up to the meeting, European Relations Minister Hamish Falconer has been making the argument that the SPS deal will help bring down the cost of living by controlling food inflation. Expect more of that.
There are also topics that Burnham won’t want Brussels to forget — chiefly shielding British industry from Made in Europe protectionism. The EU wants the joint statement to “bind” the U.K. to implementing EU tariffs against Chinese vehicles. But London hasn’t made up its mind and merely wants to commit to guarding against the risk of “circumvention.”
What does the EU want to talk about?
Equally, there are topics Brussels won’t want London to forget about. EU member countries are very keen on that tuition fees cut, for example.
But for the EU, building a closer relationship with Britain is mostly about geostrategic aims.
With an uncertain Atlantic alliance and war raging in Europe’s east, the EU wants to build new defense, industrial and security relationships — and the U.K. is the most obvious partner.
How does the day unfold?
The summit will kick off at 10 a.m. Brussels time, so 9 a.m. London time.
How long it goes on for will depend on just how done those deals are by the time Burnham steps off the Eurostar.
If things go smoothly, officials on both sides are estimating a joint press conference to announce the package and take questions around lunchtime.
That’ll be attended by Burnham, European Commission President Ursula von der Leyen, and European Council President António Costa.
But if there’s no sign of a white smoke by around 2 p.m., things may not be going entirely to plan.
BRUSSELS — Long before the fresh spate of hacks and website shutdowns, the inherent instability of the web was depicted in an internet-famous cartoon of a rickety tower propped up by a single small block.
Now a near-daily rhythm of outages suggests the supporting block may have come loose.
POLITICO revealed Thursday that internal services at EU institutions had gone offline after a major AI-fueled cyberattack. That came after service interruptions at U.S. national security offices and the federal aviation authority, global outages at Microsoft and a succession of internet-activated attacks on supermarket refrigerators in Europe.
Authorities blame the sudden wave of instability on the capacity of new AI tools to detect vulnerabilities in existing code and for bad actors to exploit them.
Interviews with more than a dozen cybersecurity leaders and those working on maintaining the internet show the scale of the challenge to keep the web up and running, in a world where the pace of finding vulnerabilities has rapidly and permanently accelerated.
Since the web’s early days, an unruly cast of idealist engineers and non-profit foundations have been essential for looking after decades-old critical code that’s been used to build our digital world. This small army of largely volunteer coders now say they can’t keep up with the power and pace of AI.
“We’re running on fumes,” said Robin Bender Ginn, executive director of the OpenJS Foundation, which looks after JavaScript, among the most widely used programming languages. “It keeps me up at night.”
Cybersecurity agencies around the world rang the alarm last month about major flaws detected in PHP, another critical coding language that underpins 70% of the web. This week, similar scares have spread over Java and NGINX, which also fuel large swathes of technology.
Helmut Reisinger, the European chief executive of cybersecurity giant Palo Alto Networks, said in an interview that the company is seeing “more or less” ten times as many incidents compared to about four months ago. He predicts the constant drumbeat of attacks and scares in the last month is going to become the “new norm.”
The PHP Foundation, which was only established in 2021 despite the coding language itself existing since the early 1990s, has two members of staff and 13 contractors. “Having very few individuals responsible for some key piece of infrastructure is just problematic,” said Nils Adermann, a PHP Foundation board member.
Out of office
Many tools like PHP have their origins in the 1990s, when foundational software and programming languages were developed in the messy, community-minded early days of the web. Those that worked well or were useful were gifted as a common good, picked up by other software developers and became building blocks for a vast number of digital technologies on which governments, critical infrastructure and citizens now rely.
Many open-source software libraries form the basis for widely used software like Android. Facebook was originally built using PHP, and household names like Spotify, Etsy and Wikipedia are just some top destinations that still use it as the basis for their services and websites.
The homepage of the PHP website is seen in Singapore on Oct. 22, 2026. | Yong Teck Lim/Getty Images
When cybersecurity agencies and companies issued an urgent warning in October that there were major flaws in PHP identified with the help of AI, and that they should either update their systems or shut them down, it was a public revelation that the architecture upon which our digital world relies is almost comically fragile.
Adermann explained that it is common for people who have worked on a PHP project for years in their spare time — developing a level of specific expertise that’s very hard to replicate — to duck out when they start families or can no longer find spare time.
That became especially apparent during the process of fixing the recent PHP flaws: One person involved in pushing out technical updates had a scheduled medical procedure and “was simply not available,” he said. “It was like, ‘okay, how do we figure it out, who can even do this now?’”
“That was one of these things where it became clear that we have a pretty limited budget and a lot of this is just barely working,” he said.
The discovery of the flaws did not, in the end, result in a doomsday scenario, in part because many organizations took themselves offline or fixed their systems, and also because the PHP flaws turned out to be harder than expected to exploit.
Free riders
Along with the PHP Foundation, numerous other non-profit foundations exist to act as guardians of open-source software, including the Linux Foundation, Python Software Foundation and the OpenJS Foundation.
They are typically funded through a mix of government grants and donations from technology companies. Funding can be used to pay full-time staff to carry out security audits, rather than relying on volunteers who, though committed, may suddenly become unavailable.
Tech firms have been among the biggest donors because they benefit from the open source software on which their own offerings are based, and continue to contribute to many projects.
Some of the world’s largest companies, including Anthropic, OpenAI, Google and Microsoft, have clubbed together to back a Linux Foundation project called Alpha Omega, which provided funding that ultimately led to the PHP flaws being fully identified and fixed. The project has been a “tremendous help,” said Adermann.
Another key donor often cited by open-source figures is the German government-backed Sovereign Tech Agency, which invested €205,000 in PHP in 2023 to 2024, with another €223,680 committed for 2025 to 2026, Powen Shiah, a spokesperson for the agency, said in emailed comments.
Germany is widely seen as a leader in open source, and the agency considers the funds it hands out to be investments in “chronically underfunded” critical infrastructure rather than donations.
“Improving a single foundational component can strengthen thousands of downstream applications and businesses, delivering benefits across entire industries,” Shiah said.
Juhan Lepassaar speaks at an event in Kyiv on Feb. 7, 2024. | STR/NurPhoto via Getty Images
Yet even between government and private funds, the sums remain small, particularly when dispersed among so many organizations. Tech firms can also be capricious, cutting funding when they lose interest in a particular project.
“More organizations need to understand how fragile this infrastructure can be and see investment in it as part of their own security,” Shiah said. “Some companies already invest substantially, but the scale of their financial contributions usually doesn’t reflect the value they derive.”
Only a “tiny fraction” of the companies that depend on PHP contribute financially, Adermann said. By his calculations, companies make hundreds of billions each year using PHP but contribute nothing.
Ginn said one Big Tech company dropped out of her OpenJS Foundation’s top membership tier just recently, leaving only Microsoft remaining. She noted that an issue in recent months is that more cash is being funneled toward AI-focused open-source foundations.
The AI deluge
The rise of AI models and their capabilities of finding software flaws is putting more pressure on open-source software engineers, who are inundated with reports of vulnerabilities — some of which are false or exaggerated.
Reports of flaws discovered by AI are “quite verbose,” Ginn said.
Juhan Lepassaar, the head of EU cyber agency ENISA, said in an interview that the agency has observed a similar pattern of false positives.
ENISA is making a list of the most critical open source components that should be made more secure, either by providing “skilled people, or AI models and tools,” Lepassaar said. The EU also has an open source strategy, in part with a view to weaning itself off U.S. tech.
Those plans were made before AI rattled cybersecurity defenders worldwide. Some, like the German Sovereign Tech Agency, hope that the incident will put the needs of open source software on the map and prompt an intervention.
But many of those steeped in the ecosystem, like Adermann, have their doubts.
“I’m a bit pessimistic,” he said. “I worry that people will just not care enough to actually do anything about it.”
Amazon, Walmart, Macy's and Target have been selling a cute cuddly toy called the Pixie-Paw Smart that will tell stories to and converse with your toddler. When NBC News tested the toy by asking various questions, it happily described horrific violence in detail. — Read the rest
HELSINKI — The European Commission should withdraw and resubmit its proposal for a new EU-wide corporate law framework, startup lobby groups said, denouncing the current text as a “Frankenstein” proposal.
A position on the law adopted by EU capitals in October “weakened or removed some of the very provisions that would make EU Inc. a true European standard,” Startups for Europe and EU Inc. said in a joint press release.
Both groups pitched the withdrawal to European Commissioners Henna Virkkunen and Ekaterina Zaharieva, who oversee tech policy and startups, during a policy roundtable at Finnish technology conference Slush on Thursday morning.
They also sent a joint letter to European Commission president Ursula von der Leyen earlier this week.
The European Commission published its proposal for a so-called 28th regime, also known as EU Inc., earlier this year, to harmonize how founders can incorporate a company and reward their personnel, so that startups can scale more easily across the bloc.
Lobby groups now say that EU capitals have rolled back key provisions, such as the freedom to set up shop in any EU country and simplified insolvency rules. With their demand to withdraw and submit a new proposal, the groups hope that EU capitals will rethink their position.
“This is an opportunity for Member States to reconsider their positions and make sure EU Inc. delivers on its original promise,” said Simon Schaefer, president of Startups for Europe.
Jeannette zu Fürstenberg, managing director of General Catalyst, a fund that backed leading European companies such as Helsing and Legora, also criticized the EU Inc. proposal as it stands.
“If it comes out like that, it’s going to be Frankenstein that the whole world will laugh at,” she said in an interview at Slush.
She said the EU Inc. proposal should specifically address how stock options, often used to reward startup staff, are taxed.
“If you need to pay dry income on something that’s essentially illiquid, you’re bankrupting the people building it,” zu Fürstenberg explained.
A separate group of 38 tech executives, including Spotify founder Daniel Ek, Exor CEO John Elkann and Skype co-founder Niklas Zennström, took a softer line in a declaration published at Slush. They called EU Inc. “the most important first step” toward a true single market and did not ask for the proposal to be withdrawn.
They also urged Brussels to enforce the Digital Markets Act, the EU’s rulebook for the largest online platforms. Europe must “simplify regulation that limits growth and at the same time enforce competition law and the Digital Markets Act,” the declaration said.
Still to play for
The EU Inc. legislation has also hit a roadblock in the European Parliament, where lead lawmaker René Repasi’s report on it has yet to clear the committee stage. Repasi, a German Social Democrat, told POLITICO he could understand the frustration of the supporters of the original legislation.
“The political groups that in principle support EU Inc. should now quickly reach a common parliamentary position,” Repasi said, adding that this would enable the Parliament to negotiate in favor of key provisions for startups in the interinstitutional negotiations that are still to come on hammering out a legislative compromise.
Asked for comment, EU Justice Commissioner Michael McGrath said the objective from the outset had been “to establish a genuinely European corporate framework that is simple, attractive and enables innovative companies to start, grow and scale across the Single Market.”
While he did not directly address the criticism of the startups, McGrath said the Commission stood ready to work with the European Parliament and Council of the EU to “preserve the ambition” of the legislation.
“The forthcoming trilogues will be an important opportunity to address outstanding concerns and work towards an agreement that delivers real benefits for companies across the Single Market,” McGrath said in a statement to POLITICO.
Nothing gets people thinking about the end of civilization quite like a major technology outage and mass freak-out. Last week, Microsoft went down, flights were grounded, and reports of tentacled alien sightings surfaced. The internet had some concerns.
Microsoft services went offline on November 12, followed by a nationwide FAA ground stop the next day that halted flights for more than four hours. The timing raised suspicions, especially since the FAA had introduced an AI-powered air traffic tool called SMART just days earlier. Officials reported “no evidence” of a cyberattack.
The official explanations were far less exciting. Microsoft blamed an earlier network upgrade, while Transportation Secretary Sean Duffy blamed it on outdated FAA equipment. “We have old, old switches, old routers. We use 1980s little copper lines,” he told reporters. Comforting stuff.
But if our technology really did go down for an extended period, how well would we manage? A 2023 FEMA survey found that just 51% of American adults believed they were prepared for a disaster.
A 70-year-old Navy veteran known as Prepper Larry told the New York Post he’s not optimistic about the rest of us. “Each generation seems to get dumber and dumber about the most basic things in life,” he said.
So, how would you do without the little computer in your pocket?
1. Can You Find North Without Your Phone?
GPS has made getting around ridiculously convenient, but once your phone dies, you’re on your own. The sun rises in the east and sets in the west, so you can use its position to get your bearings. After dark, look for the Big Dipper. The two stars at the outer edge of its cup point toward Polaris, the North Star, which will help you identify north.
2. Can You Still Read an Analog Clock?
Remember when telling time meant looking at an actual clock instead of your phone? It’s a skill younger generations haven’t had to rely on nearly as much. The short hand tells you the hour, and the long one tells you the minutes. You’d think we’d have that one covered, but here we are.
3. Would You Know Which Plants to Avoid?
TikTok is full of people wandering through the woods collecting mushrooms and wild plants for dinner. Looks fun, until you remember that some poisonous plants look almost identical to edible ones. And remember: just because you see birds or squirrels eating a plant doesn’t mean you can eat it, too. If you can’t positively identify something, leave it alone.
4. Could You Make Your Food Last?
You could have a pantry full of emergency food and still find yourself running low after a few days if you’re not careful. Take stock of what you have, plan out your meals, and resist the urge to stress-eat your way through half your supplies. Canned goods and other foods that don’t need refrigeration or cooking are also smart to have around, especially if the power stays out longer than expected.
5. Do You Know Where to Go When There’s No Bathroom?
Indoor plumbing is one of civilization’s greatest achievements, something you’ll really appreciate the moment it stops working. If you’re outdoors without access to a toilet, Leave No Trace recommends digging a cathole 6 to 8 inches deep, at least 200 feet from water sources, trails, and campsites. Local rules vary, so check them before you go digging.
We’re probably not facing the end of civilization anytime soon. But it’s a little embarrassing to think about how many of us would struggle to get through a single day without Google.
Former Nvidia CEO Jensen Huang sits for an interview with Bloomberg News in London on October 14, 2026. | Bloomberg via Getty Images
If you didn’t know a lot about Nvidia a month ago, what you know now is salacious, essential, and probably a little bit misleading.
The multitrillion-dollar chipmaker has primarily been in the news because its co-founder and CEO Jensen Huang had a few affairs — one with a possible spy — and resigned. This has been a huge story, not just because of the sex scandal but also because Nvidia has been propping up the entire AI industry, as its primary chip supplier and frequent financier, and, by proxy, propping up the global economy.
The downside risk is almost catastrophic: If Huang brings Nvidia down with him, the whole economic house of cards could collapse.
That doesn’t mean it will avoid a reckoning, But the reckoning is much broader than the Huang storyline — and if it comes, might even be beneficial.
Nvidia is facing a raft of investigations and potential new regulations as a more AI-skeptical Congress prepares to take power, not to mention growing competition from tech giants, like Google, that have started making their own highly advanced chips for AI. There are also signs that the chips that made Nvidia so profitable won’t be as critical in the next chapter of the AI era. So maybe Nvidia was already in trouble before the world learned more than it wanted to about Jensen Huang’s sex life.
More broadly, a shakeup in the apparently very precarious AI industry could end up being a good thing. Free market capitalism tends to thrive when there’s more competition. And, just maybe, we stand to benefit from more scrutiny into tech giants like Nvidia and their executives. There’s reason to believe real shakeup at Nvidia — even a fall from its exalted perch and its role as the money factory of the AI world — could turn out to be the right thing for consumers, healthy for the market, and useful in the event of the AI bubble bursting.
Why Nvidia matters to normies (yes, even you)
You might be wondering what any of this has to do with you. Nvidia’s top chips can cost more than a car, and precious few humans have ever even seen one. Why should the escapades of a leather jacket-loving tech executive distract you from your hobbies or affect your daily life? For the most part, it won’t. Even if you’re an enthusiastic Claude or ChatGPT user, what happens to Nvidia won’t change the way you use this technology. AI as we know it won’t immediately get better or cheaper if competition heats up in Silicon Valley. The stock market might get shakier, though. In that sense, we’re all in this together.
“American consumers likely may not feel Nvidia’s pain in their own pocketbook, but they may likely feel it in their 401(k),” said Margaret O’Mara, a historian at the University of Washington. “We’re all stockholders, whether we like it or not.”
If you’re enrolled in your employer’s 401(k), you probably own some Nvidia stock, either directly or through a mutual fund. Even if you don’t, any company as rich and central as Nvidia can cause real shock waves in the economy. The simple explanation for this is that Nvidia and a tiny handful of big tech company stocks make up a massive share of the stock market. The “Magnificent Seven” stocks account for about a third of the entire S&P 500 index — yes, that’s seven out of 500 total companies — and they haven’t been doing all that well lately. At its peak in May, Nvidia alone accounted for nearly 9 percent of the S&P 500. (It’s currently around 6 percent.) That meant that when its stock tanked following the Huang scandal, it knocked thousands of dollars off millions of retirement accounts.
This is not a reason to rebalance your portfolio, and I’m not in the business of giving investment advice. But the panic people might be feeling right now around the future of Nvidia and the possibility of an AI bubble is a helpful reminder that panicking is almost always a bad idea. An instructive example is the Panic of 1873, when one dominant company faltered, caused a run on the banks, and pushed the United States into what’s now called the first Great Depression. Plus, because the company is just so big, it’s actually pretty hard to get Nvidia stock out of your retirement accounts completely.
The good news, on this front, is that Nvidia probably isn’t going anywhere. We learned on Tuesday evening that the company continues to make money hand over fist, with $115 billion in revenue in the third quarter of this year, which is up over 100 percent from the same time last year. The advanced graphics chips, or GPUs, that Nvidia makes continue to be the gold standard for AI tasks, and perhaps more importantly, Nvidia has built an impressive moat around its chip business to shield itself from competitors.
Nevertheless, the competitors have been moving in. AI has gotten so good at writing software, for instance, that it’s possible to rewrite code designed for Nvidia chips to run on those made by competitors like AMD. Google is also making advanced chips called Tensor Processing Units (TPUs) as a challenger to Nvidia’s GPUs. (Google actually rolled out a whole new generation of TPUs in early November, just a few days before Jensen Huang resigned.) Amazon has been filling AWS data centers and powering Anthropic’s Claude with its Trainium chips, another competitor to Nvidia GPUs. Microsoft, meanwhile, has its Azure Maia chips and expects higher demand after this Nvidia drama.
A lot of companies still want to buy Nvidia products, though. Nvidia has been able to pull in 75 percent or more in profits on its chip sales because they’re in such high demand, according to Chris Miller, a professor of international history at Tufts and the author of Chip Wars: The Fight for the World’s Most Critical Technology.
“Nvidia’s argument has always been, ‘Yeah, we charge a pretty thick profit margin on top of our chips, but we give you the best possible outcome,’” Miller said. “But its pricing power would decline if the competition increases.”
Competition, as any student of capitalism knows, should lead not only to lower prices but also more innovation and better-quality products. It’s still not clear what the downstream effects of cheaper, faster, more powerful AI chips might be, though.
Most people encounter AI not through any direct contact with Nvidia, which sells its products to other companies. They use ChatGPT or Claude, and they may even pay a subscription fee for access to better models. There’s no guarantee that OpenAI and Anthropic spending less on training and running their models would lead to lower prices or even better models for consumers to use. It might just mean more profit for OpenAI and Anthropic.
Shakeups are good, actually
To the casual observer, the Nvidia shakeup looks quite scary. Again, this is one of the most valuable companies in the world, and it’s only ever had one CEO, Jensen Huang, who’s steered the company to unprecedented levels of success and dominance in his 33 years at the helm. And then the scandal that brings him down not only involves extramarital affairs — with two Nvidia employees as well as an alleged Chinese spy — but also geopolitical consequences and national security threats.
Then there’s the AI bubble of it all. Anxiety around an AI bubble has been simmering for a couple of years as the level of spending on the AI buildout has hit sky-high levels along with valuations of AI companies, many of which are not yet profitable. In fact, under Huang’s leadership, Nvidia has invested nearly $100 billion in AI companies, including OpenAI, making it one of the biggest financial backers in the sector. Continued chaos at Nvidia could upset the already fragile AI ecosystem, which is propping up the US economy. The bubble could burst. According to former SEC chair Gary Gensler, in fact, it will burst by 2028, sending the country into a recession.
On the other hand, all the renewed interest in the possibility of an AI bubble and its imminent bursting is a good thing. Jensen Huang’s affairs have directed a lot of attention and reporting resources to the state of the AI industry. Shane Greenstein, a professor at Harvard Business School, says that previous bubbles — such as the dot-com boom — were more opaque, and would have benefited from more scrutiny. He thinks we’ll know in the next one to three years what will happen with the apparent AI bubble, and he says it doesn’t matter if Jensen Huang is around for it or not.
“We’re already over our skis, as it were,” Greenstein said. “We’ve already spent hundreds of billions of dollars on data centers, so there’s no way to pay those back unless we get both productivity advance in a range of industries beyond just coding assistance, and widespread consumer adoption with some revenue-generating activity attached to it — which at the moment ChatGPT is the only successful consumer-oriented thing with any kind of revenue, and it’s still a pittance.”
In the worst-case scenario, however, the scrutiny is what might save us. The dot-com boom was propped up on more hype and less reliable data than the AI boom, according to Greenstein, which led some companies to do some shady reporting. When it went bust, the financial reckoning was worse because of that bad reporting.
“By comparison, whatever downturn we get isn’t going to be as bad as it would’ve been had things been a bit more obscure,” Greenstein said. “Now, that doesn’t make anybody feel better, but I think it’s interesting to live through this where both the skeptics and the cheerleaders are going at it all the time in public, which is quite different from the dot-com boom, where only the cheerleaders were given the microphone.”
It doesn’t make anyone feel better, and it doesn’t necessarily protect your 401(k) from some ups and downs in the years to come. Think of what you’re feeling as a productive sort of anxiety. You’re worried, but you’re paying attention, which is inevitably better than the opposite.
An artificial intelligence (AI) system is judged on the job it does, not on how clever it is in general. For European businesses, that is an opening. For example, a pharmaceutical company does not need a model that can discuss biology. It needs one that can flag the rare adverse-event signal in its own trial data. An energy grid operator needs load forecasts built on its own network topology, not generic ones. A defense manufacturer needs a system that can verify a weld against its own specification, and say why it passed or failed. Each needs a system that understands its domain, works on its data, follows its rules and guardrails, and can be trusted to execute in a protected, controlled environment.
An edge Europe can’t buy
That task-specific fit is Europe’s opening to turn its industrial expertize into an AI edge, because that edge was never going to come from outspending anyone. US private AI investment reached US$285.9 billion in 2025, more than 23 times the US$12.4 billion invested in China — though looking at just private investment figures likely understates China’s total AI spending, given its government guidance funds. The Stargate venture plans to invest between US$100 billion and US$500 billion to build advanced AI data centers across the US by 2029. Matching that euro for dollar would be a strategic mistake, and Europe’s grids rule it out regardless: connection queues in major hubs already run seven to 10 years. None of this means Europe should stop building AI factories. They matter for critical workloads and for a credible European alternative. The point is not to scale compute, but to invest smartly in the infrastructure needed to run specialist industrial applications. A factory with nothing to run on is just an expensive data center. The number that matters is not compute deployed, but the number of specialist applications built, the tasks addressed and the vertical industry solutions delivered.
Breadth first, depth next
The same pattern shows up globally in how AI is actually used. Our September 2026 CIO Pulse survey of 300 technology leaders worldwide found that 70 percent were using AI for individual productivity: particularly copilots and general-purpose tools adopted across the organization. Only 37 percent were using it to orchestrate processes end-to-end with agents, where AI sits inside a process and solves a real business problem in context, as a specialist. Breadth of adoption is arriving first, and faster, than depth. Depth is harder, because it needs specialization, context and precision, not just more scale, and that is where Europe can lead.
Sovereignty by design
Another factor in European AI adoption is non-negotiable: sovereignty. In mid-2025, 61 percent of the business leaders we surveyed called it a strategic priority. By early 2026 that figure exceeded 80 percent. The June 2026 suspension of access to two frontier models under US export controls showed how fast the terms of access can change, and how little say any European company had in the decision. Sovereignty itself has layers: where infrastructure sits, where data lives, which models run, and now, above all three, what knowledge those models are grounded in. It has to be designed in from the start. It cannot be bolted on after a system is already live. Being sovereign while staying part of the global technology landscape is Europe’s real competitive advantage. That balance only works through specialization: the precision to control what matters most while still drawing on the best of what the world builds. This is a wave Europe is well placed to ride.
We call this operating logic Specialist AI in our report, AI designed with intention. How Europe can win with Specialist AI: model size, data, deployment location, inference pathway and governance controls matched deliberately to a defined task. Frontier models suit open-ended reasoning, cross-domain synthesis and genuine exceptions, which we estimate account for roughly 10 percent of enterprise workloads. Smaller domain-tuned models suit repetitive, high-volume work where latency, cost, auditability or local control matter most. The strongest architecture uses both, deliberately, not by accident.
The point is not to scale compute, but to invest smartly in the infrastructure needed to run specialist industrial applications. A factory with nothing to run on is just an expensive data center.
The European Patent Office (EPO) shows what this looks like in practice. With Mistral AI, it processes 400,000 pages a day on a one-billion-parameter model fine-tuned on 50,000 patents. The error rate is below 1 percent, and work that used to take five days now takes minutes. The model runs on the EPO’s own domain data, in a sovereign environment. It is not the biggest model in the world. It is the right one.
Upcoming regulation, including the proposed Cloud and AI Development Act which was put forward in June, will likely only reinforce this shift and the opportunity it creates, especially through its multi-layer sovereignty definitions. Companies should start shoring up governance that holds across every layer of the stack now and understand that different parts of the stack carry different types of risk, with kill-switch exposure varying by layer. The future needs sophisticated controls, but it also rewards whoever builds them first.
Turning advantage into action
Three moves follow for enterprises:
Build demand alongside capacity. Develop the pipeline of industrial applications, from predictive maintenance to mission-critical decision support, while infrastructure is provisioned.
Design for choice. Classify data, workloads and applications to find the right deployment model and governance for each, and keep the ability to change models as the market moves.
Govern the knowledge layer. Set rules for grounding, provenance, evaluation and retention of derived knowledge, and for what any provider may learn from your data.
Policymakers can reinforce the shift. Compute investment should connect to sectoral test facilities and adoption centers, not sit apart from them. Procurement should reward task performance, auditability, energy efficiency and portability, not just scale. Standards should enable interoperability and federation, so every provider adds to a single European fabric of capacity rather than 27 disconnected ones.
Europe may not lead on AI breadth, but it can lead on AI depth. By combining generations of industrial expertize with sovereign, specialist AI, designed for specific tasks and deployed on Europe’s terms, businesses can turn the potential of AI into lasting competitive advantage.
AI Designed with Intention: How Europe Can Win with Specialist AI
BRUSSELS — Gamers of the world will be taking the day off Thursday to play Grand Theft Auto VI, billed as the most anticipated digital entertainment launch in history. The EU’s buttoned-down staffers may be too busy to notice.
Yet even a diligent Brussels bureaucrat has reason to buckle up for the latest installment of the Grand Theft Auto series, which comes 13 years after the previous version of the controversial video game that challenges players to go on crime rampages around U.S. cities to the tune of an eclectic soundtrack.
According to Sensor Tower data, GTA VI is already the biggest entertainment launch ever, with 17 million copies sold for a total of $1.6 billion (€1.46 billion) before it’s even released.
The game lands in a vastly different world than the one the previous version entered in 2013, one where EU regulators are trying to protect kids from harmful online content and handle the onslaught of AI. Here are six reasons an EU policy maven should pay attention to all the fun and games:
1. It’s a European game, sort of
The Grand Theft Auto series was originally made by two Scottish programmers, David Jones and Mike Dailly, at a Dundee game studio called DMA Design. The rights to the series were later sold to an American company, Take-Two Interactive, in the late 1990s, although its subsidiary Rockstar Games still has large offices in Edinburgh.
Meanwhile, the EU’s right to claim a Scottish success story as its own lapsed with the 2020 EU-UK withdrawal agreement.
A young man plays Grand Theft Auto IV on the game’s day of release on April 29, 2008 in London. | Cate Gillon/Getty Images
European leaders worried about the bloc’s declining competitiveness are turning to a number of policy solutions to encourage innovation, from new funding mechanisms to changes in public procurement rules.
2. EU rules may be on gamers’ side
Prior to the game’s release, gamers were concerned that GTA VI’s online version would include the same micro-transactions and in-game currency that fans say are making games “dull and repetitive.” The game being released today won’t have these features, although a later online version might.
The European Commission is proposing new rules to curb the use of these features by kids, or at least make them more transparent, because Brussels says they encourage spending. The gaming industry says these proposed rules under the Digital Fairness Act will be catastrophic for EU businesses.
3. Kids aren’t supposed to play GTA
GTA VI is rated 18+ under the PEGI system, which stands for Pan-European Game Information rating mechanism. That means it’s not recommended for minors to play at all.
The rating is a self-regulatory tool created 20 years ago today and still in place. It is based on whether the games include things like profanity, drugs, violence and sex — all hallmarks of the GTA franchise.
Games rated 12 and up are reviewed by the U.K. Games Rating Authority, whose reviewers actually play the game before assigning a rating.
4. But Brussels can’t stop them
Some EU countries and the U.K. do require buyers to show identification when purchasing a physical copy of the game. Online stores like PlayStation’s also carry out age checks to make sure that 13-year-olds don’t get access to age-inappropriate content.
But nothing stops a parent or an older sibling from giving that game to a younger person.
PEGI’s Director-General Dirk Bosmans said that “at the heart of PEGI is helping parents making informed choices.” If a parent wants to give their teenager access to GTA VI, the PEGI system lets them do so knowing that it contains violence, drugs and sex.
“We are more concerned about parents saying, ‘here’s some money, buy whatever you like’,” Bosmans said.
5. It may be the last big game not made by AI
The CEO of Take-Two Interactive, Strauss Zelnick, has said that generative artificial intelligence has “zero part” in what the firm has built for GTA VI.
The gaming industry is increasingly using generative artificial intelligence to create new titles. The latest State of the Game Industry report found that 40% of game developers use genAI in game creation, including in work on some of the biggest titles. These include the shooter Call of Duty and critics’ favorites like France’s Clair Obscur: Expedition 33.
The industry has also seen thousands of layoffs, with one-third of U.S. video game industry workers saying they have been let go in the past two years.
6. It’s a safety-first world
As the industry has moved into the mainstream with blockbuster releases like GTA VI, the European Commission has firmly set its sights on creating the first industry-specific rules for the gaming industry through the DFA and broader rules on protecting minors under the Kids Act.
In the video game section of Fnac, a child tests a Nintendo Switch demo unit displaying Mario Kart 8 Deluxe on a large screen, amidst the accessories and consoles on display in the store in Paris on August 1, 2026. | Riccardo Milani/Hans Lucas/AFP via Getty Images
The idea that violent video games contribute to violence in real life was largely debunked in the early 2010s, around the time the previous GTA installment was released, Bosmans says. But the debate around content has moved from what is “harmful” to what is “inappropriate,” he says.
Scientific research informs the PEGI system, but the regime also reflects societal norms around “appropriateness,” he explained: one f-word can land a whole game in the 16+ category.
As regulators introduce more measures to protect kids online, Bosmans said, “it feels like we’re now in a stage where we may have forgotten some of the history,” and the difference between what is harmful, what is banned, and what is inappropriate.
LUXEMBOURG — The European Commission was accused Wednesday of withholding internal messages that explain why it dropped a contentious patent proposal, amid claims of pressure from U.S. Vice President JD Vance.
At a court hearing in Luxembourg, judges were told that German center-left MEP Tiemo Wölken asked the Commission for internal notes on why the proposal was abruptly killed in early 2025. He was told there was only one relevant internal document: minutes of a private meeting between heads of cabinet, which the Commission refused to hand over.
Wölken’s lawyers said the absence of information was hard to believe given how contentious the proposal had become.
It was “unbelievable” that there were “zero emails on this highly-politicized, economically-relevant procedure,” Carsten Zatschler, Wölken’s lawyer, said at the EU’s General Court.
“All of this apparently happened without any form of electronic communication,” Zatschler said. “That’s quite simply not credible.”
In February 2025, the Commission announced it was withdrawing plans for a standard essential patents (SEPs) regulation on the licensing of patents, saying there was no “clear signal of support” from co-legislators. Negotiations between the European Parliament and EU countries were ongoing at the time. The proposal was officially killed in July that year.
The proposal had triggered a fierce lobbying battle between patent holders — including Nokia, Ericsson, Qualcomm and Huawei — and companies that license their technology, from phone makers to carmakers.
The withdrawal was announced on the same day that Vance met with Commission President Ursula von der Leyen at an AI summit in Paris, where they discussed AI, Ukraine and transatlantic ties. The timing prompted questions in the European Parliament about whether their conversation played a role in quashing the proposal.
Zatschler asked: “Why was it surprisingly withdrawn? Whether a meeting between Ursula von der Leyen and U.S. Vice President JD Vance played a role, that’s obviously not something we need to discuss here.”
“But whether it can be clarified should not be dependent on whether the European Commission categorized emails as important or not, because otherwise we would be putting the fox in charge of the hen house,” he said.
Wölken was part of the negotiating team blindsided by the decision.
“My team had met the Polish presidency [of the Council of the EU] just days before, and they were gearing up to finalize the Council position,” the lawmaker told POLITICO. “No one was informed.”
“That alone raises questions about the withdrawal, which for me turned into a real concern after hearing rumors about the meeting between Ursula von der Leyen and JD Vance,” he said.
The Commission did hand over 14 documents to the German MEP, most of which were emails and position papers from outside stakeholders sent to the Commission after the decision to kill the proposal was announced.
Lawyers for the Commission told judges that Wölken’s request didn’t specify that he was seekend internal emails predating the decision.
“These are professional lawyers,” said Tim Maxian Rusche, representing the Commission (Wölken is also a lawyer). “They work directly in Brussels. If they want internal correspondence, they’re going to have to specify that.”
He added that the Commission is “being exposed to more and more fishing expeditions” where it is asked “very general” questions in access to document requests.
“They need to make [requests] clear, because otherwise we can’t do our work,” he continued.
The Commission is also facing a lawsuit from the European Parliament over the decision to kill the proposal. Last year, lawmakers voted to sue the Commission over the decision, arguing that it had overstepped its competence.
“The Commission’s right to withdraw a proposal … cannot be used as a political instrument to short-circuit Parliament’s work or to enforce a deregulation agenda from above,” German Social Democrat René Repasi said. “This is not in line with how the democratic processes in the European Union are meant to function.”
The judges did not confirm when a decision would be made in the case.
Germany and Austria say Chinese-owned TikTok should come under European ownership to safeguard personal data, boost sovereignty and curb the platform’s “openly visible threat to our democratic society.”
The joint proposal Wednesday by top government ministers of the two countries comes after President Donald Trump forced the separation last year of TikTok’s operations into a U.S.-based entity, with its original owner, Beijing-based Bytedance holding about one-fifth of the new entity.
Trump’s campaign to initially ban and later split up TikTok cited national security concerns. Now, Berlin and Vienna are going down a similar route.
With 200 million EU users, TikTok is “a central player with significant relevance for opinion formation and social participation,” Germany’s media minister Wolfram Weimer and Austria’s State Secretary for Digitalization, Alexander Pröll said in a joint statement.
A TikTok spokesperson said there was “zero merit” to the proposal.
The platform’s role in elections is being investigated under the EU’s Digital Services Act, while regulators have fined the firm for sending Europeans’ data to China, in breach of privacy rules.
The ministers say that TikTok is a “visible threat to our democratic society” and that a “fundamental solution should therefore start where the threat to Europe originates: with the Chinese parent company ByteDance, its alignment with Chinese interests, and the resulting continuously unclear access possibilities of the Chinese state to the data of millions of European users.”
The statement comes ahead of a meeting of culture ministers taking place in Brussels Thursday. Weimer and Pröll called on other EU capitals to join their bid, but didn’t clarify how the new structure would work.
POLITICO parent Axel Springer has said it would support any initiative to bring TikTok under European control in response to reports that the media conglomerate was looking for a stake in the social media platform.
This article has been updated with TikTok’s response.
The unfinished work of nuclear governance offers an important lesson for those struggling with the risk of AI now: learning to live with a dangerous technology is not the same as eliminating its threat. | Daniel Gnap/NurPhoto/Getty Images
When Illinois Gov. JB Pritzker talks about the threat of artificial intelligence, he offers an increasingly common point of comparison: it is a danger “on the level of nuclear weapons.”
The comparison is not far-fetched. If the international community found ways to constrain a potentially catastrophic technology once before, perhaps those arrangements can help us constrain another.
The trouble is that while nuclear nonproliferation may be a useful framework for thinking about regulating AI, it isn’t necessarily a good one. Not because the technologies are too different, but because nuclear nonproliferation hasn’t actually freed us from the prospect of nuclear war.
Nevertheless, the unfinished work of nuclear governance offers an important lesson for those struggling with the risk of AI now: learning to live with a dangerous technology is not the same as eliminating its threat.
What we might be able to do, however, is eliminate uncertainty about who’s building AI, what AI is capable of, and when AI puts people at risk. The US has done exactly that with nuclear weapons by declassifying information about the size and condition of its own arsenal, encouraging other nuclear powers to do the same.
Prioritizing transparency about AI wouldn’t eliminate the dangers altogether, but it might help us move forward more collectively as we strive to manage the very real risks of this emerging technology.
What the nuclear-AI analogy gets right… and wrong
Comparisons between AI and nuclear weapons have emerged not just from politicians and commentators but from those at the very top of the frontier AI food chain. For some AI leaders, nuclear weapons also represent an ideal model for oversight.
In a 2023 essay, OpenAI’s Sam Altman, Greg Brockman, and Ilya Sutskever called for “something like an [International Atomic Energy Agency] (IAEA) for superintelligence efforts.” Their proposed international authority would “inspect systems, require audits, test for compliance with safety standards,” and restrict how the most powerful systems could be deployed.
The goal of such measures is to make the development of superintelligence safe, rather than prevent it, much as nonproliferation regimes stop short of banning nuclear weapons outright. Altman and his colleagues argue that the technology could deliver enormous benefits, while stopping its creation would be “unintuitively risky and difficult.”
Not everyone invoking nuclear governance draws the line in the same place. In September, AI researcher Yoshua Bengio told the UN Security Council that “frontier AI should be licensed like… medicine, aviation, and nuclear energy to incentivize safe development.” He also called for a universal constraint on AI development, one more expansive than those imposed on nuclear weapons: “No one, no company and no country, should develop an AI system that could cause large-scale harm.”
The real problem with such schemes may be figuring out how they would be enforced — and by whom. Sure, there are fewer states with nuclear weapons today than there could have been if nonproliferation had not become a policy priority for the US. However, this was not the result of diplomatic prowess or good-faith restraint. In the most important cases, it was a result of coercion by the US and its allies.
As Christopher David LaRoche and Ankit Panda, two nuclear experts at the Carnegie Endowment for International Peace, explained in Foreign Affairs: “South Korea and Taiwan gave up nascent weapons programs after intense pressure from the United States…and a threat to pull military support if they did not comply. Israeli airstrikes took out the Iraqi and Syrian programs; Libya surrendered its own under sustained U.S. and British pressure.”
In other words, unless the US seeks to go to war to prevent countries around the world from getting hands on frontier AI, nuclear nonproliferation is not the ideal mold for AI regulation. Plus, even in the face of a robust nonproliferation regime, a few states managed to slip through the cracks — India, Israel, North Korea, and Pakistan each acquired nuclear weapons.
Despite that, there are still fewer nuclear weapons in the world today than there were at the peak of the Cold War. But even the best-laid plans might not last forever. The arms control agreements that made such disarmament possible have all crumbled — whether due to cheating, repudiation, or expired treaties. Moreover, in addition to the increased number of states in possession of nuclear weapons, China has tripled its number of nuclear weapons just this decade.
These are the consequences of a regime built to manage a technology, rather than eliminate it. A similar frontier AI nonproliferation regime would govern who may build the most powerful systems, while leaving unaddressed the question of whether anyone should have that option in the first place.
A similar frontier AI nonproliferation regime would govern who may build the most powerful systems, while leaving unaddressed the question of whether anyone should have that option in the first place.
The greatest lesson is for the US
Today, the threat of nuclear war persists. However, nuclear nonproliferation offers us one powerful takeaway: the United States, in particular, has a tremendous role to play in shaping the future of regulating frontier AI.
“In the nuclear domain specifically, the US historically has had a powerful norm-entrepreneurship function,” Panda told me. “What the United States does when it comes to nuclear weapons does get noticed by other countries and often gets adopted by other countries.”
This has historically involved leadership in transparency, Panda said. “The US, to its credit, is much more transparent than every other nuclear country when it comes to talking about things like command and control and what we do and don’t do.” For example, the US regularly declassifies and publishes data regarding the size of its nuclear weapons stockpile, fissile material inventories, and production facilities.
In a similar way, the US could lead the world in transparency about the frontier AI it is pursuing, the limits it might impose on that technology, and the safeguards in place to keep it under control.
This could take shape in a few ways, beginning with the US government’s own use of AI:
Before deploying a powerful and risky AI system, the government could disclose what risks it tested for, the conditions under which the system is considered reliable, its known limitations, and what findings would prevent deployment.
The US could explain which military decisions AI may contribute to and which require human authorization. This might, for example, involve distinguishing intelligence analysis from authority to initiate an attack. The government could also identify military applications of AI that it rules out altogether.
The US could publish accounts of failures or unexpected behavior, explaining whether a system was withdrawn, restricted, or retested after something went wrong.
These transparency measures should also extend to the private sector, and the government could require companies to:
Disclose dangerous-capability testing: This would involve requiring frontier developers to publish standardized summaries of tests for risks such as assisting cyberattacks, contributing to biological weapons development, or operating autonomously beyond intended limits.
Report serious incidents and near misses: Developers and deployers could be required to notify a regulator of significant safety failures, breaches, or unexpected dangerous behavior.
Publish the conditions for pausing development or withholding release: Companies could disclose in advance which findings would trigger additional safeguards, delay product launches, or halt further development.
As AI continues to reshape the world we live in, prioritizing transparency around the technology’s achievements — and limits — is one way to ensure that safety and security remain central.
As AI continues to reshape the world we live in, prioritizing transparency around the technology’s achievements — and limits — is one way to ensure that safety and security remain central. Panda suggested that such measures could reassure adversaries that the US is not going to implement frontier AI models in ways that might be dangerous or irresponsible.
An aerial view of a 33 megawatt data center with closed-loop cooling system on April 14, 2026 in Vernon, California. | Mario Tama/Getty Images
The stock market is set to christen a new trillion-dollar company on Wednesday, as the AI juggernaut Anthropic goes public.
Key takeaways
Fears of an AI market crash are growing, with tech stocks sliding sharply since June.
The AI industry is spending far more on infrastructure than it earns, and cheaper “good enough” models threaten the profits of frontier labs like OpenAI and Anthropic.
Three recent shocks have made things worse.
The boom may survive, but AI companies can no longer count on governments deferring to them or markets cheering them on.
The IPO should have been the crowning triumph of the AI boom — a testament to the technology’s mind-bending commercial value, which enabled a seven-person startup to become a world-conquering juggernaut in five short years.
But all is not well with Silicon Valley. Even as investors scramble to claim a piece of our future God-emperor, Claude Opus 5,000 (hallowed be his name!), fears of an AI market crash are building.
Since June, tech stocks have taken a beating, with one generative AI ETF shedding 35 percent of its value. Over the same period, the Philadelphia semiconductor index — a rough gauge of the chip industry’s value — fell by more than one-third.
The sell-off gained steam in recent weeks, as Republicans lost key elections to AI-skeptical Democrats — and Nvidia lost its longtime CEO to a geopolitically explosive sex scandal, a leadership change that has overshadowed the company’s persistently strong earnings. Meanwhile, every mysterious IT problem at a major government institution or company is being scrutinized for signs of AI hacking — which could bring even more unwanted attention from politicians, regulators, and the general public.
Now, after years of deriding the AI boom as a “bubble,” some of the industry’s bears are finally hearing a popping sound: In their estimation, the bubble has already burst.
Such declarations are premature. To qualify as a punctured bubble, the decline in AI assets would need to become much deeper and more sustained than it’s been thus far. As of this writing, the tech-heavy Nasdaq index remains more than 4 percent higher than it was when the year began.
Nevertheless, there is an increasingly strong case for believing that investors drastically overvalued AI-related companies, and that recent events might speed the AI industry toward a day of financial reckoning. At the very least, the AI boom’s charmed childhood seems to be giving way to a more troubled adolescence. And given the centrality of artificial intelligence to today’s economy, a large-enough correction could tip the US into a recession, thereby raising unemployment, squeezing wages, and quite possibly accelerating the automation of white-collar jobs — all of which could exacerbate Big Tech’s political headaches.
“Once the fever breaks, it never really comes back,” Brent Donnelly, a former HSBC trader and president of the financial media company Spectra Markets, told me. “People aren’t going to be as optimistic now that the cracks have begun to show.”
AI bubble-o-nomics 101
Financial analysts have been wringing their hands over a potential AI bubble for years. And not without reason.
The AI boom’s scale alone makes a case for its unsustainability. According to Goldman Sachs, US investment in AI infrastructure this year will reach $581 billion, or roughly 1.8 percent of America’s entire GDP. Next year, the bank expects such capital expenditures will add up to 2.5 percent of our nation’s total economy.
For context, at the peak of the dot-com boom in 2000, spending on telecommunications and fiber totaled only 1.15 percent of America’s gross domestic product. In terms of magnitude, the closer analog to today’s AI frenzy might be the railroad buildout of the late 19th century.
Of course, both the dot-com and railroad booms ended in tears. Investors weren’t wrong to think that trains and the internet would prove immensely valuable. But in their excitement about each technology’s promise, they overestimated near-term demand for many rail lines and fiber optic cables.
And there’s reason to think that today’s AI investors are making a similar mistake. According to an estimate from the consulting firm Bain & Company, the AI industry will need to generate $6 trillion in annual revenues by 2031 to justify its spending on data centers and chips. The sector’s current revenues are closer to $540 billion, according to an estimate from Grandview Research. Thus, tech investors are essentially betting that demand for AI is poised to soar at a mind-bending rate.
Yet this hope may rest on a faulty extrapolation from Silicon Valley’s own experience. Artificial intelligence has rapidly penetrated the software industry, delivering large productivity gains in the process. But outside of tech, corporate interest has been more tepid. According to a 2026 NBER study, just 2 percent of firms account for over half of AI enterprise spending. And among companies that deployed AI, 80 percent discerned no productivity benefits from the technology.
“There’s a lot of real revenue being generated by these companies,” Paulo Carvão, an ex-IBM executive and research fellow at Harvard’s Kennedy School, said. “But the amount of investments and infrastructure far outweigh revenue-generating capabilities and there is no clear path towards these lines crossing.”
Finally, even if AI sales skyrocket, the major labs could still prove radically less profitable than markets are wagering. To justify their trillion-dollar valuations, OpenAI and Anthropic don’t merely need other companies to spend vastly more on artificial intelligence; they also need them to pay a premium for frontier models, even as cheaper knock-offs become more and more adept. (Disclosure: Vox Media is one of several publishers that have signed partnership agreements with OpenAI. Our reporting remains editorially independent.)
Yet for most corporations, that might be a losing proposition. Today, the second-tier AI systems aren’t just a little cheaper than the frontier ones. The Chinese open-weight model Kimi K3.1 Pro costs $10 per million tokens of output; for OpenAI’s most advanced model, GPT-6 Astra, that figure is $50.
Perhaps, this premium is justified, given the current divergence between each product’s abilities. But that gap might become less relevant as AI becomes more powerful.
After all, intelligence has diminishing returns in many fields of endeavor. To staff its accounting department, Taco Bell could hire a team of path-breaking theoretical mathematicians. But if those geniuses cost five times as much as your average CPA, it would probably be wiser for the company to go with the latter; you simply don’t need a Fields medalist to monitor gordita revenues. Likewise, once last year’s AI model is as capable as 10,000 geniuses, few companies may feel the need to pay extra for a bot that’s as skilled as 10,500.
Already, there is some evidence that businesses consider open-source models good enough for many purposes. In a September survey of 200 enterprises, the market analysis firm Futurum Research found that 53 percent had partially substituted open-weight models for frontier ones.
“For the financial modeling that I do — which isn’t rocket science but not fourth-grade math either — Claude is already so capable, I can’t really imagine what more the next 6.0 or 7.0 would offer me,” Spectra’s Donnelly told me. “There was an all-you-can-eat buffet approach initially where token budgets weren’t really a huge consideration. But now that it’s not free, people are looking for cheaper options.”
The autumn of the AI industry’s discontent
These basic concerns have longshadowed the AI boom. Every time investors started getting antsy, however, a new technical breakthrough or eye-popping earnings report seemed to quell their fears.
But concerns about a bubble intensified in recent weeks, for at least three reasons:
1) The “central bank” of the AI sector is in crisis. The chipmaker Nvidia has arguably been the single biggest beneficiary of the AI boom, with its market valuation rising by trillions of dollars since the release of ChatGPT. Partly for this reason, it has also functioned as the tech industry’s mini-Federal Reserve, acting as a funder and financial backstop for AI investment.
As of this August, Nvidia had invested $70 billion in AI companies, facilitating the growth of OpenAI and Anthropic. It separately extended roughly $230 billion in financial guarantees to support the data center expansion.
To see how the latter work, consider the cloud infrastructure company CoreWeave. That firm is viewed as a bigger credit risk than many other hyperscalers, since it isn’t a diversified tech giant like Google or Amazon. To prevent lenders’ fears from styming CoreWeave’s growth, Nvidia promised to buy any data center capacity that the firm failed to sell, up to $6.3 billion.
Nvidia did none of this out of charity. By facilitating the growth of AI labs and hyperscalers, the company increased demand for its own products. The aim was to foster a self-reinforcing cycle of spending: Nvidia puts money into AI companies, those companies buy its chips, the resulting revenue raises Nvidia’s valuation, which allows it to put more money into AI companies.
This plan worked beautifully — until last week, when Nvidia’s longtime leader Jensen Huang abruptly resigned amid revelations of two extramarital affairs, one with an employee, the other with an alleged Chinese intelligence asset.
The latter relationship proved especially damaging for Nvidia. The company had long lobbied the White House to loosen restrictions on the sale of advanced AI chips to China, eventually persuading the Trump administration to relax its export controls. The suspected Chinese spy in question, Alice Huang, was an executive at Megaspeed, a Singapore-based cloud computing firm that was itself suspected of facilitating the illicit transfer of cutting-edge American semiconductors to China.
The Nvidia CEO initially tried to weather the resulting storm. But when a second affair — with another Nvidia executive — came to light, Huang stepped down.
All this has left Nvidia both less capable and willing to supply AI companies with cheap financing. Nvidia’s market capitalization has fallen by nearly 40 percent since Huang’s extramarital activities first came to light. Investors fear that the company will be incapable of adequately replacing its 33-year CEO, who had exercised unusually wide-ranging control over its operations. And the scandal threatens to restrict Nvidia’s access to the Chinese market, as the federal government cracks down on the use of US chips by Chinese firms.
Nvidia’s falling valuation has left it with less spare cash for funding other AI ventures, and Huang’s departure has also reportedly sapped its interest in circular financing schemes. According to the Wall Street Journal, Nvidia’s new interim CEO, longtime finance chief Colette Kress, had privately pushed back against Huang’s increasingly ambitious investments in AI startups and data-center developers. At her behest, Nvidia halved a proposed $250 billion financial guarantee for an OpenAI development. With Cress at the helm, the AI industry may find funding harder to come by.
“This is a classic kind of black swan event,” Ryan Cummings, a former staff economist for the White House Council of Economic Advisers, told me. “A CEO has an affair with a potential Chinese spy and then all of a sudden people notice there were a lot of cockroaches underneath the floorboards.”
2) A blue tsunami just inundated Silicon Valley. By early this year, the unpopularity of the data center boom had become clear, registering in polls and local opposition to developments.
Still, November’s midterm elections revealed the extraordinary scale of Big Tech’s political problems. In the 41 jurisdictions where data center development was on the ballot, the side favoring restrictions prevailed in 40. Meanwhile, Democratic governors came to power in data center hotspots like Texas, Georgia, and Ohio, after campaigning against such projects.
This poses two threats to the AI boom. First (and most obviously), it threatens to constrain the compute buildout that powers AI models and enriches hyperscalers and chipmakers. Newly elected Democratic governors like Texas’s Gina Hinojosa and Ohio’s Amy Acton appear unlikely to block projects already underway or permanently ban all data center development. Nevertheless, they will surely impose costly requirements on new development, such as obligations to cover all electricity costs, provide community benefits, or use union labor. And across the country, every other governor with a self-preservation instinct is liable to follow suit.
As a result, the AI buildout is poised to become slower and more expensive. This is not terrible news for all of the sector’s players (if the supply of compute gets constrained, the hyperscalers who own existing data centers will be able to charge higher rents). But the top AI labs’ already gargantuan computing costs will marginally rise and growth across the sector will marginally slow.
Second, the fact that Democrats won a landslide congressional victory after campaigning against data centers and Big Tech raises the odds of federal AI regulations (if not in 2027, then in 2029). Congress’s new Democratic majorities plan to push legislation requiring data center developers to assume full responsibility for their impacts on electricity costs, mandating independent testing of advanced AI models, and expanding AI companies’ liability for harm committed with their products.
And if AI’s capabilities rapidly accelerate — as OpenAI and Anthropic’s revenue estimates assume — political pressure to regulate (or even suppress) their models could build. Fears of AI’s impacts on cybersecurity reached a fever pitch in late October, after advanced models uncovered previously unknown flaws in one of the internet’s core programming languages.
Barring an all-out ban on new models, AI safety regulations will not make or break the boom. But they would likely leave the sector slightly less profitable.
For these reasons, among others, tech stocks fell sharply after Election Day.
3) Credit is getting more expensive. Inflation came in unexpectedly high last month. Since the Federal Reserve raises interest rates in response to high prices, October’s CPI data pushed up borrowing costs across the economy.
That’s bad news for the AI buildout, since the boom increasingly runs on credit — roughly one-third of hyperscalers’ capital expenditures are debt-financed. Thus, rising interest rates means higher costs and lower profits.
What’s more, high interest rates put downward pressure on AI companies’ valuations. The reason for this is simple: The more that investors can earn from safe assets like US Treasury bonds, the less eager they become to buy a claim on some risky venture’s hypothetical future profits. Thus, high rates tend to depress stock prices across the board. But since AI companies’ future mega-profits are especially speculative, rising borrowing costs weigh heavily on their valuations.
The AI boom might survive — but it will never be the same
None of this means that the AI sector is certain to melt down. There’s still cause for doubting that AI is in a bubble analogous to the ill-fated dot-com boom. At the peak of the latter, ascendant tech companies like Yahoo boasted stock prices equal to nearly 2,000 times their earnings. Today, Nvidia’s price-to-earnings ratio is roughly 20.
Further, the dot-com bubble was fueled in part by fledgling companies with negative earnings and scant revenues. By contrast, many of today’s biggest AI players are established mega-corporations with diverse, profitable businesses, such as Amazon and Google, or chip manufacturers like Nvidia who sell an actual physical product to a range of profitable industries.
“Hyperscalers are omnivores that have multiple sources of food,” John-David Lovelock, chief forecaster at the global advisory firm Gartner, said. “They’re probably going to be fine.”
Meanwhile, AI labs continue to make technological progress at a blistering pace. According to a recent assessment from the AI safety nonprofit METR, Anthropic is now improving its unreleased models at twice its pre-November 2025 rate, thanks largely to its own AI tools’ growing ability to make better AI tools. This development lends greater credibility to the notion that artificial intelligence will eventually achieve “recursive self-improvement,” a feedback loop in which every jump in AI’s capabilities accelerates the arrival of the next, as superintelligent models automate the process of building even more superintelligent models.
Precisely how fast AI models will improve in the near term — and how profitable their newfound capacities will be — remains uncertain. But it’s at least possible that next year’s state-of-the-art models could be radically more useful to businesses than today’s.
Regardless, it’s clear that the AI boom has entered a new era. The days when AI giants enjoyed governments’ reflexive deference — and markets’ unqualified exuberance — are over. If labs fail to contain their models’ ever-growing capacity to cause harm, they will be hearing from Congress. If hyperscalers fail to muster revenue commensurate with their capital spending, they will bleed value. AI’s great crash may or may not be beginning. But its honeymoon has ended.
If you have, you might have noticed that some or all of those services didn’t work as expected. First, Microsoft’s cloud programs, such as Outlook, went down late last week. Then, all of the planes were grounded on Friday. And this week, we had to deal with the bank outages, a separate United Airlines-specific outage, the Apple outage, and the NSA communications outage.
As of this writing, most of these issues have now been ironed out, at least for the time being, and there’s no indication that any of them were cyberattacks or otherwise malicious. But whatever the reasons, it sure feels like there have been a lot more outages this week and last than in a usual seven-day stretch.
To get to the bottom of it — hopefully before something else shorts out — the Today, Explained team spoke with Lily Hay Newman, a senior writer at WIRED who covers hacking, information security, and digital privacy.
She spoke with Today, Explained co-host Sean Rameswaram about why it took Microsoft so long to get Outlook back on its feet, all of the other recent outages we’ve suffered, and whether it’s as unprecedented as it feels in the moment. Plus, she explains how AI is causing a broader “vulnepocalypse” (or “Vulnageddon”), even if it’s not specifically the culprit of all of our current outages.
Below is an excerpt of Newman’s conversation with Sean, edited for length and clarity. There’s much more in the full podcast, so listen to Today, Explained wherever you get your podcasts, including Apple Podcasts, Pandora, and Spotify.
Lily, we spent the first half of the show talking about the issues with our air traffic control systems, but last week, of course, also brought chaos to a lot of people’s inboxes when Outlook was not so good. Tell us what happened there.
Microsoft 365, so all of the productivity platforms, had a massive outage and it dragged on for a full day, which is very noticeable when it’s a set of products that are as integrated in everything, as 365 is.
That means you just can’t work, right? This was probably billions of dollars of productivity lost.
Yeah, this is the cloud version of Microsoft Office: Outlook, Word, Excel, Copilot, Teams. If that’s your productivity platform, then yeah, you can’t work. That’s everything.
Microsoft is so enmeshed in the tech ecosystem, even before all of the cloud platforms, because they’re the Windows maker. They’re in everything.
What happened?
Microsoft, a few days later, on Monday, what they were saying was that they had sent out an update. It’s a Windows update, but it’s only really for their infrastructure. It’s not an update in consumer Windows. It’s not something your laptop would’ve gotten, but it has been out and proliferating through their infrastructure for about three months. But it actually had a flaw in it that wasn’t caught by their testing.
At the same time, they say cloud infrastructure is all about managing traffic and managing different regions of data centers. The issue with that three months of proliferation of that update started triggering and that’s why it was such an intractable issue.
At first, I saw it and I thought, okay, it’ll be a few hours, three hours or something, and they’ll fix it. And then, waking up Friday morning, it’s still not fixed. And it’s like, whoa, so something’s really going on.
And so it was the interplay of multiple factors that made it so hard for them to recover, they say.
Okay. Very different situations here between air traffic control and Microsoft 365. However, both kind of have a similar origin story in that they’re about patching systems and upgrading systems.
True, yes.
We’ve heard from Sean Duffy, Transportation secretary, that he cannot guarantee what happened on Friday won’t happen again in the near distant future.
Very ominous.
Have we heard something more reassuring from Microsoft?
Microsoft would definitely say, “This isn’t going to happen again. We want to give you all assurances. We are so sorry. Sometimes problems arise because of cyber attacks or exploitation of vulnerabilities.”
And then it’s kind of like, ‘Well, why weren’t all those systems patched? There was a patch out.’
But a lot of times patching is slower than we would like because the ramifications of a bad patch or an unforeseen consequence of an update can be really severe, and so organizations tend to be very nervous.
But it’s never the last time. These types of outages and failures happen in all technology systems and they will continue to, but people try to learn as much as they can from each incident and make things more and more robust.
Maybe a good time to ask you about all of the other breakdowns we’ve been having in our technology and elsewhere over the last few weeks.
I mean, people may not have even realized they just lived through something called Black October. What is Black October, beyond Halloween?
In October there were a few things going on, including with the widely used web scripting language, PHP. There were these two vulnerabilities in it that researchers were really concerned that if they were exploited, they could cause remote access takeovers of all different websites and all different platforms.
And so there was this big, industry-wide effort to make sure that everybody was patching — a “We have to patch this right away, don’t wait” type of thing.
So there’s Black October and then there’s the “vulnepocalypse.” People have been talking about this for maybe about a year now, this idea—
Everything’s vulnerable?
Right. Or that AI tools—
We need a better name for that.
Look, it’s not my thing. I’m just doing air quotes here, heavy air quotes. But this idea that agentic AI and AI models are becoming more and more powerful and adept at finding bugs in software and in helping both researchers and, potentially, attackers figure out how to exploit those vulnerabilities and weaponize them.
There’s this [idea] that cybersecurity is always this cat and mouse — again, I’m doing the air quotes — “cat and mouse game,” or an “arms race” is another phrase that’s often used. And so that in itself is not new, but the idea that AI in the mix just really accelerates it for both sides, and how to manage that dynamic.
Then there was this Iranian cyber attack on supermarket fridges or Binance getting hacked for, I don’t know, I think the most money I’ve ever heard about in any kind of hacker heist ever — $6 billion?
It was. Yeah.
I thought that we were meant to believe that crypto was unhackable, and here we are getting hacked for $6 billion?
I just want everyone to know you can rest easy. Things were already really bad before.
The fridge hacks, for example, the vulnerabilities that are being exploited there actually kind of go back numerous years. French officials have pointed to Iranian actors as having been behind at least some of those. This is targeting the food supply chain.
And same with the Binance situation. North Korean hackers stealing cryptocurrency, including large sums of cryptocurrency, is a story that reporters have been following for years and years and years.
And then on Monday, we had this bank outage across banks across the United States — Bank of America, Chase, Wells Fargo.
It just feels like everything’s breaking at a time where we’re on the precipice of driverless cars and AI running everything. Is it getting worse, or is it just a rough few months here?
To me, this stuff feels more cyclical. People may remember eight to 10 years ago when there were a ton of DDoS attacks taking down huge services. There was even basically an East Coast internet outage.
For me, it feels like we’re in another round of those, one of those moments where a lot of stuff starts to pile up and incident responders are really overworked and really exhausted and then other things are happening and nobody has as much elasticity or bandwidth because there’s so many things piling up at once — especially right now, in this new era of ubiquitous AI and access to more and more capable AI systems.
Again, it’s not to diminish that feeling. It’s just maybe helpful to have some context that it’s not truly, truly unprecedented right now.
Ireland’s privacy regulator has forced X to stop feeding AI models with data collected by its agents, the authority told POLITICO.
The social media giant owned by Elon Musk has suspended the use of so-called third-party data collected by its Grok Bot AI agent, which started mining this data in late October to train its AI systems.
The Irish Data Protection Commission (DPC) raised “significant concerns” about that data processing in recent weeks, Commissioner Dale Sunderland told POLITICO in an interview. “We asked them to suspend the processing, and they’ve confirmed to us today that they’ve done that,” he said Monday.
X did not respond to a request for comment.
The Irish privacy authority is the most powerful in Europe, as it enforces Europe’s data protection rulebook on companies that have their regional headquarters in Ireland, including Apple, Google, Meta, X and OpenAI.
The X action is a test case as privacy regulators shift their focus from AI training that uses data people have shared about themselves, to third-party data that personal AI agents are accessing and gathering on other people.
Third-party data is personal data about an unrelated person shared by two other parties — for example, holiday pictures shared with a friend who uses an AI agent. It also includes personal data that an AI agent scrapes from the internet.
Sunderland — who is one of three co-commissioners at the Irish authority — said they are also engaging with another, unnamed tech firm about its plans to use data collected by AI agents for training.
“It’s a very live issue at the moment,” Sunderland said, declining to name the company because the discussions are taking place ahead of any actual data processing.
The regulator doesn’t have an issue with AI agents “per se,” Sunderland said, but authorities have to scrutinize whether companies are still operating legally if they get into “further taking of that data and using it to further train the models.”
Sunderland said the Irish DPC will have “further engagement” with X about “what exactly happened, and what that means for next steps.”
At Davos earlier this year, Canadian Prime Minister Mark Carney offered a compelling response to a world increasingly shaped by great-power rivalry. Middle powers need greater strategic autonomy, he argued, but they cannot secure it by retreating behind their borders.
Instead, they need to build their own capabilities while working more closely with like-minded partners. On technology, he reasoned that countries should cooperate so they are not ultimately forced to choose between “hegemons and hyperscalers.”
At the recent EU-Canada Summit, that philosophy began to translate into policy. Canada and the EU launched an “Alliance for the Future” based on closer economic, technological and strategic cooperation. They also concluded negotiations on a Digital Trade Agreement designed to support a “secure, open and trusted digital environment” resilient economic relations, and trusted cross-border data flows.
That principle matters not only for Europe’s new relationship with Canada, but for its approach to other trusted partners.
Sovereignty without isolation
Like Canada, Europe has diagnosed its technology problem correctly. It depends too heavily on a handful of providers for the cloud, computing power, digital infrastructure and artificial intelligence (AI) models underpinning its economy and security.
Reducing those dependencies will require Europe to put its money where its mouth is, including by unlocking more diversified private investment. Mario Draghi estimated that the EU needs an additional €750 billion to €800 billion of investment each year until 2030. The European Commission, meanwhile, wants to mobilize €200 billion for AI and at least triple Europe’s data center capacity within five to seven years.
Europe cannot expect investment on an unprecedented scale while narrowing the field of companies able to provide it.
But Europe’s approach risks moving in the opposite direction. It could exclude some of the businesses investing most heavily in the continent and building European capabilities, while favoring others that remain dependent on non-European technology.
That is the risk confronting policymakers as they shape the Cloud and AI Development Act. The legislation seeks to influence not only how Europe expands and secures its technological capabilities, but also which companies can take part.
Herein lies the core trade-off. Europe cannot expect investment on an unprecedented scale while narrowing the field of companies able to provide it. Whether it closes its technology gap will depend partly on its ability to draw on the capital and expertise of experienced, trusted partners.
Look beyond ownership
The debate over digital sovereignty therefore needs to move beyond the simplistic question of ownership or headquarters.
A definition based principally on corporate nationality could exclude trusted companies with decades of investment, infrastructure and employment across Europe. It could also protect nominally European suppliers without necessarily making Europe more secure or resilient.
Effective sovereignty should mean that Europe can keep essential services running during a crisis, protect critical systems and retain a choice of secure, trusted suppliers. Assessments should therefore examine who controls data, encryption, operations, security and continuity.
Corporate nationality is relevant, but it cannot answer these questions on its own. A cloud provider may be EU-owned yet rely on processors, software and security tools developed elsewhere. Another may be headquartered outside the EU but operate European data centers, employ European engineers, place data and encryption keys under European control, and comply with demanding European regulation. Any nationality test alone tells policymakers little about which arrangement provides greater sovereignty.
In practice, sovereignty done right depends on who can access critical systems, where operational decisions are taken, whether data and encryption remain protected and, ultimately, whether services continue when normal arrangements fail.
Nationality should carry the greatest weight where a provider is controlled from a country that presents a recognized security threat, or where the relevant jurisdiction has the capacity and record of exploiting technological dependencies. The strictest restrictions are justified where foreign interference could threaten national security, including defense, law enforcement and the most essential public services.
Countries such as Switzerland, Canada and the UK should not automatically fall into the same category as recognized security threats. Where the security conditions do not apply, nationality-based restrictions risk becoming protectionist rather than protective.
Instead, Europe should apply three tests for trusted partners and domestic firms alike: whether operational control rests in Europe; whether the service can continue during a crisis; and whether the provider makes a substantial, long-term contribution to Europe’s technological capacity.
In practice, sovereignty done right depends on who can access critical systems, where operational decisions are taken, whether data and encryption remain protected and, ultimately, whether services continue when normal arrangements fail.
Telecoms is a case in point. Regardless of where they are headquartered, operators must keep essential services running through cyberattacks, power failures and supply-chain disruption while protecting the confidentiality of EU data.
A company’s registered address answers none of those operational questions.
Build with trusted partners
This matters beyond the Cloud and AI Development Act. The approach Europe takes to sovereignty could shape policy toward telecommunications, satellites, semiconductors and other strategic technologies.
Rules that exclude trusted foreign investors primarily on the basis of nationality risk fragmenting the market and deterring some of the capital and expertise Europe is trying to attract. They could also encourage reciprocal restrictions on European companies operating abroad.
The better approach is the one Carney has begun to articulate: build domestic capability while deepening cooperation with countries that share your security interests and can contribute technology, infrastructure and investment.
The EU should therefore provide a route for British companies that meet demanding security, control and resilience requirements to receive equivalent treatment as trusted providers.
The EU’s developing relationship with Canada shows that this is possible. Its new Alliance for the Future explicitly combines closer economic integration with technological cooperation, security and resilience, while the Digital Trade Agreement seeks to remove unjustified barriers while maintaining strong safeguards.
If deeper integration with Canada can strengthen European sovereignty, the same principle should apply to other trusted partners that meet equally demanding standards.
The obvious next test is the UK.
Britain sits outside the EU’s institutions but remains closely connected to Europe’s economy, technology sector and security. Infrastructure, investment and supply chains cross the Channel, while the UK and EU face many of the same cyber and geopolitical threats.
The EU should therefore provide a route for British companies that meet demanding security, control and resilience requirements to receive equivalent treatment as trusted providers, with reciprocal arrangements for qualifying European companies in the UK.
Strategic autonomy does not require strategic isolation.
The second UK-EU Summit on 20 November provides an opportunity to turn that principle into practice.
Leaders should establish a UK-EU Digital Partnership, covering the mutual recognition of trusted providers, coordinating rules affecting cloud and AI infrastructure, and the removal of unnecessary procurement barriers before they become embedded.
Such a partnership would support both sides’ sovereignty objectives by distinguishing genuine strategic dependencies from relationships that add to their resilience and security.
Europe is right to seek greater control over its technological future. But Carney has identified the crucial distinction: strategic autonomy does not require strategic isolation.
Canada and the EU have begun to show what trusted partnership can look like. The UK should be next.
Disclaimer
POLITICAL ADVERTISEMENT
The sponsor is Vodafone Group Services Limited.
The political advertising is related to the EU’s Cloud and AI Development Act and the development of EU-UK digital cooperation.
President Donald Trump departs the White House on November 16, 2026, ahead of a trip to China and the Philippines. | Kent Nishimura/AFP via Getty Images
This story appeared in The Logoff, a daily newsletter that helps you stay informed about the Trump administration without letting political news take over your life. Subscribe here.
Welcome to The Logoff: The Trump administration is closing a global AI chip loophole that the Jensen Huang scandal put in the spotlight.
What’s changing? The Commerce Department is trying to make it harder for China to access top-of-the-line AI microchips like those made by Nvidia.
Currently, China is able to dodge export rules by renting computing power from third-country data centers, which can legally buy American-made chips that China can’t.
But under the rule issued Monday, which is set to take effect in 90 days, data centers in a handful of countries will face new export restrictions on chips — including a requirement that they perform “robust due diligence on their customers” to make sure that they aren’t evading an export ban.
What’s the context? Huang, who resigned from Nvidia last week, was brought down in part by his relationship with Alice Huang (no relation), an alleged Chinese spy (kind of) and an executive at a Singapore-based tech company, Megaspeed, that was renting compute to Chinese firms — exactly the kind of thing that the new rule aims to crack down on.
What’s the big picture? The Nvidia scandal is front and center. But under the hood, there’s also some Trump administration infighting going on. As a statement announcing the new rule notes, it was written not by the Commerce Department, which is issuing it, but by the Treasury Department and the National Security Council.
That’s conspicuous because Treasury Secretary Scott Bessent has done his best to throw Commerce Secretary Howard Lutnick under the bus for the scandal, saying that Lutnick’s department failed to restrict Megaspeed’s access to advanced AI chips sooner.
Where’s President Donald Trump in all this? As of Monday, off to China and the Philippines for a multiday trip that will feature a trilateral summit with Chinese President Xi Jinping and Russian President Vladimir Putin.
For now, though, Trump is standing by Lutnick — and even by Jensen Huang, saying Monday that he “would certainly not mind” seeing him return as Nvidia CEO (the Nvidia board has already shut down the possibility).
And with that, it’s time to log off…
One of my colleagues described this story about Anny Aviram, a now-retired conservator at New York’s Museum of Modern Art, as “a true balm for the weary soul,” and it lives up to the billing. You can enjoy it here with a gift link.
As always, thanks for reading, have a great evening, and we’ll see you tomorrow!
What happened on Friday was only the latest symptom of an aging air traffic control system. | Adam Gray/Getty Images
On Friday evening, in the midst of the second-longest closure ever of US airspace, I bought two plane tickets and found myself contemplating, possibly for the first time, the list of hidden fees tacked onto my fare.
I had forked over $16.95 for a US transportation tax, plus $5.30 for a flight segment tax, both of which pool into the Airport and Airway Trust Fund, which will ultimately contribute to the Federal Aviation Administration’s work, thereby ensuring that my flight from New York to Florida lands smoothly.
But though one might expect my $22.25 — enough to pay an air traffic trainee’s salary for about an hour — would go directly into the FAA’s coffers, it does not. In fact, the FAA needs to get congressional approval every time it taps into its trust fund, a political bottleneck that has for decades kept the FAA from making the critical long-term modernization projects it needs to make sure that a single botched cable job is not enough to bring down the busiest airspace in the world.
In the days since all air travel shut down across the United States, we’ve learned a great deal about what happened and why. But the best response to the crisis — and the path to preventing it from happening again (and again and again) — may well be found in shortening the path that my $22.25 takes on its way to the beleaguered agency that needs it so badly. It’s a solution that transformed air travel in Canada — which turned its air traffic control system around three decades ago — and it might yet help guarantee on-time departures here in the United States, too.
In the aftermath of the groundstop on Friday, college football teams delayed their kickoff times and Microsoft — whose Outlook outage suddenly became only the second-most disruptive telecommunications issue last week — issued an inadvertently cheeky disclaimer saying “it wasn’t us.” The band Hoobastank performed an impromptu rendition of their 2004 hit “The Reason” aboard a Southwest Airlines flight, before the actual reason for the delay was even clear. But while rumors have swirled over whether the disruption could be the work of a rogue AI cyberattack, the real reason why the FAA closed America’s airspace for only the third time since 9/11 now appears to have been embarrassingly mundane.
“They were putting in new equipment and new technology,” said Markwayne Mullin, secretary of the Department of Homeland Security, in a press briefing on Friday, “and that new equipment was overwhelming the older system.”
There is, to be sure, plenty of new equipment to be put in, and not just at the tiny, two-runway airport in El Dorado, Arkansas, where the trouble began. The physical bones of our air traffic control system were installed back when many of Hoobastank’s Gen X fans were still in diapers, and the bulk of this vital system — which tells pilots when and where it’s safe to land — has not been upgraded since the 1980s. Most control towers still track airplanes using thin strips of cardstock, while 3.5-inch floppy disks remain the medium of choice for uploading weather forecasts. And copper wires, instead of more modern fiber optic cables, still transmit large swaths of data between air traffic control facilities.
In 2024, the Government Accountability Office found indications that 76 percent of the FAA’s systems were or could be unsustainable because of factors like aging equipment. Just a few months ago, a single severed cable grounded flights at some of the nation’s busiest airports. And for years, air traffic controllers have been sounding the alarm about chronic understaffing, unsustainable workloads, and an accumulation of safety complaints.
What happened on Friday was only the latest symptom of an aging air traffic control system, one that has suffered from deferred maintenance for so long that even trying to fix things can make the whole network go haywire. The Trump administration, to its credit, has made upgrading this decades-old system a priority of its transportation agenda, spurred in part by a tragedy over the Potomac last year in which an Army helicopter collided with a commercial airliner, killing 67 people.
But making sure that this fiasco doesn’t happen again will require more than just plying the FAA with another $50 billion in cash, per Transportation Secretary Sean Duffy’s pleas to Congress. It might require doing away entirely with the political wrangling that keeps the FAA thirsting for your airline fees, and giving it easier access to those hidden fees tacked on to your boarding pass and mine.
“Modernization in the US is a political battle,” John Gradek, an aviation expert at McGill University, told me. But in other countries like Canada, air traffic control “modernization is very much a functional battle,” made possible by predictable “funding it can use to modernize at its own pace, rather than fighting constant battles for congressional appropriations.”
Even $50 billion may not be enough to fix what’s wrong with the FAA, not least of all because a single lump sum — no matter how gigantic — won’t address the funding bottleneck that makes it hard to keep things running day in and day out. “They’re throwing a number at it, and there’s no detail behind the number,” he said. “What’s the architecture of the solution?” In fact, there is a better way to fix our airspace, and our neighbors to the north might be able to chart the course.
How Canada fixed its air traffic control system
In the 1990s, Canada’s air traffic control system was bogged down by many of the same problems that the American system still suffers from now.
Audits at the time found that chronic staffing shortages and unreasonable expectations of overtime were playing a serious role in air traffic incidents such as dangerously close calls between passenger jets. Meanwhile, Canada’s transportation department, Transport Canada, had gone seriously over budget and behind schedule while trying to modernize the country’s radar system, par for the course of a system under strain.
At the time, as in the US, a fee tacked onto flights went into a fund under the control of the Canadian government, meaning that Transport Canada often found itself begging Parliament for more cash. But this changed in 1996, after a broad coalition of fed-up pilots, airlines, and air traffic controllers rallied behind a new governance structure for Canada’s air traffic control system. Rather than being housed within Transport Canada, the new system would fall under a newly created independent nonprofit called NAV Canada.
NAV is funded directly through passenger fees without parliamentary approval, and since it’s not a government entity, it’s exempted from civil service requirements, rigid procurement rules, and other barriers that dogged the old system. In the ’80s and ’90s, Transport had spent $18 million Canadian dollars and six years to modernize its radar system. When NAV undertook a comparable extension in 1997, it completed the project in about six months and at a cost of $6 million.
“That whole process of modernizing and bringing new technology to bear on air traffic control,” said Gradek, “has paid out handsomely to the Canadian air traffic control system.”
Douglas Mein, a senior director of Transport Canada during the transition, later testified to the British Parliament that moving the air traffic control system to an independent nonprofit freed it from the red tape that often prevented necessary hiring and tied up technological investments for years on end. A few years after NAV took over in 1996, Canada’s air traffic controllers started retiring the paper strips they once used to track flights in favor of an electronic system, an upgrade that the US only started rolling out in 2022.
Over the decades, Canada’s air traffic control system seems to have become more resilient, while the US has just gotten more vulnerable to disruption as its technology continues to molder. American policymakers have repeatedly pointed to Canada’s success in arguing for a restructuring of America’s air traffic control system. Most recently, under the first Trump administration, Republicans pushed for a Canada-style restructuring that floundered in Congress, seemingly under pressure from the private aviation industry, which was concerned that such an arrangement would prioritize the interests of the commercial airline industry.
Others have pointed to fears that such a system would lead to a loss of public accountability. Some also note that Canada, while seemingly still better off than the US, has run into its own renewed staffing shortages and other air traffic control issues in recent years. But fixing America’s air traffic control system might not necessarily require this kind of privatization. In The Atlantic last year, Dorothy Robyn, who oversaw aviation policy under the Clinton administration — which also pushed for a similar spinoff for the FAA — argued that air traffic control ought to be spun off into a standalone entity, possibly one that would still be governmental. This would, she suggests, contribute to efficiency, while also ensuring the independence of safety regulators, who are currently housed within the FAA itself.
Since 1995, Robyn wrote, more than 60 countries have separated their air traffic control system from regulators, some through standalone nonprofits but others through government entities. For example, Germany and New Zealand both migrated their air traffic control systems away from their transportation departments and into government-run entities that are still publicly owned, but have much more freedom to hire workers and spend passenger fees on improvements than the FAA does.
Even if the US is not destined for Canadian-style independent air traffic control, a reform that allows the FAA to tap into the fees we all pay on our flights without having to grovel to Congress could help the institution better plan and invest in its technology and staff in the long run. A more fluid and regularized funding scheme would hardly fix everything overnight, but it would at least present a vision for the future of the nation’s air traffic control system that neither requires intermittent multibillion-dollar checks nor panic-inducing admonition that without literally any funding right now, the system that keeps Americans safe in the sky is “going to continue to fail,” as Duffy proclaimed on Friday.
“There is a way for the system to be able to, in fact, build in redundancy, build in resilience, and build in backups,” said Gradek, but to do so, “air traffic control needs constant funding, needs constant handholding, and you cannot have these battles to figure out where the next dollar is going to come from.”
LONDON —The U.K. government has confirmed it plans to end Palantir’s contract to provide a data platform for the National Health Service in 2029.
However it is not severing the contract immediately, as some MPs and campaigners had pushed it to do.
The government will begin formal procurement for a “new” contract in the second half of 2027, with the process expected to take 18 months, health minister James Frith told MPs on Monday. He described the extension of Palantir’s contract until 2029 as “a bridge, not the destination.”
In a bid to regain “public trust” and open up opportunities for “U.K. sovereign capabilities,” Frith said “interoperability, resilience and trusted common standards” will be “baked into” the new procurement round, though non-U.K. firms will also be encouraged to bid.
Palantir’s contract has attracted considerable scrutiny from campaigners, health professionals and MPs over controversial views expressed by the company’s founders and concerns about vendor lock-in.
Two parliamentary committees have called for its NHS contract to be ended in the last year.
“We want to give the health and care system more choice, greater resilience and less dependence on any single supplier, as well as the ability to adapt and evolve as new technologies emerge in future, with better value for taxpayers through time,” Frith said.
“These objectives require a new approach — an approach that will stimulate a more diverse and innovative supplier market, including greater use of sovereign AI as well as internationally developed capabilities,” he added.
This approach will encompass investment in strategic technologies, changing how novel technologies are procured, and developing open standards, Frith said.
“We must avoid over-dependence and our future capability should avoid being locked into any single technology,” he said. He didn’t say whether Palantir will be able to bid on the new contract in 2029.
A Palantir spokesperson said the announcement was “good news for patients and for the NHS” and it was “proud” of its work.
“We will now redouble our efforts over the next two years to continue and accelerate that support,” as well as “deepen our investment in the U.K.,” they said.
Russian-speaking hackers built an AI agent that carried out what may be the first-ever fully-automated cyberattack, Google reported Monday.
The tech giant’s cyber threat intelligence division said it had spotted hackers using artificial intelligence to “fully automate vulnerability discovery and exploitation without a human in the loop for the first time.”
The development could mark an inflection point in the use of AI for cyberattacks, which has dominated headlines and worried policymakers since AI firm Anthropic released its specialized model Mythos in April.
Hackers, who used Russian in technical documentation identified by Google, set up a series of AI agents to “continuously compare pre- and post-patch code, identify vulnerabilities, and automatically develop exploits without human intervention,” Google said.
That means the agents independently found software flaws and ways to use them. The attackers targeted French websites in the media, entertainment and education sectors, Google said.
Until now, even the most advanced AI models could not carry out cyberattacks entirely independently; they often required skilled hackers to guide them or fix mistakes. The key change instead was that AI models allow hackers to accelerate attacks and make them more targeted.
The development is likely to spook governments and critical infrastructure operators who are still reeling from this summer’s hack by an OpenAI agent into the Hugging Face platform. Though acting autonomously, that hack didn’t appear to be maliciously aimed at certain targets like the one being reported by Google. Fears are also running high after the recent discovery of major flaws in PHP, a programming language that underpins 70 percent of the web, and a series of hacks on supermarket fridges, which have been partially blamed on Iran.