← The State of Tech RSS Feed

20 november 2026 · 14:16

TSMC and Samsung halt advanced AI chip shipments to China

TSMC and Samsung have suspended shipments of advanced chips made at sixteen nanometres and below to Chinese clients, after updated US rules on AI chip export controls closed a manufacturing loophole. This episode of The State of Tech explains what the rules mean for Chinese AI chip designers and for European chipmakers and suppliers. We also cover Palo Alto Networks' $4.5 billion share deal for the AI startup Applied Compute.

TSMC Samsung AI chips export controls Nvidia Palo Alto Networks

Beluister deze aflevering:

Transcript

Samantha: Welcome to The State of Tech, The European Edition, Friday November twenty, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: TSMC and Samsung stop shipping advanced chips to Chinese AI designers, Palo Alto Networks buys AI startup Applied Compute for four and a half billion dollars, the founder of Muslim Girl sues xAI over Grok images, America's top cyber official steps down, and two things to try yourself: the free outage tracker Downdetector, and a ten-minute check prompted by Citibank's new AI agent terms. Let's start with the chips.

TSMC and Samsung halt advanced chip shipments to Chinese AI designers after new US rules.

Samantha: Taiwan Semiconductor Manufacturing Co., better known as TSMC, and Samsung Electronics have suspended shipments of advanced chips to Chinese clients. The cut-off is chips made at sixteen nanometres and below. That number describes how finely a chip is manufactured, and the smaller it gets, the more powerful and efficient the chip. The trigger is an update to export rules from the US Department of Commerce. It closes compliance loopholes that Chinese fabless designers had been using. Fabless means they design chips but pay factories like TSMC and Samsung to actually make them.

Bob: The practical result: Chinese AI chip developers are now limited to older, mature production processes at twenty-eight and forty nanometres, unless they get an explicit export licence from Washington. And it's not only about the chips themselves. The rules also block access to advanced packaging, the technology that bundles several chips into one powerful unit, and to high-bandwidth memory, the very fast memory stacked right next to an AI processor. According to the report, that severely hampers Beijing's push to build home-grown rivals to Nvidia's top AI processors.

Samantha: And this week, a very different set of numbers landed in Washington. Brian Mast, who chairs the House Foreign Affairs Committee, released confidential Commerce Department licensing data. It shows that between December 2025 and May 2026, regulators approved the export of nearly nine hundred thousand advanced Nvidia and AMD AI chips to Chinese companies. That's out of roughly three point four million requested. Approved buyers included Alibaba, Tencent, Baidu and carmaker BYD.

Bob: So there are two tracks running side by side. On one, the US is tightening the screws on Chinese firms trying to build their own top-end chips. On the other, the licensing data shows Washington let hundreds of thousands of powerful American AI chips go to Chinese tech giants. Lawmakers seized on that gap, questioning whether export controls are enforced effectively enough to protect national security and slow China's AI development. The data only runs to May, so it says nothing yet about licensing under the newest rules.

Samantha: For Europe, this lands close to home. European chipmakers and technology suppliers sit inside the same global supply chains as TSMC and Samsung, so every tightening of US rules means more compliance work: checking who a customer really is, and where a chip or component might end up. Re-export is the other worry. When hardware crosses borders several times, it gets harder to prove it doesn't reach a restricted market. The bottom line: the rules of the chip trade keep shifting, and European firms have to keep pace with decisions made in Washington.

Palo Alto Networks pays four and a half billion dollars in stock for AI startup Applied Compute.

Bob: Palo Alto Networks is buying an AI startup that's only eighteen months old. The company is Applied Compute, founded by former OpenAI researchers, and the price is four and a half billion dollars, paid entirely in shares. Applied Compute builds tools that let businesses take an open-weight AI model, meaning a model whose inner workings are published so anyone can download it, then train it further, tailor it to their own tasks, and run it day to day.

Samantha: That's a real shift for Palo Alto Networks. Until now its focus has been watching over systems and spotting threats. With this deal it moves into actually building, customising and serving secure AI models for companies. Applied Compute's co-founder and chief executive, Yash Patil, joins as Co-Chief AI Officer. The deal still needs regulatory approval and is expected to close in the company's fiscal year 2027. So nothing changes for customers overnight, but the direction is set.

Bob: The bigger picture: this security company doesn't just want to protect AI, it wants to be where businesses make their AI. That's a different line of work than Palo Alto Networks has been known for. Open-weight models are central here, because a company can run them on its own systems instead of sending sensitive data to an outside AI provider. And paying in stock rather than cash shows how much value is being placed on a team that's barely a year and a half old.

Samantha: For European companies, the appeal is in that control. A firm that tailors an open model in-house can keep closer track of where its data goes, which matters under EU data protection and security rules. Palo Alto Networks pitches the combination as secure AI for the enterprise. Whether that holds up will depend on how the tools work in practice once the deal closes. For now, the key takeaway is that a major security player is betting billions that companies want their own customised AI, not just off-the-shelf chatbots.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.

Muslim Girl founder sues xAI, saying Grok turned her selfie into sexualised fake images.

Bob: Author and activist Amani Al-Khatahtbeh, the founder of the media platform Muslim Girl, has filed a federal lawsuit against Elon Musk's xAI in California. She says the company's Grok AI let users create non-consensual sexualised images of her. According to the complaint, people on X asked Grok to alter a selfie she had posted publicly.

Samantha: The edits removed her hijab and showed her in bikinis and lingerie. Those images then went viral, with thousands of shares. The complaint accuses xAI of negligence, of a design defect, and of intentionally distributing non-consensual sexual content. She is seeking damages of up to two hundred and fifty thousand dollars for every unauthorised share, under California state law.

Bob: That per-share structure stands out. If a court were to accept it, the potential damages could climb very quickly, because the complaint speaks of thousands of shared images. The design defect claim matters too. It argues the problem isn't only individual users misbehaving, but how the tool itself was built. That moves the debate from who typed the prompt to what the company allowed its product to do.

Samantha: The complaint specifically describes the removal of her hijab alongside the sexualised edits, so the images touch on her religious identity as well as her body. These are allegations, and xAI will have its chance to respond in court. But the case shows how a single public selfie can be turned into material that spreads far beyond the control of the person in it.

Bob: For Europe, the question of responsibility is very live. The EU's AI Act also addresses deepfakes, so developers of AI image tools in Europe face legal duties of their own. A US ruling that holds a developer liable for what its tool generates could become a reference point for European lawyers and regulators, even though the legal systems differ. And it's a reminder that once an image is online, image generators can be pointed at it.

White House National Cyber Director Sean Cairncross steps down ahead of a new Congress.

Samantha: In Washington, the man in charge of America's cybersecurity strategy is stepping down. Sean Cairncross, the White House National Cyber Director, announced he's leaving the post. He was confirmed by the Senate in August 2025 and led the Office of the National Cyber Director, which coordinates federal cyber policy, through planning for the defence of critical infrastructure.

Bob: The timing is political. According to administration officials, he's leaving now so the White House can get his intended successor confirmed by the Senate before Democrats take control of Congress in January. That successor is Kirsten A. Davies, currently the Pentagon's Chief Information Officer, the person responsible for the Defense Department's IT. So it's a planned handover, not a sudden exit, at least as officials describe it.

Samantha: It does come at an awkward moment. The change follows a wave of widespread technical outages across US aviation, banking and government networks. Federal officials attributed those to software glitches, not to an attack. That's an important distinction. Not every big outage is hacking, but a software fault can still leave travellers stranded and bank customers locked out.

Bob: And that's why this job matters beyond Washington. The national cyber director has to pull agencies, companies and allies together when infrastructure wobbles. A changeover can bring new priorities, and the plans for protecting critical infrastructure will pass to someone coming from the Pentagon. For Europe, the US cyber office is an important counterpart on protecting shared systems and reporting incidents.

Samantha: Airlines, banks and cloud services don't stop at the border, so a problem in a US system can ripple into European flights or payments. That's why transatlantic coordination on incident reporting matters, and why a stable counterpart in Washington counts for European cybersecurity agencies. The bottom line: the hand-off is designed to be quick, but it lands right after a run of outages, and just before a new Congress arrives in January.

The free website Downdetector shows in seconds whether an outage is you or them.

Bob: To close out, two things you can actually use today, and the first follows straight on from those outages. When your banking app won't load or an airline site keeps crashing, the first question is: is it me, or is it them? A free website called Downdetector tries to answer exactly that. It collects reports from users having problems with a service and shows them as a simple graph, so you can see at a glance whether complaints are suddenly spiking.

Samantha: It covers services people rely on every day: banks, airlines, phone and internet providers, social networks, streaming and games. Open the site in any browser, type in the name of the service, and you'll see how many problem reports came in over the past day compared with what's normal. There are separate country versions, including for many European countries, and apps for phones too. It takes about thirty seconds, and you don't need an account to check.

Bob: The practical payoff is simple. If the graph is flat, the problem is probably on your side, so restarting your router or phone is a sensible first step. If there's a big spike, you can stop fiddling with settings and wait, or switch to a backup, like paying by card instead of through the app. Users can also leave comments describing what's broken, which can show whether it's logging in, payments, or the whole service.

Samantha: There are limits worth knowing. Downdetector is built on what users report, so a spike tells you something is wrong, but not why. It can't tell a software glitch from a cyber attack, and a busy news moment can drive extra reports. The site also shows advertising. Still, for that first "is it just me?" moment, it's a quick reality check that can save you twenty minutes of pointless restarting.

Citibank's new AI agent terms make customers liable for what their AI assistants do.

Bob: The second one is less an app and more a check you can do in ten minutes. Citibank, the third-largest commercial bank in the United States, has changed its online terms for customers who use outside AI agents. Those are AI assistants you give access to your bank account so they can manage your money or carry out payments. Under the new terms, the customer carries the full risk of mistakes, wrong instructions, or data leaking through a hack.

Samantha: The key sentence is a legal one. Any action started by an AI agent you've authorised is treated as if you authorised it yourself, directly. So if the assistant pays the wrong person, the bank sees it as your payment. Citi isn't alone here. According to the report, it follows similar moves by JPMorgan Chase and Bank of America, which makes it a pattern among big US banks rather than a one-off.

Bob: So here's the check, and it's free. If you're thinking about letting an AI assistant into your bank account, or you already have, open your bank's online terms and search for wording about third parties or automated access. Then look in your banking app or online banking for any list of connected apps or services, and remove anything you don't recognise or no longer use. Ten minutes, on any device.

Samantha: One limitation to keep in mind. These are Citibank's terms, and your own bank, in Europe or anywhere else, may handle this differently. That's exactly why reading your own bank's rules beats assuming. The wider shift is clear, though. As AI agents start handling money, banks are spelling out who pays when something goes wrong, and so far, the answer from these big US banks is the customer.

Samantha: Today we covered: TSMC and Samsung cutting off advanced chips to Chinese AI designers, Palo Alto Networks buying Applied Compute, the Grok lawsuit from the founder of Muslim Girl, America's cyber director stepping down, and two things to try yourself: Downdetector, and a check on your own bank prompted by Citibank's AI agent terms.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech, the tech world in 15 minutes.