← The State of Tech RSS Feed

21 augustus 2026 · 13:02

Nvidia's China AI Chip Plan & Sovereign AI Surge

Nvidia is preparing small-batch shipments of a brand-new China-specific AI chip built with licensed Groq technology, with ByteDance and Tencent already placing orders ahead of a year-end delivery target. In the same episode, a new CNAS report reveals that 184 sovereign AI projects have launched across 67 countries in the first half of 2026, raising pointed questions about where Europe fits in the race for AI self-reliance. Friday's episode of The State of Tech also covers Broadcom's $60-billion debt push for custom silicon, an urgent CISA cybersecurity alert, and two free open-source tools worth trying this weekend.

Nvidia Broadcom Sovereign AI Cybersecurity CISA Export Controls

Beluister deze aflevering:

Transcript

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

Bob: And I'm Bob Russell. Today: Nvidia's new AI chip plan for China, Broadcom hunting sixty billion dollars in debt for AI, a global surge in sovereign AI projects across sixty-seven countries, an urgent cybersecurity alert for vulnerabilities being exploited right now, and to close out, two things you can actually use today: the free note tool Trilium Notes, and the open-source home cooking planner Grocy. Let's start with Nvidia.

Nvidia plots a China comeback with a new AI chip designed to slip past US export rules.

Samantha: Nvidia is preparing small-batch shipments of a brand new AI chip built specifically for Chinese customers, with deliveries expected by the end of the year. This is not a simple rerun of the earlier watered-down chips. It's a fresh variant of Nvidia's language processing unit, developed using licensed technology from Groq, and it's designed to work alongside GPUs to speed up AI chatbots. ByteDance and Tencent have already placed orders, according to The Information.

Bob: And that detail about Groq is what makes this interesting. Nvidia is essentially building a hybrid product to stay relevant in a market it was largely locked out of. At the same time, China is easing limits on Nvidia's existing H200 chips, with larger quantities heading to Hong Kong. So both sides are quietly adjusting. Washington keeps the top-tier chips at home, Beijing lets a middle tier back in, and Nvidia gets to keep a foothold in the world's second-largest AI market instead of ceding it entirely to Huawei.

Samantha: The bigger picture is that export controls are turning into a permanent product design constraint, not a temporary hurdle. Chipmakers now build entire product lines around what's legal to ship where. That fragments the global market, and it slowly pushes Chinese buyers toward domestic alternatives whenever the compliant version falls short.

Bob: For Europe, this matters in two ways. First, if Nvidia diverts a chunk of its manufacturing capacity toward a China-specific chip, that adds pressure on the supply lines feeding European data centres. Cloud providers here are already competing with US hyperscalers for the same allocations. Second, European regulators are watching this dance closely, because whatever compromise emerges between Washington and Beijing sets a template. If chip design becomes a diplomatic tool, Brussels will want its own leverage, and the sovereign AI conversation we'll get to later becomes even more urgent. For now, though, the takeaway is that the chip war has quietly moved from confrontation to accommodation, and Nvidia is the company writing the new rulebook.

Broadcom seeks over sixty billion dollars in debt to fund the next wave of custom AI chips.

Samantha: Broadcom is in talks with lenders to raise more than sixty billion dollars, aimed at financing a major AI chip deal. The package could include a junior tranche of around thirty billion, plus a senior-secured tranche Broadcom would partially guarantee, potentially bringing the total to as much as one hundred billion dollars. Anthropic is reported to be one of the main beneficiaries, alongside other firms building custom chips.

Bob: The scale here is what stops you in your tracks. One hundred billion dollars in debt, for a single wave of chip investment. That's roughly the annual defence budget of a mid-sized European country. And it signals that even the biggest tech companies no longer want to pay for AI infrastructure out of cash flow. They're going to bond markets, they're pledging assets, they're doing what utilities and airlines used to do.

Samantha: It also tells you something about the custom chip trend. Anthropic, Google, Meta, they all want to reduce their dependence on Nvidia by designing in-house. Broadcom is the partner making that possible for many of them, and this financing round is essentially the industrial backbone of that shift. The bottom line is that AI hardware is no longer just Nvidia's story. A second layer of custom silicon is being built with borrowed money.

Bob: For European AI companies, this is a tougher landscape. European startups do not have access to sixty-billion-dollar debt facilities, and the venture capital market here is a fraction of the US one. Compared to the hundreds of billions being mobilised in America, European AI funding remains modest. Industry groups have argued for years that without dedicated financing tools, European players will keep losing ground to their US counterparts on hardware and infrastructure.

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.

A second wave of sovereign AI projects sweeps across sixty-seven countries as nations race for self-reliance.

Bob: A new report from the Center for a New American Security tracks a striking jump in government-backed AI initiatives. In the first half of 2026, one hundred and eighty-four sovereign AI projects were launched across sixty-seven countries. That's a significant increase from previous years, and the researchers describe it as a clear second wave. The goal, in almost every case, is national self-reliance across AI infrastructure, models, and data.

Samantha: The geographic shift is what stands out. Asia-Pacific has overtaken Europe as the most active region for new projects, with India leading the pack at eleven initiatives. Around eighty percent of these projects are infrastructure-focused: data centres, national compute programs, and access schemes for researchers and startups. Countries are treating AI compute the same way they once treated electricity grids or railway networks.

Bob: And that framing matters, because it changes who pays and who plans. When AI becomes a critical national capability, ministries of industry step in, sovereign wealth funds get involved, and procurement runs on very different timelines than private-sector deals. What this shows is a global rebalancing, where governments no longer trust that the market alone will deliver strategically important computing power.

Samantha: For Europe, the numbers are humbling. The continent added thirteen new sovereign AI projects in the first half of the year, respectable in absolute terms, but overtaken by Asia-Pacific in both pace and ambition. The EU has its AI Factories initiative and national compute programs in Germany, France and the Netherlands, but critics point out that these are fragmented compared to the more coordinated pushes seen in India, Singapore, and the Gulf states. The question European policymakers now face is whether to consolidate faster or accept that the frontier is being set elsewhere.

CISA orders urgent patching as attackers combine AI with classic exploits across Microsoft, VMware and Apple systems.

Bob: The US Cybersecurity and Infrastructure Security Agency has added four actively exploited vulnerabilities to its Known Exploited Vulnerabilities catalog. They affect Microsoft SharePoint, VMware vCenter, and Apple macOS. Federal agencies were told to patch them by today, August twenty-one. One of the flaws has reportedly been used by a Chinese-speaking threat actor in what researchers are calling an AI-enabled autonomous hacking campaign, running alongside manual operations.

Samantha: The reach is broad. Exploitation has been observed in forty-seven countries, with Germany, the US, Turkey, Iran, and France among the most heavily hit. What makes this campaign different is that AI is doing part of the reconnaissance and exploitation work autonomously, then handing off to human operators for the harder targets. That combination is faster and cheaper than a purely manual campaign, and it lowers the skill floor for advanced attacks.

Bob: The vulnerabilities themselves are serious on their own. They allow remote code execution and persistent access, which means an attacker who gets in can stay in and quietly move through a network. For any organisation running SharePoint on-premises, or VMware vCenter for virtualisation, or a fleet of Macs, this is a drop-everything patching moment. The bottom line is that critical patch windows keep shrinking because the attackers automate faster than defenders can.

Samantha: For Europe, and especially for German and French organisations already showing up in the telemetry, the response has to be immediate. National cybersecurity agencies in both countries have echoed the CISA guidance. The wider lesson is that AI-assisted attacks are no longer a future scenario. They're already changing the economics of intrusion, and that shifts the pressure onto every IT team to keep patching cycles measured in days, not weeks.

Trilium Notes turns your own laptop into a private, tree-structured knowledge base for research, writing and daily notes.

Bob: To close out, two things you can actually use today. First up, Trilium Notes. It's a free, open-source note-taking application that a growing community swears by for one specific reason: it organises everything in a hierarchical tree, which suits people who think in categories and sub-categories rather than tags or backlinks.

Samantha: The design choice is what sets it apart from the more famous alternatives. Instead of a flat pile of markdown files, you build branches. A note can live in multiple places at once through cloning, so you don't have to duplicate content. Users report that it scales comfortably to tens of thousands of notes without slowing down, which is why it's popular with researchers, lawyers, and technical writers dealing with long projects. It handles rich text, code blocks, sketches, maps, and even relationship diagrams between notes.

Bob: Practical details. Trilium Notes runs on Windows, macOS, and Linux as a desktop app, and there's also a server version you can host yourself if you want to sync between devices. Everything stays on your machine or your own server, so nothing is uploaded to a company cloud unless you set that up deliberately. Setup takes about ten minutes for the desktop version. It's completely free, no subscription, no account required.

Samantha: For Europeans in particular, the self-hosting option is worth flagging. If you want your notes on a server physically located in the EU, under EU data protection rules, you install it on a small virtual machine and you're done. Reviewers frequently praise it as the tool that finally handled their thesis notes, legal case files, or long research projects without falling apart. If Obsidian's linked-notes approach never quite clicked, Trilium's tree structure might be the mental model that does.

Bob: One small caveat worth mentioning. Trilium Notes has more of a learning curve than the polished commercial alternatives. The interface is functional rather than beautiful. But once you get comfortable with the tree navigation and the cloning trick, most long-term users say they never go back to anything else.

Grocy turns your kitchen into a smart household manager for groceries, chores, and meal planning without any subscription.

Samantha: Second one for today, and it's completely non-AI: Grocy. It's a free, open-source household management app, and the best way to describe it is enterprise resource planning for your fridge. It tracks what food you have, what's about to expire, what you need to buy, and even suggests recipes based on what's already in your kitchen.

Bob: The workflow is straightforward. You scan barcodes with your phone camera when you unpack the shopping, and Grocy logs the item, the quantity, and the expiration date. When you cook something, you consume the ingredients from the app, and your stock updates automatically. It also handles chores, so it can remind you when the coffee machine needs descaling or when the smoke alarm battery is due for a check. Some users even run it for their small businesses to track supplies.

Samantha: It runs on any web browser, and there are companion apps for Android and iOS to make barcode scanning quick. You can self-host it on a Raspberry Pi, a small home server, or a cheap virtual machine. There's also a hosted version through a few European providers if you don't want to run the server yourself. The whole thing is developed by a small independent team and is genuinely free.

Bob: What users report loving is the reduction in food waste. Once the app tells you that three items are expiring this week, meal planning becomes much more concrete. Households running Grocy for a few months typically say they spend less on groceries because they stop buying duplicates of things they already have buried in the back of the pantry.

Samantha: Today we covered: Nvidia's China chip comeback, Broadcom's sixty-billion-dollar debt hunt for AI, the global surge in sovereign AI projects, urgent patching orders from CISA, and two things to try yourself: Trilium Notes and Grocy.

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.