Dell NVIDIA agentic AI on-premises: EU data sovereignty win

Samantha: Welcome to The State of Tech, The European Edition, Tuesday May nineteen, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: Dell and NVIDIA bring agentic AI to your desk, Google unveils a major Gemini upgrade at I/O, Anthropic briefs G20 regulators on cybersecurity risks from its frontier model, a Korean court limits a Samsung strike to protect chip supply, plus two things you can actually try this week, a Korean startup teaching robots by watching master craftsmen, and a smarter customer service AI for businesses. Let's start with Dell.

Dell and NVIDIA launch desk-side agentic AI, letting companies run autonomous workflows locally without the cloud.

Samantha: Dell Technologies and NVIDIA have launched something called Dell Deskside Agentic AI. In plain English, it's a powerful machine that sits in your office and runs autonomous AI agents locally, no cloud connection required. It's an expansion of their AI Factory partnership, aimed at teams who want AI workflows that don't ship sensitive data offsite.

Bob: And the reason this lands today is cost and control. Cloud bills for running AI have been climbing fast. Companies are seeing latency problems, vendor lock-in, and growing unease about where their data actually lives. A local box that can run agentic workflows, meaning AI that takes multi-step actions on its own, changes that calculation.

Samantha: The target audience is workgroups inside larger firms. Think legal teams processing contracts, engineering teams running simulations, or finance teams crunching forecasts. Instead of sending all that to a hyperscaler, the work happens on a machine down the hall.

Bob: There's a market shift here too. For two years the story has been: AI lives in the cloud, only the big three can host it. Dell and NVIDIA are pushing back, arguing serious AI workloads can sit on-premises again. If that catches on, the cloud growth curve for AI flattens.

Samantha: The trade-off is upfront capital expense versus monthly cloud spend. For organisations with strict data rules, that's an easy decision. For others, it depends on usage patterns.

Bob: And it puts pressure on cloud providers to sharpen their pricing on AI services, because suddenly there's a real alternative for customers who want to keep AI in-house.

Samantha: One caveat: local AI hardware is only as good as the models running on it, and the software support around agentic workflows is still maturing. Early buyers will be doing some of the integration work themselves.

Bob: True, but for organisations that have been waiting for a credible non-cloud option, this is the most serious offering yet. The agentic AI conversation just got a new venue, and it's not a data centre in Virginia.

Google reveals Gemini 4.0 at I/O 2026, promising Android phones that complete complex tasks on their own.

Bob: Over to Mountain View. Google opened its I/O developer conference with a big Gemini update, widely expected to be Gemini 4.0. The headline feature: AI that can carry out complex, multi-step tasks across different Android apps, with the user's approval.

Samantha: So instead of asking your phone to set a timer, you ask it to book a restaurant, message three friends, and add it to your calendar, and it actually does all of that across separate apps. That's the promise. Whether it works smoothly in daily use is the next question.

Bob: Google is positioning this as the foundation layer for Android going forward. Every new feature in the coming year will lean on it. That's a strategic bet against OpenAI and Anthropic, who don't own a phone operating system.

Samantha: There's also a competitive angle around timing. OpenAI is rumoured to have its next model close to launch. Google wants developers building on Gemini before that happens.

Bob: The user impact comes down to whether app automation actually saves time or just creates new failure modes. Letting AI act on your behalf is convenient until it books the wrong restaurant or messages the wrong person.

Samantha: The other thing to watch is battery life and data use. Running agents continuously in the background is expensive on both fronts, and Google has been quiet about that side of it.

Bob: It always is at launch. We'll see the real numbers once people use it for a few weeks.

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.

Anthropic briefs G20 financial regulators on cybersecurity risks from its unreleased Claude Mythos model.

Bob: Anthropic has been talking to G20 financial regulators, including the Financial Stability Board, about the cybersecurity risk profile of its advanced Claude Mythos model. That's an unusual move, and it tells you how seriously regulators are taking frontier AI.

Samantha: Alongside that, Anthropic clarified its Project Glasswing plan, which lets partners like Amazon, Microsoft, NVIDIA, and Apple use a preview version of Mythos for defensive cybersecurity work. The worry: if a model is smart enough to defend systems, it's smart enough to attack them.

Bob: And financial systems are the obvious concern. A model that can find vulnerabilities in banking infrastructure is both a defensive tool and a weapon, depending on who's holding it. Anthropic is trying to get ahead of regulators by briefing them directly.

Samantha: It's a clever positioning move. Instead of waiting to be regulated, Anthropic walks into the room with the regulators and frames the conversation. It also makes life harder for competitors who haven't done the same.

Bob: The risk profile of frontier models being discussed at G20 level shows how AI has moved from a tech industry topic to a financial stability topic in under two years. That's a fast escalation.

Samantha: The unresolved question is whether briefings like this become standard practice or stay voluntary. Right now it's goodwill from Anthropic. Tomorrow it might be a requirement.

Bob: And once it becomes a requirement, smaller AI labs without policy teams will struggle to keep up. That further concentrates frontier AI in the hands of a few well-resourced players.

A South Korean court forces Samsung's union to keep chip lines running during a planned strike.

Samantha: A South Korean court has ordered Samsung Electronics' labour union to maintain minimum staffing at its semiconductor plants, even if the planned general strike goes ahead. The ruling classifies key chip processes as maintenance operations, which under Korean labour law must continue.

Bob: This matters because Samsung memory chips end up in roughly half the world's smartphones, servers, and laptops. A real shutdown, even short, ripples through global supply chains within days. The court has effectively limited the union's leverage.

Samantha: It's also a precedent. By labelling chip production as essential infrastructure, the court is signalling that semiconductors now sit alongside power and water in terms of strategic importance. That's a meaningful shift in how the industry is treated legally.

Bob: The union still has options, but a full stoppage isn't one of them. Expect more targeted actions, slowdowns, or selective walkouts that put pressure on Samsung without triggering the court order.

Samantha: For Samsung, it buys breathing room. For workers, it raises the bar on collective bargaining in strategically important industries, not just in Korea.

Bob: It also reinforces why the US, Europe, and Japan are all pushing for domestic fabs. One labour dispute on the other side of the world shouldn't be able to stall your industrial base.

Samantha: To close out, two things worth your attention this week, a fresh approach to teaching robots, and a smarter way for companies to handle customer calls.

A Korean startup is recording master craftsmen at work to teach robots human-level dexterity.

Bob: South Korean startup RLWRLD is taking an interesting angle on training robots. Instead of writing rules or generating synthetic data, they're filming skilled human workers doing their jobs. Hotel staff folding napkins, logistics workers handling fragile goods, that kind of detailed expertise.

Samantha: They're building what they call universal AI brains, software that lets robots perceive, decide, and act with human-like skill. The data comes from people who've spent years mastering small physical tasks that are surprisingly hard to automate.

Bob: The reason this is worth knowing about: most robot training so far has been either rule-based or done in simulation. Both have limits. Real human dexterity is messy in ways simulations don't capture. A napkin fold involves dozens of tiny adjustments a master worker does without thinking.

Samantha: RLWRLD is targeting deployment in industrial sites and potentially homes by 2028. That's ambitious, but the approach is concrete. They're partnering with companies in hospitality and logistics to collect the data on-site.

Bob: If you're curious about where robotics is heading, this is worth following. It's a different bet from what Boston Dynamics or Tesla are doing. They're going humanoid first and figuring out skills later. RLWRLD is going skill-first and worrying about the body later.

Samantha: There's also a quiet ethical question. The master craftsmen being recorded are essentially training their own replacements. The startup says it compensates workers for their time, but the longer-term picture is uncomfortable.

Bob: And it's a useful reminder that the robotics race isn't only being run in California. Korean and Japanese players are doing serious work with a very different philosophy.

WIZ.AI launches Wizlynn, a customer service AI that handles real conversations, interruptions and all.

Bob: WIZ.AI has launched Wizlynn, a customer service AI platform that goes well beyond the chatbots most of us have suffered through. It handles real-time voice conversations, copes with interruptions, speaks multiple dialects, and has compliance guardrails built in for regulated industries like banking.

Samantha: The reason this matters for ordinary users: if you've ever been stuck in a phone menu or argued with a bot that doesn't understand you, this is the category trying to fix that. Whether it succeeds in practice is another question, but the bar is finally rising.

Bob: For business listeners, Wizlynn is a deployable platform, not a research project. It integrates with existing enterprise systems, which is usually where these tools fall over. The compliance angle, especially for financial services, is the differentiator. Most AI chatbots can't handle regulated conversations safely.

Samantha: The honest test is what these calls sound like in six months. The promise of natural AI conversation has been made before. The difference now is the underlying models are genuinely better at handling messy real-world dialogue.

Bob: If you run a customer service team, even a small one, it's worth requesting a demo this week to see where the technology actually sits today versus the marketing.

Samantha: Today we covered: Dell and NVIDIA's desk-side agentic AI, Google's Gemini 4.0 reveal, Anthropic briefing G20 regulators, the Samsung strike ruling, a Korean startup teaching robots from craftsmen, and Wizlynn for smarter customer service.

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.