Latest / UpNext AI / South Korea’s AI Buildout, Google’s SpaceX Compute Deal, and Multimodal Lifelong Learning | UpNext AI – June 8, 2026
Transcript
- Welcome to the UpNext AI podcast. It's Monday, June 8th, 2026, and here's what matters in AI today. We start in Seoul, where Nvidia is using CEO Jensen Huang’s Korea trip to make a broader point: South Korea is becoming one of the world’s centers of AI infrastructure, robotics, and what Nvidia calls AI factories. This comes from Nvidia’s own running update page, so the framing is promotional, but the details are still useful. Nvidia says NAVER is building a full-stack Nvidia AI factory in Korea with Nvidia DSX, and plans to expand its GAK Sejong facility to 55 megawatts and beyond to gigawatt scale. Nvidia also says SK Telecom plans to build a gigawatt-scale AI cloud in Korea using the Nvidia DSX platform to support sovereign, physical, and agentic AI services. The company also used the trip to spotlight industrial partnerships. Nvidia and LG announced plans to build an AI factory supporting robotics, autonomous driving, data center technologies, and GPU cloud services. And Nvidia says its collaboration with Doosan is expanding across physical AI, robotics, and AI factory infrastructure. So the bigger takeaway here is not one product launch. It’s that Nvidia is presenting South Korea as a full-stack AI ecosystem: memory, cloud, manufacturing, robotics, gaming, and national-scale infrastructure all at once. If you want a snapshot of where the next wave of AI deployment is headed, it looks a lot less like one chatbot and a lot more like regional compute capacity tied directly to industry. From there, to the economics of that buildout. TechCrunch reports that Google will pay SpaceX 920 million dollars per month for compute. According to TechCrunch, the agreement runs from October 2026 through June 2029 and gives Google access to approximately 110,000 Nvidia GPUs, CPUs, memory, and related components. The report says a Google representative described the deal as a short-term bridge agreement driven by stronger-than-expected demand for recently launched AI products, specifically surging customer demand for its agent platform and Gemini Enterprise. That makes this one of the clearest recent signals of how hungry major AI platforms still are for capacity, even when they already own enormous amounts of infrastructure themselves. TechCrunch notes that Alphabet has already committed to more than 180 billion dollars in capital expenditures this year and said that figure could rise again in 2027. What stands out is the word bridge. Even at Google’s scale, demand can outrun supply. And when that happens, the answer is no longer just buying more chips over time. It’s signing massive compute agreements to keep the platform moving now. The bottom line is simple: AI growth is still running hard into infrastructure constraints, and the spending required to smooth that out remains enormous. One paper worth your attention today is ProtoAda, posted earlier this week on arXiv. The full title is “ProtoAda: Prototype-Guided Adaptive Adapter Expansion and Geometric Consolidation for Multimodal Continual Instruction Tuning.” In plain English, this is about how a multimodal model, one that works across text and images, can keep learning new vision-language skills over time without having to retrain the whole system from scratch each time. The paper proposes two main ideas. First, prototype-guided adaptive adapter expansion, which means adding smaller task-specific adaptation modules as new capabilities arrive. Second, geometric consolidation, which is meant to organize those additions so newer learning interferes less with older skills. The problem it is trying to solve is catastrophic forgetting, where a model picks up something new and in the process gets worse at things it already knew how to do. The authors position ProtoAda as a way to let multimodal systems acquire new capabilities incrementally while reducing that tradeoff. We do not have benchmark details in the supplied summary, so the practical gains are still directional here. But the reason the paper matters is straightforward: if multimodal assistants are going to live for a long time in production, they need a cheaper and safer way to keep learning. Bottom line: ProtoAda is an attempt to make multimodal models more maintainable by teaching them new skills in smaller pieces instead of retraining the whole machine every time. ...Are you building apps with voice? Elevate your app's voice capabilities with ElevenLabs. Their API is a game changer for embedding dynamic, responsive voice interactions in your applications, providing unprecedented realism, flexibility and latency. In fact, you're listening to one of their voices - right - now. If you are a developer looking to elevate user experience with natural voice interfaces, this is your solution. Visit up next dot fm slash eleven to check out their latest offerings. ... First, Simon Willison has released an alpha package called micropython-wasm, along with a Datasette Agent plugin, to run Python code inside a WebAssembly sandbox. His goal is to give plugin-style code tighter limits around file access, network access, memory, and CPU use. He’s explicit that this is still an alpha and not something to trust blindly, but it’s a practical developer story worth watching because safe code execution keeps becoming more important as agents get more capable. Next, Microsoft’s Xbox Games Showcase was mostly about games, but The Verge says it also highlighted continued confusion around Microsoft’s exclusivity strategy. The recap points to a mix of platform decisions, including Gears of War: E-Day not coming to PS5, while other major first-party titles such as Fable and Halo: Combat Evolved are still coming to Sony’s console. Also, Ars Technica reports that a school shooting survivor has sued the maker of an AI gun-detection system after the system failed to detect a weapon in a January 2025 shooting at a Nashville high school. According to Ars, the lawsuit argues the company knew or should have known about operational limitations tied to things like camera placement, angle, lighting, and weapon visibility. And finally, The Decoder reports that Perplexity is introducing what it calls Search as Code, an architecture meant to let AI models write their own search pipelines instead of relying on fixed APIs. We only have the summary here, so the implementation details are still thin, but the concept is notable because it pushes search from a fixed tool call toward something more composable and agent-like. Before we wrap up, a quick note: this podcast is generated with the assistance of AI and is intended for informational purposes only. All referenced articles, research, and commentary remain the property of their original authors and publishers. If you enjoyed this episode, don't forget to subscribe, rate, and leave us a review! And that's your briefing for today. Full source links are in the episode notes, and we'll be back tomorrow with what's up next!