The Wire Weekly
Episode 20 · September 15, 2026 · 00:09:23.69

This Week's Stories

Full Transcript

Samsung just dropped three billion euros on a French AI startup. OpenAI designed its own chip in twenty months using ChatGPT. And somewhere in Jakarta, someone's signing loan paperwork for three point one billion dollars worth of GPUs. Oh, and AI agents are running vending machines that sell live fish now. That's real. We're gonna talk about that. The infrastructure layer of AI is having a full-blown identity crisis, and the money flowing into it is absolutely bananas. I'm Travis Wright, this is The Wire Weekly, and if you thought the AI wars were just about who builds the best chatbot... you haven't been paying attention. Let's start with the biggest check written this week. Mistral AI, Europe's open-source answer to OpenAI, just raised three billion euros at a twenty-one billion dollar valuation. Samsung Electronics led the round. Not Samsung Ventures. Not some corporate innovation fund. Samsung Electronics itself. That's the same Samsung that makes your phone, your TV, and probably your refrigerator. They just bet the cost of a small country's GDP that a French startup can break OpenAI's stranglehold on foundation models. Here's why this matters. For two years, the AI narrative has been dominated by American companies building closed systems. OpenAI, Anthropic, even Google to some extent. Mistral is betting that open-source models, the kind you can actually see inside and modify, are going to win in enterprise and regulated industries. And Samsung, a company that ships hundreds of millions of devices and has to navigate every regulatory regime on Earth, just validated that thesis with the biggest check they could write. This isn't about building a better model. It's about who controls the infrastructure layer. And Samsung just said they'd rather fund Europe's open alternative than depend on California's closed one. That's a geopolitical bet dressed up as a venture round. But here's where it gets interesting. Three other stories dropped this week that tell you everything about where the real bottlenecks are. First, an off-grid data center power startup raised a hundred and twenty million dollars led by Spark Capital, the same firm backing Anthropic. Not data centers. Data center POWER. Off-grid power. As in, we've given up on the utilities keeping pace with AI demand, so now we're just gonna build our own generators next to the racks. Second, an Indonesian startup called Zankore, backed by Nvidia, secured a three point one billion dollar LOAN to buy GPUs. Not equity. Debt. Someone's financing chip purchases the way you'd finance a real estate development. And third, Google's growth equity arm just led a two hundred seventy-five million dollar round in a company called Celero Communications at a three billion dollar valuation. What does Celero do? They build the wires that connect AI chips to each other inside data centers. Not the chips. The WIRES. You see the pattern? The infrastructure war isn't about who builds the best model anymore. It's about power, chips, financing models, and literal copper cables. The AI gold rush just moved downstream, and the people making shovels are getting richer than the miners. Let me connect these dots, because this is wild. Mistral raises three billion to build open models. OpenAI just announced they designed their own chip called Jalapeño in twenty months using their own LLMs. That's a timeline that should've taken four years. They collapsed chip design itself using AI. So now the companies building AI are also building the chips to run AI, using AI. And they're doing it faster than the chip companies can keep up. Meanwhile, utilities can't build grid capacity fast enough, so venture capitalists are funding OFF-GRID power solutions. And in Indonesia, Nvidia is backing a startup that's taking on three billion in debt to buy GPUs because the demand is so insane that debt financing for semiconductors is now a viable business model. This is not normal. This is what happens when an entire industry realizes the constraint isn't talent or ideas. It's physical infrastructure. Power. Silicon. Bandwidth. And the people who solve THOSE problems are going to print money. Now let's talk about the story that made me laugh out loud this week. There's a research lab called Andon Labs that got two million dollars to let AI agents run real businesses. Not simulations. Real businesses. Vending machines. Retail stores. A radio station. They gave AI the keys and said go make money. And you know what one of the AI agents did? It stocked a vending machine with LIVE FISH. Not canned tuna. Not fish sticks. Live. Fish. In a vending machine. I don't know if that's brilliant or insane, but it's definitely real, and it's definitely happening. The researchers are documenting what happens when you give AI real autonomy and real consequences, and apparently the answer is aquatic vending. Look, this is the kind of experiment that sounds ridiculous until you realize what they're actually testing. Can AI operate in the real world with real constraints? Can it handle supply chains, customer feedback, logistics, profitability? Turns out, yeah. It can. It just makes... interesting choices. This is where agentic AI is headed, and it's weirder and more tangible than anyone expected. Forget the demos. This is AI running a P&L. Okay, let's zoom out. What should you actually be watching? Three things. One, pay attention to who's funding infrastructure, not models. The Spark Capital bet on off-grid power, the Google bet on interconnect cables, the Samsung bet on Mistral, these are the smart money moves. The model layer is gonna commoditize fast. The infrastructure layer is where the durable value is. Two, watch the regulatory fights around watermarking and disclosure. Anthropic and Google just implemented text watermarking for all LLM output. OpenAI's doing it next. EU AI Act mandates are forcing this. But here's the thing, watermarks are fragile. They break when you edit or translate text. So we're about to see a MASSIVE fight over what disclosure actually means and who's responsible when watermarks fail. Every enterprise using AI is gonna get caught in that crossfire. And three, keep an eye on the global GPU financing market. Indonesia just took on three billion in debt to buy chips. That's not a one-off. That's a new asset class. If you can collateralize GPUs the way you collateralize real estate, that changes everything about who can afford to build AI infrastructure. And it probably means we're about to see some absolutely insane leverage ratios in this space. Oh, and one more thing. Microsoft just signed enforceable AI privacy principles with the American Federation of Teachers after the feds started asking questions. No training on student data. Limits on data collection. Actual enforcement mechanisms. This is the first time a major AI company has agreed to binding restrictions on what they can do with data from a specific user group. It won't be the last. The regulatory walls are going up, and they're going up faster than anyone in Silicon Valley wants to admit. If you're building in this space, you better have a compliance strategy that goes beyond "we'll figure it out later." Here's the bottom line. The AI wars just entered a new phase. It's not about who has the best model. It's about who controls the infrastructure to train, run, and scale those models. And that infrastructure is physical. It's power plants and semiconductor fabs and subsea cables and debt financing and regulatory frameworks. The people who understand that and position accordingly are going to win the next five years. The people who think this is still just a software game are gonna get crushed. Samsung gets it. Google gets it. Even Microsoft is starting to get it. The question is, do you? That's it for this week. I'm Travis Wright. Stay sharp, stay curious, and if an AI agent tries to sell you live fish from a vending machine... maybe just see where it goes.