The Wire Weekly
Episode 6 · May 11, 2026 · 00:08:01.81

This Week's Stories

Full Transcript

So OpenAI just committed over a trillion dollars to data centers while literally admitting in their financial disclosures that the underlying technology might not scale the way they need it to. At the same time, their AI coding assistants are getting rolled out to Amazon's engineers because Amazon's own tools were so bad their people rejected them. And somehow, Wall Street's fancy new AI trading bots are getting destroyed by basic market timing strategies a community college finance class could execute. That's the world we're living in. I'm Travis Wright, this is The Wire Weekly, and if this week proved anything it's that we're living through the most expensive experiment in tech history... and nobody knows if it's actually going to work. Let's start with the big one. OpenAI has now committed over ONE TRILLION DOLLARS to compute and data centers. That's not a typo. Twelve zeros. And buried in their recent statements is this beautiful little admission that the current LLM architecture might hit fundamental scaling limits. They're saying the quiet part out loud now. All that compute, all those chips, all that power... and they're not even sure the math works past a certain point. This is corporate chicken at a scale we've never seen. Because once you're a trillion dollars in, you can't exactly pivot to selling enterprise SaaS tools and call it a win. You're ALL IN on AGI or bust. And what's wild is the market is just... fine with this? We've normalized the idea that burning through essentially a mid-sized country's GDP on a hypothesis is just what frontier AI companies do now. But here's what nobody's talking about. This isn't just an OpenAI problem. This is the entire frontier AI race realizing they're in a game of who can afford to keep spinning the roulette wheel longest. Anthropic, Google, Meta, they're all making versions of this same bet. The difference is OpenAI said the quiet part loud. They told us there's a scenario where this doesn't scale, and they're betting a trillion anyway. I've been in crypto for nine years. I've seen some spectacular gambles. But this? This makes the ICO boom look like a bake sale. At least with crypto we knew the tech worked, we just didn't know what people would BUILD with it. Here, we're not even sure the foundation holds. Now let's connect some dots because three things happened this week that LOOK unrelated but they're absolutely not. First, Anthropic's AI trading agents are bombing in public tests. These things were supposed to eat Wall Street. Instead they're losing to basic timing strategies. We're not talking about getting beat by Renaissance Technologies level quant stuff. We're talking basic moving averages and momentum plays. Second, Amazon just rolled out Claude and Codex to all their corporate engineers after their own people straight up rejected Amazon's internal AI tools. Think about that. Amazon. One of the most engineering-driven companies on Earth. Built their own AI coding assistant. And their engineers said no thanks, we'd rather use someone else's. Third, Genesis AI just dropped their robotics foundation model with a working demo of robotic hands doing complex manipulation. Not a render. Not a carefully edited video. Actual hardware executing tasks that would've seemed impossible two years ago. See the pattern? The AI that was supposed to work in structured environments like trading and enterprise coding is struggling. But the AI that everyone thought was years away, the physical embodiment stuff, that's actually DELIVERING. This tells us something important about where we are in the curve. LLMs are incredible at certain things. They're transformative for unstructured tasks, creative work, research synthesis. But the moment you need them to operate in constrained systems with clear right and wrong answers... they're weirdly brittle. Meanwhile, the robotics stuff that combines vision models with physical feedback loops is leapfrogging expectations. Wall Street wants deterministic outcomes. Code either compiles or it doesn't. But robots? Robots can learn from failure states in real time. They're not trying to predict the perfect answer, they're iterating through physical trial and error at machine speed. Turns out that might be closer to actual intelligence than next-token prediction. Alright, wild card time. And this one is TOO good. There's a startup called Orbital Inc backed by Andreessen Horowitz. Their plan? Launch AI inference data centers into SPACE to bypass Earth's power grid constraints. Because why deal with terrestrial infrastructure when you can just... go orbital? The pitch actually makes sense on paper. Unlimited solar, no cooling costs in the vacuum of space, no NIMBYs complaining about your data center. Very cool. Very sci-fi. One small problem they apparently didn't think through until recently: latency. Turns out when your AI is processing requests in LOW EARTH ORBIT, you still have to beam the answers back down. And physics is undefeated. They're hitting a 550 millisecond latency wall that nobody saw coming. For context, that's an eternity in computing time. You could run inference on a potato-powered server in Kansas and get better response times. This is the kind of thing that happens when you have too much VC money and not enough engineers saying "wait, have we thought about speed of light?" It's not a bad idea, it's just hilariously premature. Maybe in ten years when we have true edge computing in orbit and we're doing batch processing for long-running simulations. But real-time inference? From space? Not yet, my friends. Not yet. So what should you actually be watching next week? First, keep an eye on Nvidia's supply chain moves. They just dropped HALF A BILLION on optical fiber infrastructure with Corning. Everyone's obsessing over which chip architecture wins, but Nvidia's betting the real bottleneck is data movement. If they're right, we're about to see a wave of infrastructure plays that have nothing to do with compute and everything to do with bandwidth. Second, watch what happens with Aptos. They just committed fifty million of their own money to build out AI agent infrastructure on their Layer 1. Not VC money, their OWN treasury. That's a bet that agentic AI needs blockchain rails, and they're willing to spend real capital to prove it. If that works, every L1 is going to scramble to copy the playbook. Third, and this is the big one, watch PwC. They just became OpenAI's first Big Four reseller. They're not consulting on AI strategy anymore, they're SELLING the actual tools. That's a massive shift. If PwC can successfully deploy AI agents across enterprise finance at scale, that's the proof point every other Fortune 500 CFO needs to cut the check. And if they can't, well... that tells us something too. The through-line this week is pretty clear. We're ALL IN on a technology that might not work the way we think it does, we're betting trillions on scaling laws that might not hold, and somehow the stuff we thought was impossible is shipping while the stuff we thought was solved is falling apart in production. That's not a criticism. That's just where we are. The frontier is messy. It's supposed to be. But if you're building in this space, if you're investing in this space, if you're making career bets on this space, you need to understand we're in the experimental phase of the most expensive experiment ever run. Place your bets accordingly. I'm Travis Wright. This is The Wire Weekly. I'll see you next Tuesday.