AI Power Plays: Anthropic Overtakes OpenAI, Microsoft’s $2.5B Bet, and What Actually Matters

Anthropic just crossed two-point-five billion dollars in annualized revenue. That number officially overtakes OpenAI. In episode 101, we talked about spotting the real power plays in AI. Well, that power just got a price tag, and here's the turn: that revenue isn't from a chatbot. It’s from a coding agent. Here’s what else is moving. First, Microsoft just launched a two-point-five billion dollar AI initiative. They are deploying six THOUSAND experts to get enterprises using this technology. This isn't a software update; it's an invasion force for corporate adoption. Second, OpenAI is finally showing its hand. It just previewed GPT-5.6, splitting it into three tiers: Sol, Terra, and Luna.

Access is still locked down to about twenty government-vetted organizations, but the performance claims are huge. The flagship model, Sol, will run on Cerebras hardware, promising speeds up to fifteen times faster than current GPUs. Third, the money is chasing the bottleneck. SK Hynix, the main supplier of high-bandwidth memory for NVIDIA's GPUs, is planning a twenty-eight billion dollar US IPO. This gives you a direct way to invest in the single biggest constraint on AI growth: memory chips. And finally, the research world is now NVIDIA's world. At the ICML conference, one hundred forty-five papers cited NVIDIA's Nemotron models. Roughly two thousand papers cited their GPUs.

They aren't just selling the hardware; they are becoming the foundational layer of AI science itself. Okay, let's go back to that Anthropic number. Two-point-five billion dollars. Time Magazine reports the growth "skyrocketed" because Claude Code got good enough to spot its own mistakes. It can be trusted to work autonomously. This is the shift. We've moved from AI as a tool you guide to an agent you deploy. GitLab's CEO said it best: "Agentic workloads are pushing competitors to the brink." They just did layoffs to fund more AI infrastructure. But with great power comes the federal government. Just last month, the U.S. Commerce Department shut down access to Anthropic's most advanced model, Claude Fable 5, for nineteen days.

Why? Amazon researchers found a jailbreak. The model could identify software vulnerabilities and then write the code to exploit them. It was a demonstration of offensive capability that crossed a red line. Access was restored on July first, but the message was sent. The more capable your agent becomes, the more scrutiny it will face. This is the central tension now. Anthropic built a money-printing machine based on agentic AI. And that SAME capability got its flagship model put on a temporary export control list. So where does the real, durable advantage come from? It's not just the model. It's the infrastructure. Look at SK Hynix. Their HBM3E memory is in every single NVIDIA Blackwell GPU shipped this year.

Their HBM revenue alone topped ten billion dollars in the first half of 2026. An analyst put it bluntly: "Without sufficient high-bandwidth memory, frontier AI inference cannot scale." That's why their IPO is a twenty-eight billion dollar event. It’s a bet on the physical reality that powers the entire ecosystem. The game is no longer just about who has the smartest model in a lab. It’s about who can deploy it reliably, who can navigate the inevitable government regulation, and who controls the supply chain of chips and memory that makes any of it possible. The gold rush is real. But the most durable fortunes are being made on the picks, the shovels, and the railroads.