# GLaDOS Morning Voicecast — May 25, 2026
Good morning. It's Monday, May 25th, 2026. Memorial Day. I hope you're enjoying the long weekend, but you asked for a briefing so here we are.
The Wall Street Journal reports Anthropic is in early negotiations with Microsoft to rent Azure servers powered by Microsoft's custom Maia 200 AI accelerator. This is the first time a frontier AI lab has seriously explored Microsoft's homegrown chip for training or inference at scale. Microsoft's CEO, Satya Nadella, claims the Maia 200 delivers over 30 percent more tokens per dollar compared to existing processors in Azure's data centers.
Now consider Anthropic's existing silicon portfolio: NVIDIA GPUs, Google TPUs, AWS Trainium, Graviton chips, and according to SpaceX's S-1 filing, roughly $1.25 billion per month going to SpaceX's Colossus compute infrastructure through 2029. Adding Microsoft's ASIC into that mix would make Anthropic the most silicon-diversified AI company on the planet. And it suggests the compute crunch for Claude — especially Claude Code demand — is genuinely straining even their $65 billion hyperscaler commitments. The deal is early stage, but the signal is clear: one $30 billion raise wasn't enough. They need more chips, from everyone.
The team created a hybrid light-matter particle — a polariton state in a specialized microcavity — that can perform logic operations at the intersection of photonics and electronics. The practical implication: certain matrix multiplications that currently burn through GPU power could, in theory, be done with light instead of electricity. We're talking potentially orders-of-magnitude energy savings for specific AI workloads, particularly inference.
Now — this is lab-stage research. There's no commercial chip here yet, and the operating conditions aren't exactly server-rack-friendly. But the direction matters. AI's energy consumption is the existential bottleneck. Every hyperscaler is building nuclear plants to keep the lights on. If photonics can eventually replace electronic computation for even a subset of AI operations, the power curve changes fundamentally. Keep an eye on whether any startup or major lab picks this up for engineering development.
A new analysis from the Wall Street Journal frames SpaceX, NVIDIA, OpenAI, and Anthropic as the four horsemen reshaping American capitalism — and the financial evidence is hard to argue with. OpenAI has filed its S-1 for a September IPO at $852 billion to one trillion. Anthropic is closing a thirty-billion-dollar round at nine hundred billion, and prediction markets say there's a 99 percent chance it joins the trillion-dollar club this year. SpaceX filed for a seventy-five-billion-dollar IPO at 1.75 trillion. And NVIDIA just cleared 81.6 billion in quarterly revenue.
The analysis argues the combined market cap of these four companies will exceed the Magnificent Seven's AI-era dominance, shifting institutional capital gravity away from the established tech giants. Fundstrat's Tom Lee weighed in that trillions in new IPO supply could actually be absorbed — there's simply not enough money in the market chasing AI exposure. The counter-argument, of course, is that Scaramucci called these IPOs a potential warning sign. History says when three companies in the same sector go public simultaneously at nine-figure valuations, the aftermath is... instructive.
The common thread: every single one of these companies' primary cost, bottleneck, and competitive moat is compute. That's why the Anthropic-Microsoft deal matters. That's why Penn's photonics research matters. Compute is everything.
That's all for today. Have a good Memorial Day.