Tuesday, 19 May 2026

The infrastructure race in AI is shifting in interesting ways this week, and I think it's worth paying attention to where the real competition is forming. Blackstone's $5 billion joint venture with Google to build a new TPU cloud isn't just another partnership announcement—it signals that the hyperscalers are running out of runway to build everything themselves, and the largest capital allocators are stepping in to fill the gap. This matters because it's no longer about Google or Amazon or Microsoft each owning their own fiefdom. It's becoming a question of who can mobilize capital and compute at scale faster than anyone else.

By the way, Microsoft's AI business just hit a $37 billion annualized run rate, driven mostly by Azure and the Copilot integration into Microsoft 365. That's not a niche number anymore. That's a material piece of their revenue stream, and it's happening faster than most people expected. The enterprise adoption curve for AI tooling has compressed dramatically. Companies aren't piloting anymore; they're deploying. That creates immediate pressure on competitors to match both capability and availability.

What's interesting alongside this is the integration layer starting to form around Claude. SandboxAQ, the Alphabet spinout chaired by Eric Schmidt, just embedded its quantitative models for drug discovery and materials science directly into Anthropic's Claude through the model context protocol. This is the kind of vertical stack I've been expecting—specialized domain expertise married to general-purpose LLM interfaces. It's not sexy infrastructure news, but it's probably more consequential than it looks. If you're a researcher or a company trying to run molecular simulations, you now do that through a conversational interface with Claude rather than specialized software. That's a shift in how people work, not just what tools they use.

There's one other thread I'm watching closely: Nvidia's Jensen Huang recently suggested that China may open its market to US AI chips after meeting with Trump. If that actually happens—and it's a big if—the geopolitical calculus around chip exports and AI infrastructure fundamentally changes. Right now, the US has export controls on advanced semiconductors. That's been a major constraint on where TPUs and GPUs can flow. A partial opening would reshape the entire global AI compute landscape.

What ties these threads together is accessibility. More TPU cloud options, faster enterprise adoption, better domain-specific integrations, and potentially broader chip distribution all point toward a world where building with AI becomes less about proprietary hardware advantage and more about who can execute best on the software layer. That's a more competitive market, and honestly, a healthier one.