Saturday, 6 June 2026

The infrastructure underpinning AI is becoming as strategically important as the chips themselves, and this week's moves make that shift unmistakable. SpaceX just inked a $30 billion deal with Google—roughly $920 million a month—to supply computing capacity. That's not a side project for a rocket company. It's a signal that as AI models grow more demanding and competition for compute intensifies, the companies that control the physical layer will command real economic power.

What interests me here isn't just the scale, though it's substantial. It's the timing and the pattern. SpaceX is preparing for an IPO, and having locked in a major, multiyear revenue stream from one of the world's largest tech companies is exactly the kind of story that appeals to investors. But Google isn't just buying compute—it's securing supply during a period of extraordinary scarcity. The irony is worth noting: Anthropic, which competes directly with Google's own AI division, was already a SpaceX customer. Now Google itself has moved to lock down capacity. Everyone is racing to ensure they have the resources to train and run the next generation of models, and that competition is reshaping the infrastructure market.

By the way, while companies fight over compute capacity, a different kind of constraint is emerging: security. Microsoft's threat intelligence team identified seven new attack vectors against AI agents, with a particularly sharp finding about Anthropic's Claude Code GitHub Action potentially exposing CI/CD secrets when agents process untrusted inputs. The specificity here matters. These aren't theoretical vulnerabilities—they're concrete failure modes in real workflows. Microsoft's response has been Execution Containers, a security boundary layer for Windows that tries to limit what agents can actually do. The problem is becoming clearer: as we deploy AI agents into production systems, the blast radius of a compromised agent grows exponentially. We're moving faster on capability than on containment.

Which brings me to something that feels almost quaint given the commercial frenzy: both the U.S. and China have reportedly signaled openness to cooperating on AI safety, even as their competition in AI capability intensifies. Anthropic itself has been frank about the risk that humans could lose control over advanced AI systems. On the surface, it sounds like acknowledgment of the obvious—we're building powerful tools and need to be thoughtful about it. But I find the gap between that acknowledgment and actual resource allocation telling. We're spending tens of billions on compute and infrastructure. How much are we spending on the kinds of safety research that might actually matter if things go wrong?