What we're seeing is the infrastructure reality catching up to the hype. AI companies sold the vision of intelligence as a utility—cheap, abundant, always available. But they're hitting the same constraints that plague every computing resource: hardware is finite, data centers take time to build, and physics doesn't care about your growth targets.

Anthropic is handling this better than some competitors. They're being relatively transparent about capacity constraints. They're still providing access, just with limits. But the existence of limits at all contradicts the fundamental premise of AI as an unlimited resource.

The question is whether this is temporary growing pains or a fundamental constraint on AI deployment. Optimists will say that Nvidia will ramp production, new chip manufacturers will enter the market, and infrastructure will catch up to demand. Maybe Google's memory compression breakthrough helps reduce hardware requirements.

But I'm skeptical. Every time we solve one bottleneck, the models get bigger and more demanding. GPT-5 will require more compute than GPT-4. The next generation of models will push hardware requirements even higher. We're on a treadmill where capability improvements consume efficiency gains.

What this likely means is a future of tiered AI access. If you can pay, you get priority access to the best models. Free tier users get throttled or pushed to smaller, less capable models. Enterprises get dedicated capacity if they're willing to commit to massive contracts.

In other words, AI stops being democratized and starts looking like every other scarce resource—allocated based on ability to pay.

Anthropic's usage limit changes are small and reasonable given their constraints. But they're a signal of larger issues the industry needs to address. You can't have infinite demand and finite compute resources without either limiting access or massively expanding infrastructure.

And infrastructure expansion isn't instant. Building data centers takes years. Chip fabs take even longer. If AI usage keeps growing at current rates, we're going to see more rationing, not less.

The technology is impressive. The demand is real. But the idea that everyone can have unlimited access to frontier AI models whenever they want? That's running headlong into physical constraints.

Anthropic's usage limits are the canary in the coal mine. The question is whether the industry can build infrastructure fast enough to meet demand, or whether we're headed for an era of AI scarcity.