One thing I think people keep missing is that two things can be true at the same time. 1/Compute is absolutely one of the biggest moats in AI. If you don’t have access to enough GPUs, power, networking and datacenter capacity, you’re simply not playing the same game. 2/You can still lose your shirt building that moat. The internet is a good analogy. Fiber was one of the most important technologies ever built. Without it, there’s no cloud, no YouTube, no Netflix… But a lot of the companies that financed and laid the fiber generated terrible returns. The infrastructure created enormous value, but much of that value was captured by everyone using it, not necessarily by the people who built it. The interesting thing today is that the hyperscalers are trying to avoid repeating that mistake. They’re increasingly funding their capex with long-term customer commitments before the infrastructure even comes online. But it’s the AI labs buying that compute that ultimately have to make the economics work. They’re betting their models stay at the frontier long enough to maintain pricing power and high margins. Those margins are what they’ll use to pay for the massive compute commitments they’re signing today. If frontier models become commoditized, that equation gets much harder. Compute is still a moat, but that doesn’t automatically translate into great returns. Compute is essential. Great unit economics are not guaranteed. The real question isn’t whether AI needs more compute but It’s who actually captures the value created by all this infrastructure?