VC10X Pulse - Is Meta Selling Excess Compute?

This week, reports emerged that Meta has been offering excess AI compute capacity to external customers. For some investors, that immediately raised alarm bells. Has AI demand peaked? Did Meta overbuild? Or is this simply what happens when you're deploying AI infrastructure at unprecedented scale? The answer matters, not just for Meta, but for the entire AI investment thesis. Let's separate signal from noise. So what actually happened? The key fact is this: Meta has reportedly begun making some excess AI compute capacity available externally rather than leaving it idle. That's very different from saying AI demand has collapsed. In fact, Meta continues to invest aggressively in AI infrastructure and has repeatedly guided toward enormous capital expenditures to support future AI projects. So the question isn't whether Meta still believes in AI. It clearly does. The question is: Why would a company investing hundreds of billions in AI sell excess capacity? Now, it's important to understand overbuilding isn't necessarily bad. When companies build infrastructure, they rarely build exactly enough capacity. They intentionally build ahead of demand. Think about Amazon building AWS, cloud providers building data centers, telecom providers laying fiber. For periods of time, there is almost always unused capacity. That's actually healthy because if you're constantly operating at one hundred percent utilization, you can't onboard new customers or launch new products efficiently. From that perspective, selling unused compute isn't necessarily a warning sign. It's simply monetizing an idle asset. Now, the investor question is this: Is AI demand slowing? This is where context matters. If AI demand were genuinely weakening, we'd expect to see signs across the ecosystem. Hyperscalers cutting CapEx, GPU demand slowing, memory prices falling, data center projects being canceled. Instead, we are seeing the exact opposite. Recent results from companies like Micron, NVIDIA, and major cloud providers continue to point to robust AI infrastructure demand. So one company's excess capacity is not enough evidence to conclude that AI build-out is slowing. Now, what could actually explain this? There are several plausible explanations. First, Meta may have built faster than its internal AI workloads could immediately absorb. Second, training demand is often uneven. Large clusters may sit underutilized between major training runs. Third, selling spare capacity helps offset infrastructure costs while keeping systems productive. In other words, this could simply be better capital allocation, not weaker demand. I actually think this highlights something more interesting. We are entering the next phase of AI infrastructure. The first phase was build as much compute as possible. The second phase is optimize utilization. Investors should not just ask who owns the most GPUs. They should ask who owns the highest return on those GPUs. As AI infrastructure becomes more abundant, utilization and monetization become just as important as capacity itself. Now, what should investors watch from here? Rather than focusing on one headline, watch for these indicators. Are hyperscalers