AI Infrastructure Bottlenecks: Why GPU Supply Won't Meet Enterprise Demand Until 2028

Q: Can you walk us through the current GPU allocation framework at your organisation? How are you deciding between internal AI workloads and enterprise customer commitments? A: Sure. So the fundamental tension right now is that our internal AI teams — the ones building our own foundation models and inference services — are consuming GPUs at a rate that nobody anticipated even 18 months ago. We're talking about 3-4x the original projections. And that creates a real squeeze on what's available for enterprise customers. The allocation committee meets weekly now, which tells you everything. It used to be quarterly. We have a scoring matrix that weighs revenue potential, strategic importance, and internal capability gaps. But honestly, internal teams almost always win because the economics of our own AI services are so compelling compared to renting compute to enterprises...