The AI buildout is a stack of mismatches
Velocity, economics, value capture and capital risk are each manageable alone. Together they explain why capacity sits idle while bills climb.
Most commentary on AI infrastructure picks one problem: power, chips, cost or hype. I think the more useful view is that the buildout compounds four mismatches at the same time. Each one could be managed alone. Together they create a system that is expensive, fragile and oddly underused.
1. Velocity
Models and usage patterns change in months. Data centres, grid connections and power contracts take years. Capacity is being planned for demand that will look different by the time it arrives.
Agentic AI makes this sharper. Gartner estimates that agentic workloads use 5 to 30 times more tokens than a chatbot to complete an equivalent task. A shift in how people use AI can change compute demand by an order of magnitude, far faster than infrastructure can follow.
2. Economics
Capacity is bought for peaks and paid for all the time. In a 2024 industry survey by ClearML and the AI Infrastructure Alliance, 68% of companies reported peak GPU utilisation under 70%. Much of the cost of AI is not the model price. It is idle hardware.
The levers that fix this are operational, not heroic: use the fleet better, route simpler work to smaller models, compress models where quality allows, and understand which workloads really need a frontier model.
3. Value capture
The companies spending on infrastructure are not always the ones that capture the value it creates. Value tends to settle with whoever owns the customer relationship or the workflow, while the capital-heavy layers carry the cost. That gap shapes who keeps investing and who pulls back.
4. Capital risk
The spending is committed now against returns that are uncertain and later. If usage patterns shift, efficiency gains arrive faster than expected, or value capture moves, some of that capital will have been placed in the wrong spot, at the wrong size, at the wrong time.
Why the stack matters
Any one mismatch is a normal business problem. Stacked, they reinforce each other. Fast change raises the risk of over- or under-building. Poor utilisation worsens the economics. Weak value capture makes the capital harder to justify. And the pressure to show returns pushes organisations to deploy AI faster than their people can absorb it.
For an enterprise, the practical lesson is to treat AI spend like any other operating cost with levers. I built AI Fleet Economics to show how utilisation, agentic mix, routing and quantisation move a monthly bill. The biggest savings rarely come from negotiating a better model price. They come from using what you already pay for.