Executive Summary
- AI inference economics will save the grid; by using accordian infrastructure, operators can handle AI-driven demand spikes and contract the footprint when the grid needs relief.
- The era of over-provisioning is over, as 2026 is bringing in the era of inference models and, with that, unpredictable spikes of demand.
- The most successful CTOs are deploying flexible, ‘breathing’ infrastructure that enables them to be truly grid-ready.
When AI burst onto the scene, businesses went bananas with the potential of big picture products and services, and got into a habit of over-provisioning. FOMO was huge back then. Now, inference economics is entering that chat, forcing businesses to recalculate their infrastructure. This year, with the state of the grid, blind bulk buying is a massive liability. Traditional capacity planning is like buying an expensive waterproof winter coat that you wear one day a year – and it’s June, so heatwaves are incoming. Inference costs have dropped 280-fold over the last two years; there is still growth in overall AI spending, because usage, in the form of inference, has outpaced cost reduction.
Fixed AI infrastructure threatens grid readiness
Training AI models is a predictable power draw that’s continuous: you can anticipate them pretty well. Inference is the wild second child – it creates massive, erratic spikes in power demand based on real-time human-to-application usage.
The grid is buckling under the pressure of these high-density workloads and then millions of users query AI applications at the same time during a peak demand time on a sunny June day. This is where fixed AI infrastructure risk triggers local grid failures. This is where the accordion model for infrastructure comes in.
What is the accordion model?
Accordion infrastructure is a dynamic design where a data centre’s power and computing footprint expands and contracts in real-time to match immediate demands and grid constraints. It is an elastic architecture, almost as if the infrastructure is ‘breathing’, so when AI inference spikes occur, the infrastructure ‘expands’ by spinning up flexible, operational-expense-driven cloud resources or drawing on local systems that are optimised. Furthermore, during times of strain on the grid or periods of low demand, the footprint ‘collapses’ by shifting non-essential workloads to operate during lower-demand times and drawing on behind-the-meter battery storage to protect the regional power grid – this is known as dynamic peak shaving.
Grid-aware elasticity makes financial sense
Over-provisioning is a costly capacity move, whereas in an accordion infrastructure model, you’re leaning into OpEx flexibility because CTOs are then aligning costs with actual inference usage. With how constrained the grid is right now, it’s uncool to be high-volume energy hogs (and they’re being penalised by grid operators); grid operators are favouring and financially rewarding developers who fluctuate their workload through flexible infrastructure. The accordion model lowers utility costs and prevents regulatory friction. And in this current market condition? It just makes sense.
The most successful CTOs are obsessed with elasticity, not over-provisioning. Being ‘grid-ready’ doesn’t mean you have huge backup generators to survive an outage, or that they’re required to grow. True grid readiness is being nimble, and having infrastructure that can ‘breathe’ and adapt to demand so you utilise capacity better and avoid the risk of outages.
