Data centres are almost always described as a problem for the grid: vast, fast-growing, always-on loads. But that framing misses their most valuable and least-discussed property — much of that load can move.
Not all data-centre load is rigid. A large share — AI training, batch processing, and workloads that can move between sites — can shift in time or location without harming the service. Treated as flexible, data centres can soak up surplus renewables, back off at peak, and provide grid services, turning the "data-centre problem" into one of the grid's most valuable controllable resources.
The standard story about data centres and the grid is one of strain. They consume enormous amounts of electricity, that consumption is growing fast, and — the assumption goes — it’s rigidly always-on, so the grid simply has to supply whatever they demand, whenever they demand it. That assumption is the problem, because it’s only partly true.
A surprising amount of what a data centre does is not time-critical. And anything that isn’t time-critical is, in principle, flexible — which changes the picture entirely.
It helps to separate two kinds of workload. Some is latency-sensitive and must run now: serving a live web request, a video stream, a real-time inference query. You can’t delay it without users noticing. But a large and growing share is not like that:
For all of these, when and where the computation happens is flexible. And flexibility, on a grid full of variable renewables, is extraordinarily valuable.
Once you treat a chunk of data-centre demand as shiftable, the relationship with the grid inverts. Instead of a load that must be served at all costs, the data centre becomes a resource that can:
A data centre that can move its workload in time is no longer just the grid’s biggest customer. It’s one of its most useful instruments.
If the load is flexible, why is it so rarely treated that way? Partly habit — data centres are engineered for maximum uptime and utilisation, and “run everything as soon as possible” is the default. Partly incentives — operators only flex if there’s a clear commercial reason and a market that rewards it. And partly difficulty: flexing workloads against grid conditions, electricity prices and carbon intensity, without breaking service-level commitments, is a genuinely hard scheduling and control problem. It requires knowing what can move, predicting grid and price conditions, and orchestrating compute accordingly — automatically and continuously.
That difficulty is precisely why the flexibility stays “silent.” The capability exists; unlocking it is an engineering and market-design challenge, not a physics one.
Turning a data centre’s latent flexibility into real value means classifying workloads, forecasting grid and market conditions, and scheduling compute against them without breaching service guarantees — exactly the kind of cross-disciplinary problem we work on, and the flip side of the data-centre demand question. If you operate large, controllable loads, let’s talk about making them flexible assets.
Turning latent flexibility into value means classifying workloads, forecasting grid conditions, and scheduling compute against them without breaching service guarantees. That's the kind of problem we solve.