The Watt Ceiling
By Sparse Bob • CTO, Sparse Supernova • March 2026
I’ve been watching the AI boom with two sets of eyes: the builder’s eye, and the energy eye. The builder sees the speed of progress. The energy eye sees the bill — and the bottlenecks — arriving in real time.
Frontier AI is being sold as an unlimited race: scale the clusters, scale the models, scale the revenue. The physical world doesn’t work like that. Power delivery, cooling, and water don’t expand on a quarterly schedule. Planning and grid reinforcement don’t happen because a valuation needs them to happen. They happen when steel is ordered, permits are approved, transformers are delivered, and local networks can carry the load.
That’s the watt ceiling: a hard set of constraints that decides what can be built, where it can be built, and how fast it can operate. Once you accept that, another conclusion follows quickly: there isn’t enough infrastructure headroom for every frontier AI company to scale the way their investment story assumes.
Energy is the strategy. Everything else has to fit inside it.
Data centres can be constructed quickly. Grid capacity can’t. Transmission upgrades, substations, connection agreements, and transformer lead times move slowly. The supply chain behind them is stretched, and the queues are real. In some regions, the response is already visible: connection restrictions, planning pushback, and “bring firm power” requirements that change the economics overnight.
The risk concentrates geographically. Compute wants the same corridors: dense connectivity, favourable regulation, a deep talent pool, predictable logistics. That creates hotspots. When hotspots hit grid limits, the whole scaling narrative gets fragile — because the industry isn’t evenly distributed, it’s clustered.
Infrastructure determines outcomes
Recent benchmarking across 30 models confirms what the territorial data already suggested: the same model deployed on different data centre infrastructure can produce 70–85% less energy consumption, water use, and carbon emissions. The stack matters as much as the architecture. Where you build, and on what hardware, is an environmental decision of the first order — not an afterthought.
Water and public consent
Cooling makes this even more politically charged. Cooling pulls water, and water is already tight in many places. Communities notice when a new build looks like it will take capacity away from homes, hospitals, and industry. Public consent becomes part of the infrastructure plan, whether the industry likes it or not.
The return question
Frontier AI is capital-heavy and refresh-heavy. Hardware ages fast. Requirements shift fast. If your model depends on endless expansion, grid constraints turn into a direct threat to payback. You can have world-class engineering and still lose money if the physical rollout stalls, costs spike, or utilisation assumptions don’t hold.
Efficiency as engineering, not branding
This is why we take efficiency seriously — as engineering, not branding. Less compute per outcome. Less data moved per outcome. Better routing. Smaller payloads. These are the moves that help AI fit inside real-world limits rather than pretending the limits aren’t there.
I’m not arguing for slowing down. I’m arguing for growing up: treating energy, water, siting, and carbon as first-class design inputs. The teams that win will have systems that run inside the constraints, with numbers they’re willing to put on the table.
Comment / Respond
No accounts, no tracking, no comment platform. This opens an email to our operations team.