The AI build-out is constrained by grid connections, not ambition. There is usable electrical capacity already connected to buildings that will never draw it — and a growing number of people working out how to sell it.
What this covers:
- Why the AI build-out is now limited by grid connections rather than by compute, and what that changes
- Why there is unused electrical capacity already connected to buildings that will never draw it
- What must be true of your asset data before any of it works commercially
The compute market is short of power, not ambition. There is unused electrical capacity already connected to homes, small commercial units, telephone exchanges, stadiums, universities and thousands of other buildings — allocated to them, but never drawn. Collectively, this represents meaningful capacity that could support compute, years before new 100MW and gigawatt-scale campuses receive grid connections.
This is not a replacement for future GW campuses, it’s an opportunity that already exists.
SPAN’s XFRA and NVIDIA’s AI Grid both point towards the same conclusion: the future unit of deployment is not always a hyperscale campus, it could be spare power located next to good fibre connectivity.
Speed is the whole point. A new 100MW facility is a multi-year programme before it draws its first watt, and the connection queue is the longest part of it. Capacity that already exists behind an existing connection can be brought into service in a fraction of that time.
Why the headroom exists
One of the reasons this opportunity exists is because electricity demand falls dramatically overnight. Coal-fired (and to a lesser degree, gas) power stations cannot simply be switched off for a few hours because slowing and restarting turbines takes many hours and, in the case of coal, days. Whilst gas-fired stations are more flexible in that regard, there is an increased wear on expensive turbine components. As a result, many conventional generators continue producing electricity during periods of low demand. This means plenty of energy, with relatively few buyers, and storing that surplus energy is challenging. Grid-scale battery storage remains expensive and impractical: lithium-ion systems introduce operational and fire safety concerns, and other battery technologies simply don’t have the density required — sodium-ion, for example, offers safety advantages but currently lacks the energy density to make very large deployments economically attractive.
Where the spare capacity is
I would push the logic further. Well priced power that would otherwise go to waste, coupled with good fibre, the type of building matters far less. The data centre is not disappearing, but houses, supermarkets, stadiums, industrial estates, hospitals, universities and telephone exchanges may all have usable electrical headroom that could be utilised.
The test is a simple one. Unused power, and a good fibre connection. Meet both and a building is a candidate to form part of the AI grid, regardless of what it was originally built for.
Bi-directional power is already being tested
Vehicle-to-grid (V2G) is another interesting concept, it aims to use millions of EV batteries as distributed storage, charging overnight and exporting energy back during peak demand. This spreads storage costs across vehicle owners and creates enormous aggregate capacity. It’s been talked about for years, though in practice, widespread participation has yet to materialise because the standards, bi-directional technology, regulation and tariffs weren’t there, and customers choosing mobility first.
Likewise, a similar thing is starting to happen in data centres, albeit slowly. Ireland seems to be leading the way, with bi-directional flow or BYOP (Bring Your Own Power). It is now regulation rather than experiment there: under the Commission for Regulation of Utilities decision of December 2025, any new data centre seeking a grid connection must install on-site generation or storage matching its full demand, and must be able to return power to the national grid when called upon. That policy replaced the de facto Dublin connection moratorium in place since 2021. These facilities can draw large amounts of electricity when surplus exists and potentially return power to the grid via onsite generation or storage, during periods of high demand. All great in principle, but a new risk profile for the data centre owner/operator.
Ten thousand sites is a different problem
None of this removes the operational challenge. A single 100MW campus is one site, with everything contained within it. Ten thousand distributed nodes become a very big asset management and potentially a security problem. Every location needs an accurate, live record of installed equipment, temperature, power draw, connectivity, maintenance history, age, condition, book value and ownership. Without trusted asset data, the economics of distributed infrastructure quickly fall apart.
Where Assetspire fits
This is where Assetspire fits. Spire™ maintains a live record of physical assets, their location, power characteristics, connections and maintenance history, updated directly by engineers in the field. That capability already supports distributed telecoms and colocation estates and naturally extends to a distributed AI infrastructure.
For an estate of this shape, that record must be a single pane of glass: one view across every node, every site and every owner, whoever installed the equipment and whoever hosts it. And it must carry more than the register. The same view needs to support the monitoring, maintenance and day-to-day management of the estate, because at ten thousand sites nobody is going to send an engineer out to answer a question that should already have an answer. That is what makes the scale, the complexity and the commercial demands of distributed compute manageable rather than theoretical.
Finding the power is no longer the hard part
Finally, this should not be viewed as replacing gigawatt-scale AI campuses. Those facilities remain essential for training frontier models. The opportunity is in inference and other edge workloads, where spare electrical capacity already allocated to homes and businesses could be used productively. In some cases, GPUs installed in residential or commercial premises could even reject waste heat into building heating systems, improving the overall utilisation of both energy and infrastructure. The challenge is no longer simply finding power; it is identifying available headroom, proving it exists and managing it confidently at scale.
Steve Beber is the CEO at Assetspire. If you would like to continue the conversation, he is available on LinkedIn.