Power is the new capacity limit

The market still talks about AI capacity in chips: how many accelerators were shipped, who has allocation, what the next generation will do. That framing is a decade out of date in its instincts. A chip with nowhere to plug in is inventory. The capacity that matters is electricity that has been secured at a specific site, delivered to a specific rack, removed as heat, and measured well enough that someone can promise it to a customer.

Call it placed, measured watts. It is now the scarce resource, and it behaves very differently from silicon.

The demand is real and concentrated

The International Energy Agency's Energy and AI report, published in April 2025, put data centres at around 1.5% of global electricity consumption in 2024, or 415 TWh. Its base case has that figure more than doubling to around 945 TWh by 2030, slightly more than Japan's total electricity consumption today, and rising to around 1,200 TWh by 2035. The IEA expects electricity demand from AI-optimised data centres alone to more than quadruple by 2030 (IEA).

The growth is not spread evenly. The United States and China account for most of it; in the US, data centres make up nearly half of electricity demand growth to 2030. Europe held 15% of global data centre consumption in 2024 and China 25%. And the IEA states the point this article is about plainly: around 20% of planned data centre projects could be at risk of delay, and wait times for transformers and cables have doubled in three years.

Chips arrive on a product cycle measured in months. Grid connections arrive on an infrastructure cycle measured in years. When the two fall out of step, the slower one sets the pace.

Every major grid is rationing connections

On both sides of the Atlantic and the Irish Sea, the connection queue has become the real allocation mechanism.

In the United States, Berkeley Lab's Queued Up 2025 edition counted nearly 2,300 GW of generation and storage seeking transmission interconnection at the end of 2024, across about 10,300 projects. That is supply waiting to reach the grid, not demand, but it is the supply new data centres are counting on.

In Great Britain, the national system operator describes a connections queue of over 738 GW against the 200 to 225 GW of clean generation needed by 2030, and Ofgem approved a reform in April 2025 that replaces first come, first served with a gated queue for projects that are ready and needed (NESO).

In Ireland, the Commission for Regulation of Utilities decided on 12 December 2025 that new data centres must bring generation or storage, onsite or nearby, to match their requested demand, and that this capacity must take part in the wholesale market. System operators must also consider whether a request sits in a constrained location (CRU).

Germany's four transmission operators have applied a maturity-based procedure since 1 April 2026, giving scarce capacity to the most realisable projects rather than the earliest applicants. The European Commission's December 2025 notice on grid connections recommends the same direction for all member states: first ready, first served, "use it or lose it" milestones and reservation fees, though it is non-binding guidance (Freshfields).

The direction across these regimes is the same. Regulators are no longer asking who asked first. They are asking who is ready, where, and what the load will actually do to the system.

The rack has become a power plant problem

Inside the building, the same shift is happening at smaller scale. NVIDIA announced in May 2025 that it is leading a transition to 800 V DC data centre power to support 1 MW IT racks and beyond, starting in 2027, because the older 54 V in-rack design was built for kilowatt-scale racks (NVIDIA). At those densities air cannot carry the heat away; liquid has to.

Liquid cooling and high-voltage distribution are not upgrades you add to a hall. They decide which halls can take the hardware at all, how much of the connected power reaches compute rather than chillers, and how quickly a site can change what it runs. A facility's useful capacity is now the product of its connection, its distribution and its cooling, and the weakest of the three wins.

Why measured watts beat nominal ones

Here is the part the chip framing misses. A grid connection is a promise of maximum import. What a site can actually sell is lower and variable: it depends on the thermal headroom at that hour, the efficiency of each rack, how the load is spread, and what the grid operator now expects in return for the connection. Ireland's requirement for matching generation, Britain's readiness tests and Germany's maturity ranking all reward an operator that can say exactly how much power it will draw, when, and what it can shed.

That is a measurement and forecasting problem before it is a procurement problem. Operators who meter at the rack, forecast their own load and place work where power and cooling are available will get more useful compute out of the same connection. They will also be the ones regulators trust with the next one.

What this means for executives

Stop treating GPU allocation as the strategic constraint. Ask instead where your compute will be powered, under what connection terms, and how confident your provider is in the numbers. A contract for accelerators without a credible account of power and cooling behind it is a contract for a delivery date, not for capacity. Ownership matters too: when connection slots are rationed, the operator holding the slot holds the leverage.

What this means for policymakers

Connection reform is working in the right direction, and the regimes now in place in Britain, Ireland and Germany share a principle worth generalising: allocate grid capacity to demonstrated readiness and measured behaviour, not to queue position. Requiring operators to publish metered consumption and flexibility, not just nameplate requests, would make those tests sharper and would let smaller operators compete on efficiency rather than on how early they filed.

What this means for CTOs

Put measurement into the architecture. Know the power draw and thermal state of every machine you run, forecast it, and let placement decisions read those numbers. Plan for liquid cooling and higher distribution voltages in any new deployment, and treat power and thermal limits as scheduling inputs, in the same class as memory and network.

Where Praecise fits

Praecise Metal provides owned machines, placed, powered and cooled to measured targets: measurement, forecasting, placement, power and thermal control, on Praecise Furcate. Praecise runs its own machines on it and builds the same capability on metal its customers own. Core clusters are measured and operated end to end for placement, power, thermal and cost, and infrastructure owners can take GPU capacity on Praecise Metal. A single ownership view of inventory, health, power and cost across every machine and cell is part of Metal; nested cells and scoped managers within it are coming.