On 10 June Gartner put a number on what operators have been saying quietly for a year: data-centre electricity use grows 26% in 2026, and AI capacity is now constrained by power availability, not by chips. The hyperscalers can buy the GPUs. What they cannot buy is a grid connection, and in many jurisdictions the wait for one is five to ten years. For a New Zealand enterprise that means the capacity you want, in the region you want, may not be there when you want it, and the price you are quoted will reflect that. The cloud is not an infinite tap this year, and your plan should say so.
What you need to know
- The bottleneck has moved. Gartner forecasts global data-centre electricity consumption rising from 447TWh in 2025 to 565TWh in 2026, with AI-optimised servers taking 31% of that power this year and passing conventional servers in 2027.
- The build-out is outrunning the wires. Goldman Sachs expects US data-centre power demand to go from 31GW in 2025 to 66GW in 2027. Grid connections in many jurisdictions queue for five to ten years, and in the largest US grid a project entering service in 2025 waited more than seven years on average.
- Operators are going around the grid. Bloom Energy's 2026 survey found roughly one-third of data centres expect to run entirely on onsite power by 2030, and the IEA counts 15 to 27GW of onsite gas that may power data centres by the same year.
- Scarcity is already in the price. AWS raised its reserved GPU capacity prices 15% in January and said out loud that the pricing follows supply and demand.
- For you the question is narrower: which region your AI runs in, whether you can move it, and how much inference you burn per unit of work. Those three are in your control. The queue is not.
26%
Forecast growth in global data-centre electricity consumption in 2026, from 447TWh to 565TWh
Source: Gartner, June 2026
66GW
Forecast US data-centre power demand in 2027, up from 31GW in 2025
Source: Goldman Sachs Research, May 2026
5-10 years
Grid-connection waiting times in many jurisdictions
Source: IEA, April 2026
The constraint moved from silicon to substations
The story of 2024 was GPU supply. The story of 2026 is the substation, and you cannot fab one. Gartner's June forecast has AI-optimised servers consuming 175TWh in 2026, up 84% on the year before, and Gartner director analyst Linglan Wang says plainly that AI capacity is now constrained by power availability.
The demand side is easy to state. Goldman Sachs Research expects US data-centre power demand to more than double in two years, from 31GW in 2025 to 66GW in 2027, taking data centres to 8.5% of US peak summer demand. The supply side is where the years go. The IEA's April update puts grid-connection waits at five to ten years in many jurisdictions, and notes that wait times for transformers and cables have doubled in three years. In PJM, the largest US grid operator, a project that entered service in 2025 had spent more than three years reaching an interconnection agreement and another four waiting to energise after approval. Substation transformers are quoting lead times over 160 weeks.
None of that is a New Zealand problem in the direct sense. It is a New Zealand problem because the regions you buy compute from sit on those grids, and scarce capacity is priced as scarce.
Operators are routing around the grid, and charging for it
The response from the industry is not to wait. Bloom Energy's 2026 power report, from a survey of 152 decision-makers taken in November, found roughly one-third of data centres expect to run on 100% onsite power by 2030, and that utilities' own delivery timelines run about one and a half to two years longer than the developers planning around them expect. The IEA sees the same thing from satellite imagery: constrained by slow grid connections, US developers are building onsite gas, and around a fifth of those projects have already broken ground. "Bring your own power" has gone from a workaround to a design assumption.
That capital shows up in your invoice. On a Saturday in January, AWS lifted the price of its reserved H200 capacity blocks by 15% with no announcement, and when asked, said the pricing "reflects the supply/demand patterns we expect this quarter." Read that as a policy, not a one-off. When a provider can sell every accelerator it can power, the price of guaranteed capacity floats up to whatever the market clears at, and the region with spare power is the one with the shorter queue, not the one closest to you.
What this does to a NZ buyer's options
Aotearoa has a hyperscale region in Auckland and a consented AI campus in Southland. Who pays for that build-out, and who benefits, is its own question. The point here is what the global squeeze does to your choices, and it comes down to three things.
Region availability. The closest region is not automatically the one with capacity. By early July, Transpower's connection dashboard listed 30 data-centre enquiries, sixteen of them in Auckland, with Datagrid's 280MW Invercargill connection still in Transpower's investigation phase. Local capacity is coming, but on the grid's timetable, which is years. In the meantime the accelerators you want may be in Sydney, in Oregon, or on a waitlist.
Price. If your business case assumed the per-token or per-hour rate you saw last year holding, it is already out of date, and a data-residency requirement narrows the set of regions you can shop across, which narrows your bargaining power further.
Lead time. This is the one that catches operations teams. A "cloud" decision used to mean an afternoon. A reserved-capacity decision in a constrained region now means a lead time you have to plan around like hardware, because underneath it is hardware, waiting on a substation.
I have spent a lot of my working life waiting on things that could not be hurried: consents, contractors, a rebuild after an earthquake. Power is that kind of thing. You cannot escalate a substation. So the operational discipline is the same one we used then. Know your lead times honestly, decide where you are going to stand before the queue forms, and design so you can step sideways if the ground moves.
Tim Hatherley-GreeneChief Operating OfficerWhat to actually do
Decide where your AI runs on purpose. Pick the region, write down why, and confirm the specific capacity you need is available on the dates you need it, not just listed. If data residency confines you to one region, know that, and plan reservations early rather than assuming on-demand will be there. This is the same decision as the sovereignty one, made with a calendar in hand.
Design for portability so the region is a choice, not a lock-in. Keep your models behind one interface, keep prompts and evaluation sets as assets you own, and keep the model and the region in configuration. It is the same discipline that makes a new model a config change, applied to where the workload lives. A team that can move inference from one region to another in a week has bargaining power on price and on availability. A team that cannot is a price taker. For work that genuinely cannot leave the building, your own hardware on your own power removes the queue from the equation entirely, which is the case where we build sovereign AI on-premises.
Treat inference efficiency as a cost lever, because it is now a capacity lever too. Every token you do not generate is capacity you do not need to find. Route routine work to smaller models, cache what repeats, batch what can wait, and measure cost per completed task rather than cost per hour. The cheapest GPU in 2026 is the one you did not need to reserve.