Power first: grid and location decisions for AI data centres through 2027

The IEA's April 2026 report on energy and AI estimated that data centre electricity consumption grew 17% in 2025, and that consumption by AI-focused data centres grew 50%. Its central projection has data centre consumption roughly doubling from 485 TWh in 2025 to 950 TWh in 2030, about 3% of global electricity demand. Capital expenditure by the largest technology companies exceeded $400 billion in 2025 and is expected to rise by a further 75% in 2026.

The physical side is tightening. Between 2020 and 2025 the power density of AI servers rose elevenfold, and the IEA expects a further fourfold rise by 2027, when a single advanced rack could draw as much at peak as 65 households. Bottlenecks in power electronics, transformers, gas turbines and high-bandwidth memory make the IEA's more aggressive near-term scenarios less likely, and it expects the memory shortage to last at least until the end of 2027.

The IEA also notes that data centres fill with servers progressively and often oversize their grid connections at the start. The megawatts requested from a network operator are a forecast of demand several years ahead. Every regulator now handling a data centre queue is trying to separate firm forecasts from hopeful ones, and the tools they have chosen (fees, milestones, generation requirements and flexibility conditions) all put a price on the difference.

Contracted demand connection offers in Great Britain rose from 41 GW to 125 GW between November 2024 and June 2025, driven largely by data centres, which account for at least 80 GW of applications. On 29 July 2026 Ofgem consulted on a Data Centre Commitment Fee, payable when a large data centre accepts a connection offer, refunded at energisation and forfeited if the project leaves the queue early. The proposed range is £237,500 to £712,500 per megawatt, around 2.5% to 7.5% of average project costs. Projects would also have to meet data centre milestones, with evidence of financial capability, commercial maturity and procurement activity, to keep their place. The consultation closed in September 2026, and Ofgem's decision is pending.

The arithmetic is direct. Under the proposed range, a 100 MW request would put between £23.75 million and £71.25 million at risk. Requesting capacity well ahead of a realistic ramp cost little under the old queue. Under the proposal it becomes a balance-sheet decision. Developers will need to size requests to the deployment schedule they can finance, and investors will want that schedule tested against chip and electrical equipment lead times.

The wider connections reform runs alongside. NESO's revised timetable put Gate 2 Phase 2 transmission and large embedded offers between early September 2026 and mid-January 2027, and Phase 2 distribution offers between mid-October 2026 and mid-March 2027. The offer dates that arrive in that window will decide which British projects can commit capital with confidence in 2027.

Ireland shows where a constrained grid leads. Data centres used 22% of Irish electricity in 2024, up from 5% in 2015, and EirGrid forecasts their demand rising from 9.4 TWh in 2025 to 14.6 TWh in 2034, which would be 31% of national demand. In December 2025 the Commission for Regulation of Utilities decided that new data centres must provide dispatchable generation or storage, on site or nearby, matching their requested maximum import capacity and taking part in the wholesale market. They must also match at least 80% of their annual demand with additional renewable electricity generated in Ireland, within a six-year glide path and against a credible plan presented to the system operators, and the operators must consider whether the requested location is constrained.

The EU is starting with disclosure. Data centres with significant energy use already report energy and water performance under the Energy Efficiency Directive. On 21 September 2026 the Commission adopted a common rating scheme for data centres above 500 kW, which is now in a two-month scrutiny period before the European Parliament and the Council. It also opened a consultation on minimum performance standards that closes on 14 December 2026, with a legislative proposal planned for the second quarter of 2027. The first ratings are expected to grade 2026 performance, using data submitted in 2027.

Both jurisdictions point the same way. A connection now comes with obligations about how power is sourced, how flexibly it is used and how performance is disclosed, and those obligations belong in the investment case from the first model.

PJM's capacity auction for the 2028/2029 delivery year, held in July 2026, cleared at the price cap of $325 per MW-day across its footprint and procured 6,831 MW less than its reliability requirement. The previous auction fell about 6,500 MW short, and PJM says these were the first auctions in its history in which the whole region fell short. It points to the continued addition of large data centre loads to its load forecast. Its responses include a Connect and Manage framework, which would let large new loads connect but operate flexibly when needed, and, since 9 June 2026, help in arranging long-term bilateral contracts between large load customers and generators. The next auction, for 2029/2030, is scheduled for December 2026.

Federal regulation is moving as well. On 18 June 2026 FERC ordered all six regional grid operators under its jurisdiction to justify their tariffs for large loads or reform them, covering study processes, cost allocation and arrangements for co-location and behind-the-meter generation, with 60 days to respond.

Developers facing slow connections are turning to on-site gas. The IEA finds that reliable on-site gas supply for critical and variable data centre load needs 30% to 70% more generating capacity than the load itself. Gas turbine orders rose 70% in 2025, and the IEA finds it unclear whether on-site generation offers a faster route to market at scale.

The IEA's policy recommendations point to the trade on offer: non-firm connections and demand response in return for faster connection. A site that can reduce its draw when the grid is stressed, through batteries, on-site generation or workload scheduling, is the site an operator can connect sooner. How much interruption a project can accept is a design decision, and it belongs before land is bought.

Across these jurisdictions the order in which a site is chosen is reversing. The traditional sequence of land, planning and then power gives way to power first: the connection date, the share of capacity that will be firm and the share that will be interruptible, the fee or security needed to hold it, and the sourcing or generation obligations attached. Land, fibre, planning consent and tenant demand are then tested against that.

The financial model has to follow. Revenue ramps are usually tied to building completion and tenant fit-out. They should be tied to energisation dates by phase, with the likelihood of each date taken from the grid evidence rather than the construction program. A model with a single go-live date is hiding three: the date the building is ready, the date the network can supply power, and the date the tenant's equipment arrives. Through 2027 the second and third are the ones most likely to move.

Disclosure belongs in the same model. Where ratings, renewable-matching obligations or reporting duties apply, the data that proves compliance (metered energy, water use, renewable supply and flexibility events) should be designed into operations before the first tenant arrives, because it is far harder to reconstruct afterwards. The records that prove compliance in Ireland, the EU and the US overlap more than the rules do, so one metering and reporting layer, designed once, can serve several jurisdictions.

Before Ofgem decides on the commitment fee, list every British demand connection position held or applied for, set each against a deployment schedule that can be financed, and calculate the fee exposure at both ends of the proposed range. Decide which positions to hold, resize or release.

By January 2027, for projects in Gate 2 Phase 2, confirm the offer date and the evidence each milestone will demand, and align board approvals and financing conditions to those dates.

For Irish projects, test whether the dispatchable generation and 80% renewable requirements can be met with named assets and contracts before any further land commitment.

For EU sites, decide by 14 December 2026 whether to respond to the consultation on minimum performance standards, and check the 2026 energy and water data now, so the first rating reflects the site as it actually operates.

For US projects, decide how much interruptible load each one can accept before PJM's December 2026 auction and the tariff changes that follow FERC's orders, since flexibility is becoming the price of an earlier connection.

Across the portfolio, replace single go-live dates in financial models with energisation dates by phase and a range for equipment arrival, including memory and electrical equipment through 2027, and report that range to investors and lenders rather than a point estimate.