A silent surge in AI computing had already begun to redraw the UK’s electricity map, and the numbers now on the table suggested that misjudging its carbon cost was no longer a rounding error but a policy liability with national stakes. Last summer’s “Compute evidence annex” from the Department for Science, Innovation & Technology projected vanishingly small emissions for AI compute by 2035—between 0.025 and 0.142 MtCO2 per year, a sliver under 0.05% of national emissions—only for the government to withdraw the document and, last week, publish a correction that reframed the outlook as a 10‑year cumulative 34 to 123 MtCO2. That step-change, roughly two orders of magnitude, arose from mixing annual and multi‑year figures. The recalibration, while overdue, still left the central question unresolved: would the grid powering these facilities be clean enough, and fast enough, to keep net‑zero goals intact as hyperscale buildouts accelerated?
The Revision: What Changed and Why
The revised DSIT range positioned AI compute as a potentially material slice of national emissions—about 0.9% to 3.4% over the coming decade—rather than a statistical footnote. However, the credibility of the new outlook hinged on one pivotal lever: electricity carbon intensity. DSIT assumed a future factor around 50 gCO2/kWh by 2030, a level compatible with a grid that sources nearly all power from wind, solar, nuclear, and interconnectors. Carbon Brief countered that such a rosy assumption would unravel if natural gas had to backstop supply more often, especially during wind lulls and winter peaks. If gas filled large gaps, intensity could climb by an order of magnitude, pushing data center emissions dramatically higher than the revised baseline.
This sensitivity played out vividly in alternative scenarios anchored to Ofgem-linked expectations that future data center load could reach 20 GW. Under a low‑gas case—just 5% of electricity from gas—Carbon Brief’s modeling put annual emissions near 3.4 MtCO2. At the other end, with 95% gas, the figure rose to 68.1 MtCO2, rivaling the yearly footprint of a small European nation. The spread underscored that modeling the data centers in isolation made little sense; the broader grid mix, the timing of clean generation additions, and the volume of firm zero‑carbon power would set the bounds. Put differently, the DSIT correction solved a bookkeeping error but did not answer the operational one: how the UK would keep marginal megawatt‑hours clean as AI clusters scaled and connection queues swelled.
The Stakes: Power Mix, Capacity, and Policy
Scale compounded the challenge. The UK today hosted about 1.6 GW of data center capacity, but more than 8 GW was in planning or under construction, with Ofgem’s connection outlook indicating eventual loads as high as 20 GW. That pipeline reflected hyperscalers racing to deploy AI training and inference capacity, often favoring power‑dense campuses near major substations. Yet siting to chase electrical headroom risked steering growth toward regions still reliant on gas peakers and constrained transmission. Campaign group Foxglove argued that the government’s arithmetic glossed over these constraints while green‑lighting a buildout that could lift national electricity consumption sharply even as legal obligations to reach net zero by 2050 remained in force. That critique converged with a broader call for transparent, scenario‑based planning that stress‑tested assumptions about grid intensity and timing.
Governance gaps had already been exposed by the DSIT misstep, and the path forward depended on treating AI capacity as an integrated energy project rather than just digital infrastructure. Effective levers existed: binding power purchase agreements tied to verifiable new-build renewables and nuclear; priority connections contingent on demonstrated clean supply; co‑location with curtailed offshore wind in the North Sea zone; and demand‑side measures like flexible training schedules to avoid peak‑gas hours. Grid upgrades—transformer additions, reinforcement, and storage—could smooth variability. Heat reuse standards could turn waste heat into urban heating assets. Emissions reporting needed to separate annual from cumulative figures and disclose marginal intensity at the hour level. Financing and permitting reforms could align the Capacity Market and Contracts for Difference with round‑the‑clock, zero‑carbon supply, while Ofgem’s connections process could gate approvals on credible decarbonization plans. Taken together, these steps had offered a route to expand AI without derailing climate targets.