OpenAI's voluntary grid pledge, and the 1.5% that load shifting buys

A non-binding promise to dial down

OpenAI calls the effort an "ambition" rather than set policy, Data Center Dynamics reported. The arrangements will differ by location. Whether OpenAI's commitments become contracts is unknown.

For utilities, regulators and host towns, the fight is now about who pays for grid capacity and how fast it can be built. A voluntary pledge costs a hyperscaler little and binds nobody. The company said US data center capacity is projected to exceed 100 GW by 2030. It also said it will provide up to 1 GW of flexible demand response. Under its Georgia Power deal, OpenAI has committed to paying the full cost of the infrastructure required to serve its data center.

Ember: shifting workloads buys about 1.5%

Ember's analysis covers nine markets in the US, Europe and Southeast Asia. It was published in October 2026 and reported by pv magazine on 3 November.

In Ireland, data centers are about 19% of peak demand. In January 2025, when system demand peaked at 6 GW, they accounted for an estimated 1.1 GW. Ember projects that share rising to 24% by 2030, with demand at 1.6 GW against a 6.5 GW peak. Load shifting could cut Ireland's peak by roughly 100 MW by 2030, or 1.5%. Ember assumes 35% flexibility for AI training, 5% for inference and none for traditional data centers. That leaves about 6.5% of global capacity theoretically shiftable, falling to around 3% after a 48% utilization rate.

Other markets range from 0.1% to 0.6%. The 1.5% figure is Ember's estimate for Ireland.

The order book: switchgear and cable

Money is moving where contracts are binding. GE Vernova agreed on 29 October 2026 to buy Trench Group for $9bn.

The cash and debt funded deal brings high voltage grid equipment manufacturing in house and adds a base in Europe and Asia to its Electrification segment. Closing is expected in late 2027. On 3 November, NKT said it had won three framework agreements with Danish distribution system operator N1 for low and medium voltage cables up to 72 kV. They run six years and cover an expected 700 to 1,000 km of cable a year. N1 serves more than 800,000 homes and businesses over a network exceeding 56,000 km.

Teesside: 3.2 GW and a queue

In the UK the constraint shows up in land deals. National Grid agreed to buy 29.5 hectares (73 acres) at Dorman Point on land owned by Teesworks.

The Guardian reported the deal on 1 November 2026; Sky News first reported it. Planning applications cover nearly 500,000 sq metres. Ben Houchen, mayor of the Tees Valley combined authority, told Sky News that up to 3.2 GW of capacity at the site would let the region compete for AI projects. That is several times the 200 MW to 500 MW a single large data center needs. The government published a draft national policy statement on data centers last week, describing them as fundamental to the economy as the railways and electricity networks. Campaign groups have protested plans from Auchtertool, near Edinburgh, to Brick Lane in east London.

What is still unresolved

The timing gap explains the pressure. A large data center can need 200 MW to 500 MW and be built in 18 to 24 months.

Conventional grid expansion takes four to 10 years, according to Ember. Behind-the-meter generation and storage could deliver nearly five times more relief, with potential reductions of up to 2.7 GW in Texas, Ember found. That route depends on permits, and on equipment that is itself in a queue. OpenAI said it will share additional commitments as they are finalized. Until then the program is a statement of intent.