The Nordic grid and the AI bill

Norway is about to become the most important AI-governance story in Europe, and the reason has nothing to do with models. It has to do with the grid.

The country has approximately 34 GW of hydropower, a grid operator that has reserved 3.4 GW of capacity for new data centres, a named hyperscale build already under construction, and no published national methodology for evaluating whether any of that reservation earns its social licence. Climate policy and AI policy in Norway are now the same policy. Almost no one is treating them as one file.

The evidence

Global data-centre electricity demand rose 17 percent in 2025 against a global average electricity-demand growth of about 3 percent, and total data-centre power demand is projected to roughly double by 2030 ^[1].

Norway's position in that market is unusually strong. The country has around 34 GW of hydropower capacity and approximately 87 TWh of reservoir storage ^[2]. The Nordic grid is among the cleanest per kilowatt-hour anywhere, and reservoir storage means it is also among the most dispatchable, which matters more to a data-centre operator than the carbon number.

The commitment is already made. Statnett has reserved 3.4 GW of grid capacity for new data centres — more than 8 percent of installed hydro capacity ^[3].

The build is named and underway. Stargate Norway is 230 MW, expanding to 520 MW, running on hydropower ^[4].

And the domestic reaction has started. Local disputes over grid connection, water use, and land use around new sites are now regular items in the Norwegian regional press. The window for setting a national methodology before the politics hardens is short and closing.

What has not been published

There is no national methodology that lets a municipal council, a grid operator, and a citizen assembly answer the same question in a consistent way: what is the social value per megawatt-hour of compute allocated to a given siting.

Without that, every siting decision is a bilateral negotiation between a developer, the grid, and a municipality, conducted with asymmetric information and no shared reference point. Each one drifts from the last. A municipality that negotiates hard is described as obstructive. A municipality that negotiates softly is described as captured. Neither has a framework that would let it defend its decision to its own residents in terms anyone outside the room recognises.

This is a familiar failure. It is the same structure as early wind-power siting in Norway, which produced a decade of local conflict, a suspended licensing regime, and a durable loss of public consent that the sector has still not recovered. The lesson from that episode is that a methodology published late is worth less than a methodology published early and imperfectly.

What such a methodology would have to contain

Four things, at minimum.

A definition of the unit. Social value per megawatt-hour is a defensible unit because it makes the trade-off explicit: this electricity, in this place, for this purpose, instead of that one.

A treatment of alternatives. Every megawatt allocated to compute is a megawatt not allocated to electrification of industry, transport, or heating. The methodology has to state what the counterfactual allocation was and what it was worth.

A treatment of local externalities that are not carbon. Water, land, heat, noise, and the specific question of whether waste heat is recovered and where it goes. Carbon is the easy variable and it is not the one municipalities are arguing about.

And a disclosure schedule. What the operator publishes, at what interval, verified by whom. A methodology without a disclosure schedule is a scoring rubric no one ever fills in.

Why this matters for the Nordic AI-capital thesis

The Nordic AI investment case rests on Norway remaining a credible clean-compute jurisdiction. The sovereign fund's holdings in the AI supply chain, the estimates of what AI could add to Nordic GDP, and the recovery of Nordic venture funding to roughly USD 1.8 billion in the third quarter of 2025 all assume it.

Credibility here is not a technical property. It is a political one. If domestic social licence for grid allocation to compute erodes before the disclosure methodology exists, the damage lands on the investment case, and it is not damage the technical build can repair. Transmission capacity can be added. A withdrawn licensing consensus cannot be re-manufactured on the same timescale.

The bear case

If Statnett and the responsible ministry publish a national methodology for compute-siting social value by the end of 2026, the window holds and this issue was an early warning that turned out to be unnecessary. That is the good outcome and we would report it as such.

If they do not, then work under Pillar VI, Climate and Sustainability, becomes a post-mortem rather than an instrument. A post-mortem is a legitimate thing to publish. It is a poor substitute for arriving before the decisions were made.

What ImpactLab is doing

A working paper is in draft, under the provisional title "The compute clause: a disclosure methodology for AI-siting social licence".

Named civil-society and municipal co-authors will publish with the paper. We are seeking a municipality that has already negotiated a siting agreement to co-author the section on what the negotiation actually looked like from the municipal side, because that account does not currently exist in public.