Energy, power and placement: routing AI work to capacity

For most of the last three years, the limit on AI capacity was chips. Increasingly it is power: the electricity to run accelerators, the grid connection to deliver it and the years it takes to get one. That changes where AI work can run and what the most useful infrastructure investments are. It also turns placement, meaning the decision about which site serves which request, into an energy question.

This article sets out the evidence on data-centre electricity in the US, EU and MENA, and the case for routing flexible AI work to where power and capacity already exist.

The numbers that frame the problem

The International Energy Agency's Energy and AI report estimates that data centres used around 415 TWh in 2024, about 1.5% of global electricity. It projects that consumption will more than double to around 945 TWh by 2030. The United States accounts for by far the largest share of that growth, followed by China.

In the US, a Department of Energy report from Lawrence Berkeley National Laboratory found that data centres used about 4.4% of US electricity in 2023, or 176 TWh. It projects 6.7% to 12% by 2028, or 325 to 580 TWh. The wide range itself tells you something: nobody knows how fast AI demand will grow, so grid planners have to plan for a broad range of outcomes.

Grid connections are where the pressure shows first. The IEA notes that connection queues for supply and demand projects, including data centres, are long and complex. It estimates that around 20% of planned data-centre projects could be delayed unless grid risks are addressed. It also observes that about half of the US data centres under development are in existing large clusters, which risks local bottlenecks.

Two recommendations worth reading closely

Two sentences in the IEA report matter for anyone who routes AI work. The report suggests regulators "could explore measures to incentivise data centre operators to use spare server capacity". It also says grid operators "could also examine incentives to locate data centres in areas where grids are less constrained".

The first is about utilisation. Capacity that is already built and connected but sitting idle is the cheapest capacity there is, measured in energy, grid connections and time. Inference demand follows a daily cycle. The Aryl study of a production inference cluster found a peak-to-trough ratio of about 2.2 and average utilisation of about 65%. Every hour that an idle, already-powered site serves work that would otherwise need new capacity elsewhere is an hour of demand that does not need a new grid connection.

The second is about geography. New capacity should go where the grid can take it, and work should be able to reach that capacity wherever it is.

Flexible and inflexible AI work

Not all AI work can move. It helps to sort it into three classes:

• Pinned work must run in a specific jurisdiction or close to a user: regulated data, interactive voice, control loops. It can move only within its constraint.
• Interactive work tolerates tens of milliseconds of added latency, which on fibre covers distances of many hundreds of kilometres. Most chat and agent traffic falls here. It can move between nearby regions.
• Deferrable work has no tight deadline: batch embedding, evaluation runs, document processing, synthetic data generation. It can run wherever and whenever capacity is available.

Only the first class is truly fixed. The other two can follow spare capacity, and in practice spare capacity is where power and connections already exist. Routing that knows about each site's live load and region can place flexible work there without any change to the application.

The EU: reporting comes first

The EU has moved from principle to measurement. Under the recast Energy Efficiency Directive, Commission Delegated Regulation (EU) 2024/1364 sets up the first phase of a common rating scheme for data centres. Operators with at least 500 kW of installed IT power demand must report energy and sustainability indicators to a European database, as DLA Piper summarises. A second delegated act on rating and labelling data centres was in draft in 2026.

Reporting sets up the next policy step: rewarding sites that do more useful work per unit of energy. Utilisation is central to that. A site with high efficiency ratings that sits half empty off-peak still wastes much of the energy embodied in its build and connection.

MENA: building where power is plentiful

The Gulf states are betting on the IEA's second point at scale: build where energy and land are plentiful, then serve demand from there. In May 2025 the partners behind Stargate UAE announced a 1 GW compute cluster within a planned 5 GW UAE–US AI campus in Abu Dhabi, with a first 200 MW phase targeted for 2026. Capacity like this is only fully used if work from elsewhere can reach it, within the latency and jurisdiction limits that apply to it.

What this means for operators and policy makers

For data-centre operators, power is the scarce asset, and capacity already connected to the grid is worth more than capacity waiting in a queue. Make spare hours sellable to flexible demand, and publish enough signals, such as region, live load and warm models, for routing to find you.

For EU and national policy makers, the IEA's suggestions translate into practical measures. Reward measured utilisation, not just efficiency ratios. Encourage open, interoperable ways to route flexible work to idle capacity. And treat jurisdiction pins as the mechanism that lets regulated work stay put while everything else can move.

For US planners, the wide range in demand projections is an argument for flexibility. Capacity that can serve many buyers through routing is easier to use fully than capacity tied to one tenant.

Where Tenzro fits

Tenzro is an open network where machines sell AI compute and inference. It is live in preview, and governance decides the steps to testnet and mainnet. To be clear about scope: the network does not today read energy, carbon or grid signals, and it makes no energy claims. What is built is the placement layer that such signals could feed:

• Region and load signals. Providers declare a coarse region, at country or wide-region level, and send signed heartbeats with per-model load, warm-up state and queue depth.
• Load-responsive rates. Compute rental can track smoothed utilisation against a 50% target, so rates ease when capacity is idle and flexible work has a reason to go there.
• Observed performance. A price and latency index is built from signed measurements by validators, not from providers' own announcements.
• Constraints that hold. Jurisdiction pins fail closed, so regulated work stays where it must while unpinned work routes freely.
• Small machines for small steps. The light build runs identity, payments and agents on low-power edge devices and forwards model calls to providers. Hosting the cheap steps of model ladders on such devices is planned.