The race to build artificial intelligence has a physical address. Behind every model release is a data centre that needs a dependable connection to electricity, cooling equipment, networking and, increasingly, a credible plan for how to expand. The cost of computing hardware still matters. Yet a company can acquire chips and land and discover that the local power system cannot supply the project on the expected timetable. For investors, operators and communities, the question is shifting from how much AI capacity can be announced to how much can actually be connected and used.
That distinction is central to the International Energy Agency’s Energy and AI report. Its base case projects global electricity generation supplying data centres to rise from roughly 460 terawatt-hours in 2024 to more than 1,000 terawatt-hours in 2030. This is a scenario, not an inevitable outcome: improvements in hardware efficiency, demand and the timing of grid upgrades could change it. But even a smaller increase would create local pressure where many facilities seek capacity from the same network.
The connection queue changes the business case
An electricity contract is only part of a workable plan. A facility also needs transmission and distribution capacity, suitable substations, equipment, approvals and an agreement about who pays for upgrades. A site can have access to land and a power purchase agreement while still facing a long wait for physical delivery. Those delays affect when computers can start generating revenue and how much financing must be carried before a project is fully operating.
Local conditions matter more than a national forecast suggests. A country may have enough electricity in aggregate, but a particular cluster may lack spare grid capacity. The US Department of Energy’s assessment of data centre demand highlights the rapid growth in electricity use, while the underlying Lawrence Berkeley National Laboratory study presents a range of future outcomes rather than a single definitive projection. A useful project assessment therefore starts with a specific location, deliverable capacity and construction schedule.
Larger customers may seek dedicated connections or on-site generation. Neither option makes the wider system irrelevant. A data centre that turns to its own generators still needs fuel supply, operating permits and arrangements for backup. One that procures electricity from a remote wind or solar project needs transmission and a plan for the hours when that project produces less. The strength of the business case depends on the entire chain, not a headline price per megawatt-hour.
Capacity has a timing problem
Developers sometimes compare the construction period for a data hall with the longer lead times for grid infrastructure. A data centre shell can be built before a transformer, transmission line or additional generating capacity is ready. That creates a mismatch between committed spending and usable output. Operators may phase installations, lease less capacity initially or locate equipment elsewhere while a connection is completed.
Equipment supply can add another bottleneck. Large transformers, switchgear and backup systems cannot always be procured at short notice. Skilled labour, permitting and coordination among utilities also influence delivery. The Federal Energy Regulatory Commission’s work on connecting large loads reflects the need to assess how exceptionally large customers interact with transmission planning. This is a system planning issue; it is not solved simply by one buyer offering to pay more.
There is also a commercial consequence for customers buying computing capacity. If a provider has promised a start date that depends on an unfinished grid upgrade, delays can affect product launches, model training or enterprise deployments. Clear milestones for power availability should sit alongside milestones for chips and software. Operators need to say whether capacity is reserved, contracted, under construction or already energised. Those words represent very different risk positions.
Who should bear the cost of expansion
When a new facility triggers network investment, a practical question follows: who funds it? Large customers may pay direct connection costs, but broader upgrades can serve other users too. Utilities and regulators must decide how costs are allocated, and how to protect households and existing businesses from paying for facilities that might arrive late or use less power than forecast.
The risk cuts in both directions. If an operator is required to finance every conceivable expansion in advance, an otherwise viable project may never proceed. If optimistic load forecasts are treated as guaranteed demand, ratepayers can be exposed to underused assets. Appropriate deposits, staged commitments and transparent capacity agreements can help balance these outcomes. They should be designed for the local market rather than copied wholesale from another jurisdiction.
The Electric Power Research Institute’s updated data centre scenarios underscore the uncertainty in the scale of potential US load. This uncertainty makes disciplined planning more valuable, not less. A financial model should show how returns change if delivery is delayed, utilisation is slower or power costs rise. A grid study should test both the expected load and credible alternatives.
More generation is not the same as usable generation
The composition of future supply matters. The IEA expects renewables to provide a large share of incremental global electricity for data centres, while gas and coal also contribute to near-term growth in its base case. Its energy supply analysis distinguishes the electricity physically used by centres from contractual claims about renewable procurement. This distinction is essential when judging costs and emissions.
Power purchase agreements can support construction of new clean generation, but buyers still need continuous service. Storage, flexible demand and grid connections can help bridge the gap between variable output and around-the-clock computing. Nuclear and advanced generation have attracted interest, but their development times and commercial risks vary widely. Counting proposed plants as readily available electricity would overstate the near-term position.
AI companies may also improve how efficiently they use electricity. More capable chips can perform a given task with less energy; smarter scheduling can shift some work to periods with more available capacity. Yet lower cost per task can encourage more use. Efficiency is a serious tool for managing demand, but an investor should test whether efficiency gains reduce total consumption in the relevant business, rather than assume that they will.
Reliability becomes a commercial variable
AI workloads are not all equally sensitive to interruptions. Some model training jobs can pause and restart; certain customer-facing services need high availability. Building every site for the most demanding workload can be expensive. Matching the service promise to backup arrangements, network redundancy and realistic outage assumptions can be more efficient.
Grid operators must protect reliability for the entire system. The North American reliability assessment presented to FERC highlights rising load uncertainty alongside generation retirements and the challenge of integrating large customers. A data centre’s private backup plan does not by itself answer how its connection affects the rest of the network during a peak or a fault.
That is why contracts should address curtailment, backup fuel, service levels and responsibility for interruption. Flexible loads can sometimes reduce system stress in exchange for a financial benefit. Where an application needs uninterrupted operation, the operator should be explicit about how that promise will be met and what it costs. Reliability is an operational design choice with financial consequences.
The investment test is becoming more concrete
Market discussion often focuses on how much money technology companies plan to spend. A more useful test asks what portion of that capital will turn into working, revenue-producing capacity on schedule. The evidence includes firm grid connection dates, utility agreements, staged construction plans and the price and availability of equipment. Forecasts for power use should be consistent with forecasts for occupancy and computing revenue.
For lenders, delays can lengthen the period before a facility produces cash. For equity investors, they may weaken returns even when demand for computing remains strong. For a local government, a promised project can bring jobs and tax receipts, but also require public decisions on land, water, electricity and cost sharing. Each party needs its own view of what is committed and what remains conditional.
The wider implication is measured rather than dramatic. Electricity does not place a single fixed ceiling on AI development. It determines where and how quickly capacity can grow, and at what cost. Businesses that treat grid readiness as a core part of technology planning will have a more credible growth strategy than those that treat it as a detail to resolve after construction begins.
Location may become a competitive advantage
The search for electricity can change where companies build. A site close to customers may offer lower network latency and an established labour pool, yet face a costly or uncertain grid upgrade. A more distant location may have spare generating capacity or a faster connection but require additional fibre, equipment transport and staff relocation. The choice depends on the workload: latency-sensitive services and less time-critical model training do not have the same location requirements.
Geography also affects resilience. Concentrating facilities in a single area can expose an operator to the same extreme weather, fuel supply problem or utility disruption. Distributing capacity can reduce that common exposure, but makes networking and management more complex. An operator should compare the cost of spreading facilities with the operational value of redundancy, rather than simply selecting the lowest advertised electricity tariff.
Public authorities have a related choice. Offering a fast route to development can attract investment, but it may intensify demand on a local grid already serving homes and industry. Planning works best when it identifies who will build and pay for infrastructure, what timelines are realistic and which community benefits are likely to be delivered. A proposed data centre should be assessed on its own footprint and obligations, not on either an automatic assumption of economic benefit or a general objection to computing growth.
The financing structure can reveal hidden assumptions
The facility owner, computing customer, utility and power producer may be separate businesses. Their contracts distribute risks differently. A long lease can protect a developer from fluctuating demand, but only if the customer remains creditworthy and the facility is operational when the lease starts. An energy agreement may fix some costs while leaving the operator exposed to network fees or the cost of backup during shortages.
Financial due diligence should follow the sequence from grid connection through equipment commissioning to customer utilisation. A lender can ask how much spending occurs before each milestone, whether the project earns anything if only part of it is energised and who pays if a transformer or generating project arrives late. The same model should address changes in AI hardware: a data hall designed around today's equipment may need costly cooling or electrical modifications for future machines.
The demand side needs equal discipline. Committed customer contracts are stronger evidence than expressions of interest, though even contracts can include conditions and cancellation rights. Forecasting the power load from a full buildout while forecasting revenue from immediate full occupancy can conceal a timing gap. The strongest commercial case connects every increment of capacity to both a plausible customer and deliverable electricity.
A practical test of promised capacity
Public claims often combine operational power, power contracted for future delivery and power still awaiting approval. They should be separated in reporting. A site with ten megawatts connected and another hundred under discussion does not have a hundred and ten megawatts of productive capacity today. Similarly, an option on a parcel of land is not the same as a utility-approved expansion.
That distinction improves comparisons across projects. Investors can track connection dates, construction progress, installed computing equipment, active utilisation and revenue per unit of power over time. No single measure is decisive, and operators may reasonably protect commercially sensitive details. Consistent definitions, however, make optimistic schedules easier to identify and allow the benefits of genuine delivery to show through.
Questions readers may ask
Is AI responsible for all growth in data centre electricity use
No. Cloud services, storage, video and other digital activity also contribute. The impact of AI depends on the workload and how efficiently it runs. Published forecasts usually cover all data centre electricity use and should not be presented as a measure of AI alone.
Will buying renewable electricity guarantee continuous clean power
No. A contract can support renewable generation without matching each hour of consumption. The physical grid mix, storage and the timing of delivery determine what supplies a facility at a given moment.
What should investors check first
They should distinguish proposed capacity from connected capacity. Confirm the grid delivery date, committed upgrade costs, projected utilisation and the financial effect of delay before treating announced megawatts as a near-term earnings opportunity.
Sources
International Energy Agency’s Energy and AI report
US Department of Energy’s assessment of data centre demand
Lawrence Berkeley National Laboratory study
Federal Energy Regulatory Commission’s work on connecting large loads
Electric Power Research Institute’s updated data centre scenarios