AI growth created a power-system decision
A global cloud provider and regional utility were assessing how to co-invest in firm generation and grid infrastructure to support anticipated AI-related electricity demand. The cloud provider needed timely, reliable, scalable power capacity. The utility needed to maintain prudent, reliable service under regulatory oversight and long-term public-service obligations.
The opportunity had significant implications for infrastructure, capital allocation and growth. But the projected demand remained uncertain in timing, location, utilisation and persistence. AI demand could develop rapidly, but it could also shift with cloud adoption, data-centre efficiency, hardware evolution, customer uptake and the location of computing capacity.
The IEA projects global data-centre electricity consumption to rise from approximately 485 TWh in 2025 to around 950 TWh in 2030. That growth creates a substantial infrastructure challenge, but it does not make every projected load sufficiently certain to justify immediate, long-lived investment.iea
The decision could not be resolved through a conventional power-purchase agreement or a standard corporate joint venture. It required a partnership design that linked private demand with the wider electricity system needed to serve it.
The choice was partnership architecture
The parties could consider customer-funded infrastructure, utility-funded assets, rate-base-supported investment, direct capital contributions, capacity reservations, firm-power arrangements, development options, milestone funding, third-party participation or hybrid structures.
Each route created a different balance of capital exposure, cost recovery, regulatory scrutiny, capacity certainty, development speed, customer protection and stranded-asset risk.
The central question was which assets were attributable to the cloud provider’s load, which created broader system value, and how the costs should follow the resulting benefits. A hyperscaler might be prepared to make demand commitments or fund certain facilities in return for clearer capacity rights. A utility might be able to plan and develop infrastructure more confidently if demand were contractually credible. Neither party could assume that all costs, risks or benefits would be treated as private.
Capacity access also required careful definition. It could be physically dedicated, contractually allocated or provided through the broader grid. Each approach affected reliability, outage treatment, curtailment, reserves, operational control, regulatory feasibility and the position of other customers.
Capacity needed to work for the whole grid
AI-related data-centre demand depended on cloud adoption, model development, hardware efficiency, customer growth, load location and operating requirements. Local concentration could create requirements for transmission, substations, distribution upgrades, generation, water, land and interconnection that were not visible in an aggregate demand forecast.
Long-lived infrastructure needed credible demand support. Yet the cloud provider also needed confidence in price, timing, capacity availability and project deliverability before making binding commitments. This created a reciprocal-commitment problem: the utility could not prudently build for speculative load, while the customer could not commit fully without a credible infrastructure pathway.
Regulation introduced a further constraint. The Federal Energy Regulatory Commission launched a review of large-load co-location arrangements in 2025, highlighting the need for clear tariff treatment, grid-reliability protections and fair cost allocation. The issue was not only whether additional capacity could be developed. It was whether the resulting infrastructure structure could meet cost-causation, customer-protection and public-interest requirements.ferc
Generation and grid development also carried permitting, engineering, procurement, construction, workforce, financing, fuel, interconnection and delivery risks. Firm-generation pathways could support reliability and long-term planning, but their development timelines and risk profiles needed to be matched to the actual pace of AI-load growth.
The partnership therefore had to support data-centre expansion without transferring inappropriate costs or reliability risk to existing customers.
Testing regulated and commercial structures
Bruqe framed the work around demand credibility, reliability requirements, grid and generation alternatives, cost causation, ratepayer exposure, regulatory feasibility, capital limits, development risk, capacity rights and partner incentives.
The analysis mapped AI adoption, data-centre loads, load location, generation, transmission, rate-base treatment, tariffs, customer commitments, system reliability, licensing, fuel, construction, public acceptance, financing and governance as connected variables.
It tested utility-funded, customer-funded, regulated, corporate-procurement, third-party and hybrid structures. It also compared demand commitments, deposits, credit support, minimum bills, capacity reservations, development options, milestone funding, cost-sharing, curtailment arrangements, expansion rights, step-in rights and exit mechanisms.
These structures were assessed across plausible futures involving lower or higher AI demand, delayed load ramps, generation and grid constraints, regulatory conditions, construction delays, financing changes, local opposition, fuel or technology issues and reliability events.
The objective was not to identify a universally superior commercial model. It was to establish the reciprocal evidence thresholds that would justify progressing, deferring, redesigning or withdrawing investment.
Aligning private growth with public obligations
The work separated bilateral capacity access from system-wide reliability and regulatory legitimacy. It clarified where demand commitments, customer funding, utility investment, flexible capacity rights, governance and staged infrastructure could create a more credible partnership.
Some structures could support capacity development but required stronger evidence of demand, revised cost allocation, additional regulatory engagement or greater flexibility before further capital was committed. Others could be more robust because they better aligned the risks created by a large customer load with the parties able to manage them.
The alliance was therefore assessed as a shared-capability model—not simply as a private power transaction.
Making infrastructure conditional on evidence
The resulting framework linked infrastructure commitments to evidence of AI-load development, project feasibility, regulatory acceptability and wider system value.
The central implication was clear: credible AI power partnerships secure capacity only when they work for the grid as well as the customer.
