Power first: AI data center siting now starts at the substation

Sandeep Prakash plans power and capacity for the data center regions that AI workloads run on, and he says the first question on a new site has changed. Speaking in a recent interview on Wavelengths, the Amphenol Broadband Solutions podcast, with host Daniel Litwin of Voice B2B, Prakash said operators used to start with the fiber route and work out the electricity later. Now they start with the substation.

Prakash, a senior program manager in data center infrastructure planning who earlier worked on network coverage strategy in the wireless industry, described a shift in where the industry's shortage sits. The scramble used to be for GPUs. Now it is for the power to run them.

The bottleneck moved from chips to the substation

The reason, in Prakash's telling, is that demand changed shape. Early AI use was a prompt and a response, a linear load, he said. Model processing, video generation and low-latency inference for a hospital running AI diagnostics or a financial firm running real-time risk models draw far more compute, and that compute needs power in specific places.

A review paper posted on arXiv in September 2025 by researchers at Texas A&M University and Harvard University puts figures on that density. Modern hyperscale AI data centers typically draw more than 100 megawatts, the authors write, and some new campuses are planned at the gigawatt scale. AI computing racks run at 30 to more than 100 kilowatts each, against 7 to 10 kilowatts for a traditional server rack, according to the same paper.

For a facilities lead used to a 10 kilowatt rack, that is a different building with a different utility conversation. Prakash said the two disciplines that used to run separately, network engineers on fiber routes and latency, utility planners on generation and load forecasting, now make one location decision together. He described regional shortfalls concentrated where the compute is wanted most.

Grid queues run four to seven years in Virginia

RAND put the scale of the shortfall in a January 2025 report. AI data centers could need 68 gigawatts of power worldwide by 2027, RAND estimated, against 88 gigawatts of total global data center capacity in 2022, with 10 gigawatts of additional capacity needed in 2025 alone, more than the state of Utah's entire power capacity. RAND's extrapolation reaches 327 gigawatts by 2030 if chip supply keeps growing exponentially.

The grid is not keeping pace with that curve, according to RAND. Grid connection requests take four to seven years in key regions such as Virginia, the report says, and transmission lines face multistate permitting and local opposition. RAND also flags permits for on-site backup generators and environmental assessments as points where data center projects stall.

Prakash said utilities were not prepared for this and are now trying to shorten the time to first power. He pointed to fast-tracking interconnection and upgrading aging lines in markets that already have infrastructure, which he said is quicker than building from scratch. For the CIO signing a colocation contract, the delivery date on that contract is set by the utility, not the landlord.

Prakash's demand figure runs ahead of the IEA's published series

The International Energy Agency's (IEA) own published baseline is lower on timing. In its April 2025 Energy and AI report, the IEA estimated that all data centers together consumed around 415 terawatt hours in 2024, about 1.5% of global electricity, after growing 12% a year over the prior five years. The arXiv review cites the IEA's path from that 415 terawatt hours to roughly 945 terawatt hours by 2030.

So Prakash's figure lands years ahead of the agency's own trajectory. The direction, though, is the same, and the IEA's newer numbers show it steepening. In its 2026 Key Questions on Energy and AI report, the IEA says global data center electricity demand grew 17% in 2025, in line with its projection, while consumption at AI-focused data centers surged 50%, and its satellite tracking shows AI factories more than tripled in capacity over 18 months.

The IEA also notes the countervailing trend: energy per AI task has been falling by at least an order of magnitude a year, while video generation, reasoning and agentic tasks use hundreds or thousands of times more energy per query than simple text. Both curves belong in the same capacity plan. Prakash said the same thing from the planner's chair.

He described buffer inventories that absorb near-term swings, and long-lead components and global supplier shortages as the risks a model built on past history will miss. The planner who gets the call when a region runs out of capacity is the one who forecast off last year's workload mix.

Solar and batteries beat a grid upgrade on time to power

Prakash's answer to the queue is to stop waiting on it where possible. In regions with abundant sun or wind, he said, alternative generation can bridge a gap temporarily or replace grid power permanently, and solar converters plus storage can be installed much faster than grid infrastructure can be brought online. Clean energy, in his framing, is also the fast energy.

On the demand side, he pointed to newer GPU architectures improving power usage effectiveness, liquid cooling that cuts water draw, a concern he tied to drought-affected regions such as California in summer, and shifting workloads into hours when residential power draw is low. He cited Sweden, where he said data center heat is redirected to warm homes in winter. The IEA's 2025 report puts a number on the cooling upside: cooling takes about 7% of consumption at efficient hyperscale sites and over 30% at less efficient enterprise facilities.

Prakash was candid about the limits. He said major companies have published sustainability reports acknowledging that emissions rose in recent years partly because of AI infrastructure, and he expects net zero commitments made for 2030 or 2040 to be adjusted because they never assumed this growth. He also said several data center projects, including some in high-growth regions, have been paused pending local government permission.

Energy and compute budgets now move together

McKinsey's Technology Trends Outlook 2026, published in September 2026, describes energy and AI investment moving in step. Energy technologies drew nearly $200 billion in investment in 2025, according to McKinsey, and spending on AI infrastructure doubled in a single year. The report frames the defining questions of 2026 as who can build the hardware and staff the deployments.

The IEA's 2026 report says capital expenditure by the largest technology companies exceeded $400 billion in 2025 and is expected to rise another 75% in 2026, with the capex of just five companies now larger than global investment in oil and gas production. Prakash's three priorities line up with that money: transmission and storage to move power from where it is abundant to where it is short, geography near dense populations that need low latency, which he said includes future growth markets in Southeast Asia, and longer-lived battery storage. Nuclear and fusion, he said, sit further out.

RAND adds a consequence for the procurement director choosing between regions. If U.S. companies cannot find adequate power, the report warns, they may build data centers abroad in countries with faster permitting and more available power. RAND recommends that planners model grid supply against data center demand explicitly, factoring in reliability requirements and power usage effectiveness.

For an enterprise buyer evaluating a colocation site or a new cloud region, the practical change is the order of questions in the RFP: the interconnection date and the utility behind it come before the fiber providers on the property. The markers to watch belong to the sources: RAND's four to seven year connection queues in Virginia, and whether the 75% capex jump the IEA expects in 2026 lands in buildings that can be energized. Prakash's own marker is whether the 2030 and 2040 net zero targets get rewritten to fit the growth he plans for every day.