Every article in this series has circled the same underlying challenge: the grid was planned for slow, predictable growth, and data centers deliver fast, lumpy, enormous growth instead. This final piece examines how utilities are adapting the core machinery of grid planning — load forecasting, resource procurement, tariff design, and cost allocation — to the era of the 200-megawatt customer.
Forecasting: from extrapolation to investigation
Traditional load forecasting is an exercise in extrapolation. The forecaster takes decades of historical data — how demand responded to weather, economic growth, and population — builds a statistical model, and projects it forward. Energy efficiency trends, electric vehicle adoption, and building electrification get layered in as adjustments. The method works well when the future resembles the past.
Data centers break the method because they have no past. There is no historical relationship between GDP growth and data center construction from which to extrapolate; the current wave is unprecedented in scale and speed. A forecaster who extrapolates past commercial load growth will miss the data center wave entirely. One who simply adds up every interconnection request will overstate it wildly, because queues are full of speculative projects.
Utilities are developing a new forecasting discipline that looks more like investigative reporting than statistics. It starts with the interconnection queue, then discounts it: which projects have site control? Which have financing? Which have signed power contracts or posted substantial deposits? Planners assign probabilities to individual projects based on development milestones, producing probability-weighted forecasts rather than single numbers.
They also model scenarios explicitly. A base case might assume a moderate realization rate for queued projects; a high case assumes most get built; a low case assumes a shakeout. Resource planning then has to be robust across scenarios — procuring generation that is useful in all of them, or structuring procurement in tranches that can be scaled up or down as the load materializes. This is a profound shift from the old world, where the forecast was a number and the plan followed from it.
The forecasting challenge extends beyond the utility’s own territory. Data center developers shop for sites across regions, and a project that appears in three utilities’ queues will only get built once. Regional coordination — sharing queue information, aligning assumptions — is improving but remains incomplete. Double-counting the same speculative load in multiple forecasts is a real and recognized risk.
Resource planning: procuring for uncertainty
Once the forecast exists, the utility must procure the resources to serve it — generation capacity, and in some cases transmission, on timelines that now look uncomfortably short. Integrated resource plans (IRPs), the formal documents through which utilities lay out their procurement strategy, are being rewritten around large-load uncertainty.
The central dilemma: build generation for load that might not materialize, or risk shortfalls if it does? Building early means ratepayers fund capacity that could sit underutilized — the stranded-cost risk. Building late means the data center waits years for power, or the region faces reliability shortfalls — the delay risk. There is no risk-free option; the planning problem is choosing which risk to bear and how to price it.
The emerging answer is staged commitment. Rather than procuring the full forecasted need upfront, utilities procure in phases tied to the load’s own milestones: initial capacity when the data center signs a binding contract and posts security, additional tranches as construction milestones are met. This aligns the utility’s investment with the customer’s demonstrated commitment, so neither side bears the other’s speculative risk.
Contract structure does heavy lifting here. Long-term contracts — 10, 15, even 20 years — give the utility revenue certainty to finance new generation. Take-or-pay provisions and minimum bills ensure the customer pays for reserved capacity whether or not it uses every megawatt. Exit fees and long notice periods prevent the customer from stranding utility investments by departing early. These provisions are increasingly standard for very large loads, and regulators are generally supportive — they protect the customers who aren’t at the negotiating table.
Tariff design: pricing the large load correctly
The rate schedule is where planning meets the customer’s bill, and large-load tariffs are being reinvented. The design objectives are clear: recover the full cost of serving the load, send efficient price signals, and protect other customers from cost shifts. The implementation is where it gets interesting.
Demand charges — fees based on the customer’s peak draw rather than total consumption — are the primary tool. A data center’s cost to the utility is driven substantially by capacity: the generation, transmission, and distribution built to serve its peak. Demand charges align the bill with that cost structure. For very large customers, demand charges may be supplemented or replaced by negotiated contract demand: a fixed megawatt reservation the customer pays for regardless of use.
Time-varying rates add a second dimension. Even a flat-profile data center can shift some load — pre-cooling, batching non-urgent computation, charging batteries — and rates that reward off-peak consumption improve system efficiency. More significantly, time-varying rates expose large customers to the real cost of consuming during system peaks, which is when their flat load is most expensive to serve.
New large-load tariff classes are proliferating. Common features include: higher fixed charges reflecting the dedicated infrastructure; minimum bills set as a percentage of the contracted demand, ensuring cost recovery even if the facility idles; long contract terms with limited exit rights; and in some jurisdictions, requirements that new large loads be served by new clean generation — effectively bundling the tariff with resource procurement. Economic development riders that discount rates for large new loads, once common, are being reconsidered or conditioned on stronger cost-recovery guarantees.
Cost allocation: the fairness question
Underneath the tariff mechanics lies the fundamental fairness question: when serving new large loads requires billions in grid investment, who pays? The principle that the cost causer should pay is widely accepted; the difficulty is defining the cost causer when infrastructure serves multiple purposes over decades.
Consider a new transmission line built primarily to serve a data center cluster. It also relieves existing congestion, improves reliability for nearby communities, and creates capacity for future growth. Is the data center the sole cost causer, or one beneficiary among several? Reasonable regulators disagree, and the answer determines whether the line’s cost lands on the data center’s contract, on all ratepayers, or somewhere in between.
The emerging consensus leans toward protecting existing customers first. Mechanisms include: assigning all directly attributable upgrade costs to the interconnecting customer; using minimum bills and long terms to guarantee recovery of shared investments; and in some cases, requiring large-load customers to fund new generation directly — through utility-built resources underwritten by the customer’s contract — rather than drawing on the shared resource pool.
There is a mirror-image concern: overcharging large loads. If tariffs are punitive, development goes elsewhere — to neighboring states with friendlier rate structures — and the home region loses the investment, jobs, and tax revenue. Interstate competition for data center investment is real, and regulators know it. The art of large-load rate design is finding the rate that recovers full costs without driving away the load that justifies the investment.
Reliability and the reserve margin
All of this planning serves one ultimate objective: keeping the lights on. Grid operators maintain reserve margins — generation capacity above expected peak demand — to cover plant outages and extreme weather. Data center load growth tightens these margins from both sides: it raises the peak, and it arrives faster than new generation can be built.
Several regions are now planning explicitly around large-load-driven peak growth, accelerating generation procurement and in some cases retaining existing plants that were slated for retirement. Demand response — paying large customers to curtail during emergencies — is being extended to data centers, though their flat, critical workloads limit how much they can flex. The most reliable data centers maintain on-site backup that can carry the full load, which helps the facility but doesn’t help the grid unless coordinated.
The data center as grid resource
The planning conversation has been one-directional so far: how does the grid serve the data center? But large, flexible, well-equipped loads can also serve the grid, and planners are beginning to treat data centers as resources rather than just burdens.
The most immediate opportunity is demand response. While AI training is often described as inflexible, parts of the workload genuinely are shiftable: checkpointing allows a training run to pause and resume, batch inference can be deferred by hours, and pre-cooling can shift cooling load. For illustration, a 200-megawatt campus that can shed 40 megawatts for four hours on ten critical days per year provides the grid with 1.6 gigawatt-hours of avoided peaking energy — comparable to a mid-size battery installation, at a fraction of the capital cost.
On-site generation and storage compound the value. A data center with behind-the-meter turbines or batteries can export power during grid emergencies, functioning as a distributed peaking resource. A few pioneering arrangements already compensate data centers for grid services — frequency regulation, spinning reserve, emergency capacity — turning the facility’s power infrastructure into a revenue-generating asset rather than pure insurance.
Realizing this potential requires market and contract design that rewards flexibility. Today’s large-load tariffs mostly charge for peak demand, which penalizes the very load shape the grid wants. Inverted structures — lower demand charges for customers who demonstrate dispatchable flexibility, capacity payments for curtailable load, streamlined participation in ancillary service markets — would align the data center’s incentives with the system’s needs. The flexible-interconnection concept from our interconnection piece is one version of this bargain: faster, cheaper grid access in exchange for being a resource when it counts.
There are limits. No grid operator should plan around curtailing loads that may not materialize, and data center operators won’t sacrifice contractual uptime guarantees lightly. But the direction is clear: the most sophisticated large-load planning treats the data center as a two-way participant in grid reliability, not a one-way drain on it.
The planning paradigm shift
Step back, and the change is philosophical. Twentieth-century utility planning assumed the utility knew its customers: load grew with the economy, and the forecaster’s job was to measure the relationship. Twenty-first-century planning confronts customers the utility must investigate individually — a handful of corporate decisions, each worth hundreds of megawatts, determining the territory’s growth.
This demands new capabilities: project-level due diligence, scenario-based planning, staged procurement, sophisticated tariff design, and constant coordination with economic development authorities, grid operators, and the customers themselves. It also demands regulatory frameworks that let utilities act quickly — approving investments on shorter cycles, allocating costs fairly, and balancing economic development against ratepayer protection.
The utilities that master large-load planning will shape where the data center industry grows and how the grid evolves to serve it. Those that don’t will face the worst of both worlds: the costs of growth without the benefits, or the benefits of growth without the infrastructure to support it. Forecasting, tariffs, and cost allocation sound like back-office topics. In the age of load growth, they are where the future of the grid gets decided.