Azure AI Capacity Deals: Check Location, Delivery and Cost Terms

The most useful part of the forecast concerns who carries the risk when promised AI demand requires real construction. A utility may need to expand generation, transmission, and substations before a datacenter starts operating. If the project arrives late or consumes less electricity than expected, someone still has to pay for that infrastructure.

Microsoft has already committed to shouldering more of those costs. Its January Community-First AI Infrastructure initiative and its participation in the March Ratepayer Protection Pledge provide concrete examples of the approach Forrester expects to spread. For cloud customers, the practical task is to distinguish those operator commitments from their own contracts—and to avoid assuming either that infrastructure costs disappear inside the cloud or that every new obligation will immediately appear on an Azure bill.

Forrester’s 2027 forecast puts the risk of unused capacity at the center​

In its September 17 sustainability predictions, Forrester frames AI expansion as a resource-allocation problem. Senior analyst Abhijit Sunil writes that governments, utilities, communities, and companies must decide who gets scarce power, who pays for infrastructure, and who absorbs the risk. The forecast concerns physical capacity as well as model capability.

Forrester cites a projection that US datacenter power demand will rise from 31 gigawatts in 2025 to 66 gigawatts in 2027. Those figures describe forecast datacenter demand, not a measured increase attributable exclusively to AI. The important commercial point is the anticipated speed of expansion: utilities would have to plan for substantially greater demand before knowing precisely which proposed facilities will materialize.

Sunil identifies the mismatch directly: “Utilities cannot assume that every proposed data center will arrive on schedule or consume its promised load.” Infrastructure built for demand that fails to appear can become a stranded asset—an investment whose expected use no longer supports its cost. In the situation Forrester describes, households and existing businesses could be left paying for facilities built to accommodate a speculative datacenter project.

The firm predicts that at least two countries will turn a ratepayer-protection principle into tariffs or equivalent obligations in 2027. Here, a tariff means an electricity rate structure, not an import duty. Forrester expects long-term minimum payments, financial guarantees, energy tariffs, and requirements for operators to reduce electricity demand when grids are under strain.

These mechanisms address related but different risks:

The forecast does not establish a worldwide standard or specify a new obligation for every cloud customer. It describes an expected direction of travel: more of the cost and uncertainty associated with expansion would sit with the organization requesting the electricity. Forrester’s further prediction—that enterprises rapidly scaling AI will increasingly inherit those costs—is a commercial forecast, not evidence that a particular provider has already changed its prices.

That boundary matters when interpreting the enormous investment estimates accompanying the AI buildout. The Register reports that PwC estimated cumulative global datacenter capital expenditure through 2050 at $22 trillion to $50 trillion, with a central estimate of $31.6 trillion. The underlying methodology is not established here, so the estimate is useful only as an indication of projected investment scale, not as a spending commitment or a basis for forecasting an enterprise cloud bill.

Microsoft has already committed to paying for more than electricity consumed​

Microsoft’s involvement makes this more concrete than a general warning about AI resource consumption. The White House’s March 4, 2026, fact sheet identifies Microsoft alongside Amazon, Google, Meta, OpenAI, Oracle, and xAI as signatories to the Ratepayer Protection Pledge. The announced commitment covers new generation resources and the power-delivery infrastructure required for their datacenters.

Under the pledge, the companies agreed to build, bring, or buy new generation resources and cover the cost of associated power-delivery upgrades. They also agreed to negotiate separate rate structures with utilities and state governments and pay for the power and related infrastructure brought online to serve their facilities, whether or not they use all the electricity. That last provision addresses the unused-capacity risk at the heart of Forrester’s forecast.

A pledge is not a universal electricity tariff. The White House announcement describes commitments and future negotiated arrangements; it does not establish identical terms for every datacenter, utility, or state. Nor does the announcement independently demonstrate that household electricity bills have fallen because of those commitments. The documented development is the agreement to allocate costs differently, rather than a measured result for consumers.

Microsoft’s own January 13 initiative provides more detail about how the company intends to apply that principle. President and Vice Chair Brad Smith said Microsoft would ask utilities and public commissions to set its rates high enough to cover the electricity costs of the datacenters it builds, owns, and operates. The commitment includes the cost of adding and using the infrastructure needed to serve those facilities.

Wisconsin offers a specific example. Microsoft said it supported a rate structure for “Very Large Customers,” including datacenters, that would charge those customers for the electricity required to serve them. Its announcement also committed to earlier planning with utilities, transparency about projected requirements, and advance contracting for electricity. Microsoft described paying for transmission and substation improvements required by its expansion as a continuation of existing practice, rather than an entirely new policy.

Taken together, the January initiative and March pledge support a narrower, stronger conclusion than “AI is about to make Azure more expensive.” Microsoft has accepted the principle that its datacenter growth should fund the electricity infrastructure it requires. Whether and how that affects a customer’s price depends on subsequent commercial decisions and contract terms; neither announcement establishes an Azure surcharge, a Microsoft 365 Copilot price change, or a new customer minimum commitment.

Water and land make datacenter expansion a local decision​

Electricity gets much of the attention because it lends itself to rates, contracts, and capacity figures. Forrester’s forecast also treats water and land as constraints that can determine whether a project gets built. A company’s ability to finance a campus does not, by itself, establish that a particular community can accommodate its demands.

Sunil lists the effects residents may experience: new substations and transmission lines, water withdrawals, land conversion, construction traffic, persistent low-frequency noise, and emissions from backup generators or dedicated power generation. Forrester expects community-impact reviews to become a gate for new projects. Its prediction is that operators will increasingly need to demonstrate how they will manage those effects and who will pay, rather than relying on general promises of local benefits.

Microsoft’s January water commitments illustrate why site-specific detail is essential. The company described historical reliance on evaporative cooling, which draws on water to manage heat, particularly in hot weather. It also said it had introduced a closed-loop design that constantly recirculates cooling liquid and no longer needs potable water for cooling, with deployments in locations including Wisconsin and Georgia. That is Microsoft’s account of the design and its deployment, not an independently measured result across its entire fleet.

Microsoft’s closed-loop cooling claim has a specific boundary​

The claim concerns potable water used for cooling in a particular next-generation design. It does not establish that every Microsoft datacenter uses that design, that every facility has eliminated water demand, or that its surrounding water infrastructure needs no investment. Preserving that scope prevents an improvement in one part of a facility’s operations from becoming an unsupported claim about the whole site.

Microsoft also committed to a 40 percent improvement in water-use intensity across its owned datacenter fleet by 2030. A target expressed as intensity is different from a promise to reduce total withdrawals by the same percentage. The announcement does not justify treating the two as interchangeable, especially while the company is expanding its infrastructure.

Its Quincy, Washington, example addresses a different part of the problem. Microsoft says it partnered with the city on the Quincy Water Reuse Utility, which treats and recirculates datacenter cooling water rather than relying on the community’s groundwater supply. The company also says it funds required future improvements to that system. This combines a change in water supply with a commitment about who pays for supporting infrastructure.

The water program extends beyond cooling design. Microsoft promised to replenish more water than it withdraws in the same water districts where its datacenters use water, and to begin publishing water-use data for each US datacenter region together with replenishment progress. Those are commitments to local accounting and disclosure; the announcement alone does not demonstrate completion of every replenishment project or publication of every regional dataset.

For an enterprise evaluating expansion, the useful inference is that “the provider has a water policy” is an incomplete answer. The relevant facts are which design serves the proposed capacity, what local supply and treatment arrangements support it, and whether those arrangements are operating or still planned. Forrester’s forecast makes the same broader point about land: a site, its supporting infrastructure, and its effects on nearby communities need to be considered together.

Air-permit streamlining does not resolve every datacenter constraint​

The regulatory direction is not uniform. Alongside Forrester’s expectation of greater community scrutiny, The Register reports that the US Environmental Protection Agency proposed removing federal minimum requirements for public participation in minor-source air permitting. The report says that could allow some datacenter power projects to proceed without public notice. It describes a proposal, not a completed elimination of those requirements.

The scope is important. Minor-source air permitting concerns a particular environmental approval process. It is not a general authorization to build any datacenter, and a change to participation requirements in that process would not itself supply water, secure land, or expand transmission capacity.

Likewise, Forrester’s expectation of community-impact reviews should not be read as an announcement of a single new federal review system. It is a forecast about how development decisions will increasingly be made. Community objections and requirements involving noise, traffic, water, or utility costs can remain relevant even where one permitting procedure becomes less demanding.

The practical synthesis is that cost allocation and project approval are separate decisions. An operator can agree to fund electrical upgrades without having resolved a site’s water requirements. A streamlined air-permit process can change one administrative step without establishing that the utility can deliver power on the project’s preferred schedule.

For IT planners, that means a broad policy announcement is a weak substitute for a project-specific capacity commitment. A government may favor AI construction while a particular facility still depends on unfinished infrastructure or unresolved local approvals. The evidence supports examining those dependencies; it does not establish that any named Azure region is currently delayed by them.

Azure buyers should distinguish operator obligations from customer terms​

Forrester expects cloud and colocation economics to become more dependent on location, electricity contracts, grid maturity, and an operator’s ability to generate power or reduce demand. That is a useful procurement framework because it identifies the factors the firm believes will shape costs. It is not a new Azure pricing schedule.

There are several distinct relationships involved. A utility may impose obligations on a datacenter operator. The operator may have commitments to customers using space or capacity in the facility. A cloud provider may package its services under a different set of customer terms. An obligation in the first relationship does not automatically become an identical obligation in the last.

This distinction is particularly important for minimum payments. Forrester predicts that datacenter operators will increasingly face them, while the White House pledge already describes payment for associated power and infrastructure even when electricity goes unused. Neither establishes that an ordinary Azure subscription now requires the customer to underwrite a power plant or accept a new take-or-pay electricity arrangement.

The same discipline applies to flexible-load requirements. Forrester expects operators to face obligations to reduce electricity demand under grid stress. The evidence does not identify which Azure services would be affected, how Microsoft would implement any such obligation, or whether a customer workload would experience an interruption. It would therefore be premature to turn that forecast into an outage warning.

Infrastructure commitments belong beside price in capacity planning​

The procurement implication is an inference from the documented risks: a large AI deployment should be evaluated against the capacity and commercial terms the provider actually offers, rather than against the assumption that future infrastructure will arrive wherever and whenever needed. Location, delivery timing, and contractual flexibility become relevant inputs alongside the quoted service price.

An enterprise that needs capacity in a particular geography should establish whether the offer depends on existing facilities or future expansion. If expansion is involved, the useful distinction is between a target and a contractual commitment. Forrester’s stranded-infrastructure argument explains why that difference can matter to both parties: providers and utilities may seek firmer commitments precisely because they are being asked to spend before demand is certain.

Electricity exposure also deserves precise language. A fixed customer price, a price that can be revised at renewal, and an explicit energy-cost adjustment are different commercial arrangements. The reporting does not establish which of these will become more common in Azure contracts, but it does provide a reason for buyers to identify which arrangement they are being offered rather than assuming that every infrastructure risk is either fully absorbed or automatically passed through.

The supplied announcements offer no basis for changing an existing Microsoft 365 Copilot deployment solely because of this forecast. The immediate audience is the team making a substantial capacity, location, or long-term spending decision. Its task is to test the assumptions in that decision, not to predict a price increase that Microsoft has not announced.

What this means for Azure and enterprise AI planning​

Before making a substantial new AI capacity commitment, buyers should confirm the delivery and pricing assumptions in the actual offer. Existing customers whose plans do not depend on a major expansion have no specific operational change to make on the strength of Forrester’s forecast alone.

A practical review should keep today’s terms separate from future expectations:

• Confirm whether the proposed capacity is available now or depends on future construction, utility upgrades, or local approvals, and distinguish a delivery target from a contractual commitment.
• Identify the geography the business needs and whether the provider’s offer commits to supplying capacity there, rather than treating a global expansion announcement as evidence of availability in that location.
• Examine minimum-spend obligations, reservation terms, cancellation exposure, and any energy-related price adjustments in the customer agreement; do not assume an operator’s utility obligations automatically become yours.
• Ask how any power-demand reduction requirements are handled for the offered service, while avoiding an unsupported assumption that they necessarily mean customer downtime.
• Evaluate water and community-impact claims at the relevant site or region, distinguishing deployed cooling designs and operating infrastructure from future targets and replenishment promises.

These checks do not require an IT department to become a utility regulator. They require procurement, finance, and technical teams to agree on what has actually been purchased: capacity available today, a commitment to future supply, or a plan whose delivery still has dependencies. The distinction is especially valuable when an AI business case assumes rapid growth.

Forrester’s prediction is that those dependencies will become more visible in datacenter economics during 2027. Microsoft’s commitments already show one part of that transition: expansion comes with responsibility for the infrastructure and local resources it requires. For enterprise buyers, the concrete consequence is to make capacity location, delivery commitments, and cost allocation explicit before signing a long-term AI deal—not after the project’s growth assumptions have become contractual obligations.