AI Data Center Power: Behind-the-Meter, Nuclear PPAs and Grid Architecture

Two announcements in the same 48 hours made the central problem of the AI build-out unusually concrete. On September 30, 2026, Constellation and Amazon announced a 20-year power purchase agreement tied to a roughly $3 billion upgrade of the Calvert Cliffs nuclear plant in Maryland. On October 1, JERA, Dell Technologies and RHAELM announced a memorandum of understanding for a 400 MW AI campus in Japan that plugs straight into an existing power station. One deal buys electrons through the grid for decades; the other tries to skip the grid queue entirely. Both are answers to the same bottleneck.

That bottleneck is AI data center power: not chips, not capital, but how many megawatts you can energize, and how soon. This post works through the electrical and thermal architecture behind both deals, the arithmetic that connects megawatts to GPUs, and the trade-offs between behind-the-meter supply, nuclear PPAs and conventional grid interconnection. You leave with a decision framework, worked (clearly labelled illustrative) numbers, and a list of failure modes that the press releases do not mention.

What this covers: what each deal actually says, the reference power chain from substation to rack, MW-to-GPU-to-PUE arithmetic, how PPAs differ from physical delivery, where behind-the-meter designs break, and a practical checklist.

Context and Background

For most of the cloud era, power was a line item. A hyperscale facility drew 20 to 60 MW, utilities could serve it in a couple of years, and the siting question was mostly about fiber and land. AI training and inference clusters changed the unit of planning. Campuses are now specified in hundreds of megawatts and discussed in gigawatts, and a single site can look like a mid-sized industrial city to the local utility. The grid, which was built around slow, predictable load growth, has interconnection queues and transformer lead times that now dominate project schedules.

Two reference points frame the scale. The International Energy Agency’s analysis of energy and AI projected that global data center electricity demand would roughly double by 2030 from a 2024 base, with AI-optimized servers the fastest-growing component; read the primary source at the IEA Energy and AI report rather than relying on secondary summaries, since the headline figures are scenario-dependent. For the U.S. grid specifics, our earlier analysis of AI data center power, grid constraints and nuclear in 2026 covers interconnection queues and utility forecasting in more depth, and this post builds on it rather than repeating it.

The three ways a campus gets electrons

Operators have, in practice, three supply strategies, and the two news items sit at different points on the spectrum.

The first is grid-connected supply with a financial hedge. The campus connects to the transmission or distribution system, buys power at wholesale or retail rates, and signs a power purchase agreement (PPA) with a generator to lock price and claim clean attributes. Electrons are fungible on the grid, so the PPA is largely a financial and accounting instrument. The Amazon and Constellation deal at Calvert Cliffs is of this type.

The second is co-location at a generator, sometimes called front-of-the-meter or behind-the-meter depending on where the metering point sits. The data center is built next to an existing plant and takes power through a short private connection. Variants of this appeared in earlier nuclear-adjacent deals in the U.S., which drew regulatory scrutiny over how much grid cost-sharing the co-located load avoids.

The third is fully islanded or on-site generation: gas turbines, fuel cells or reciprocating engines at the campus, with the grid as a backup or not at all. The JERA, Dell and RHAELM project is described as behind-the-meter and directly connected to an existing generation asset, which places it in the second category with elements of the third.

A correction worth making up front

Some early coverage and internal notes blended the two announcements into one “400 MW at Calvert Cliffs” story. The primary sources do not support that. JERA’s own release describes the Chiba Project at JERA’s Chiba Thermal Power Station in Japan, with up to 400 MW, a total capital deployment above US$15 billion (about 2.3 trillion yen), and operations targeted around 2028. Constellation’s release describes Calvert Cliffs in Lusby, Maryland, with 190 MW of new generating capacity, 690 MW of total power under the agreement, a 1,790 MW plant, more than $3 billion of investment, and new capacity arriving in 2030 to 2032. They are different continents, different physical architectures and different commercial logics, which is exactly why comparing them is useful.

The Reference Architecture: From Source to Rack

The short answer to “how does an AI campus get its power” is a chain of conversions, each with losses and failure modes: a generation or grid source, high-voltage interconnection, step-down transformation, uninterruptible power and storage, rack-level distribution, and a cooling plant that removes nearly every watt as heat. Behind-the-meter designs shorten the first link; PPAs leave the physical chain unchanged and change only who pays whom.

Figure 1: Reference AI data center power chain. A grid path and an on-site generation path both feed the campus high-voltage bus, which splits into IT load and the cooling plant.

Figure 1 shows the two supply paths converging on one campus high-voltage bus. The grid path runs through a utility substation and, critically, through an interconnection process that can take years. The on-site path connects a generator, here gas or nuclear, directly to the same bus. From the bus, transformers and switchgear feed two consumers: the uninterruptible power supply and battery energy storage system (UPS and BESS) that protect the GPU racks, and the cooling plant that removes the heat those racks produce. The two are coupled, since a cooling outage becomes an IT outage within minutes at current rack densities.

Why the interconnection step dominates schedules

A new large load needs a utility to study the impact, build or upgrade a substation, and sometimes add transmission. Transformer and switchgear manufacturing capacity is limited, and large power transformers have had multi-year lead times in recent years; check current figures with the manufacturer or utility because they move. The critical point is that the electrical equipment on the utility side is not the part you control. A GPU cluster can be ordered and racked in months, but a 400 MW grid connection that is not yet built cannot be accelerated by capital alone.

This is the thesis the JERA release states directly: a behind-the-meter connection to an existing generation asset can deliver computing capacity “years ahead of a conventional grid-connected development schedule.” The power plant already has a switchyard, a fuel supply, a cooling-water arrangement, and an interconnection to the grid. The data center borrows all of that.

Voltage levels and what changes between them

Understanding the chain helps when reading deal announcements. Generation at a plant typically steps up to transmission voltage, in the hundreds of kilovolts, to move power over distance. A data center campus connecting at transmission or sub-transmission level steps that down to a medium voltage such as 13.8 kV or 34.5 kV for on-campus distribution, then to roughly 400 to 480 V (or higher in newer designs) at the rack-row level. Every step is a transformer with a few tenths of a percent to a couple of percent of loss, and every step is a place where redundancy is bought.

Collocating at a generator can remove one or more of those stages. If the plant’s switchyard is already at a voltage the campus can tap, the long-distance transmission leg disappears. That is not only a schedule gain; it removes transmission losses and reduces exposure to grid congestion. It also moves reliability responsibility onto the campus owner and the plant, because the grid is no longer a second, independent source unless the design explicitly keeps one.

Redundancy tiers still apply

Uptime-style tier concepts, N, N+1 and 2N, apply to the power path whether supply is grid or on-site. A behind-the-meter campus with a single generating unit and no grid tie is effectively N for supply, regardless of how redundant the internal switchgear is. A realistic design keeps the grid connection as a standby source or uses multiple generating units with enough spinning margin to survive a unit trip. Training workloads tolerate checkpoint-and-restart interruptions better than inference serving, which is one reason some operators split campuses by workload class and assign different redundancy to each.

From Megawatts to GPUs: The Arithmetic Behind a 400 MW Campus

A megawatt figure on a press release is facility power, the total drawn at the campus boundary. The number of accelerators it supports depends on Power Usage Effectiveness (PUE), the per-server power envelope, and how much headroom the electrical design reserves. The direct answer: at an illustrative PUE of 1.2, a 400 MW facility delivers about 333 MW to IT equipment, which at an assumed 1.8 kW per accelerator including its share of CPU, memory, networking and storage supports on the order of 185,000 accelerators.

Figure 2: Where 400 MW of facility power goes at an illustrative PUE of 1.2. IT load dominates, and all of it eventually becomes heat the cooling plant must reject.

Figure 2 traces the flow. Facility power splits into IT power and overhead; the IT power feeds servers and network, of which the GPUs are the largest consumer; and every watt of IT load becomes heat that returns to the cooling loop, which itself draws power. The loop closes: cooling overhead is driven by IT load, so improving cooling efficiency reduces total facility demand without touching compute.

Worked example (illustrative, not from either announcement)

PUE is defined as total facility energy divided by IT equipment energy. The numbers below are assumptions chosen to show the method; neither JERA nor Constellation has disclosed GPU counts, PUE targets or rack power for these projects, and I have not found them published.

Facility power: 400 MW (JERA’s stated “up to” capacity).
Assumed PUE: 1.2, a figure achievable with liquid cooling in favorable climates; air-cooled legacy sites often run 1.4 to 1.6.
IT power: 400 / 1.2 = about 333 MW.
Assumed all-in power per accelerator: 1.8 kW. This folds a roughly 1.0 to 1.4 kW accelerator thermal design power together with host CPU, memory, NICs, switches and storage share. Current flagship parts vary, so treat this as a placeholder.
Accelerators supported: 333,000 kW / 1.8 kW = about 185,000.

Now vary the PUE. At 1.1, IT power is 364 MW and the same arithmetic yields about 202,000 accelerators. At 1.4, IT power is 286 MW and the count falls to about 159,000. The swing between 1.1 and 1.4 is roughly 43,000 accelerators from the same 400 MW grant, which is why cooling design is a revenue question, not a facilities footnote. Each 0.1 of PUE at this scale is worth about 30 MW of power that could instead run compute.

Energy and cost, same caveats

A 400 MW load at 100 percent utilization consumes 400 MW times 8,760 hours, or about 3.5 TWh per year. Real campuses do not run flat out; a 70 to 85 percent average load is a reasonable planning band, giving roughly 2.5 to 3.0 TWh. At an assumed delivered price of $70 per MWh, 3.5 TWh costs about $245 million per year; at $100 per MWh it is $350 million. Against a stated capital deployment above $15 billion for the Chiba project, annual energy is on the order of 2 percent of capex, which is why operators will pay a premium for speed: delaying a $15 billion asset a year costs far more than the electricity premium.

Compare the Calvert Cliffs numbers. A 190 MW uprate at an assumed 90 percent capacity factor generates about 1.5 TWh per year (190 times 8,760 times 0.9). That is the energy of roughly a 170 MW continuous load. The 690 MW total under the agreement is a larger commitment that includes existing output, so Amazon is not buying only the increment.

Why the GPU count is the wrong headline

Power-limited sites compete on tokens or training FLOPs per megawatt, not per GPU. Efficiency generations matter: if a successor accelerator delivers substantially more throughput per watt, the same 400 MW buys more capability even with no change in site count. Planners therefore model a campus as a power envelope and refresh the silicon inside it over a 4 to 6 year life, while the electrical and cooling infrastructure, designed for 15 to 25 years, stays. That asymmetry, short-lived compute inside long-lived power, is what makes locking in cheap, firm electricity for 20 years rational.

Nuclear PPAs: What the Contract Actually Buys

A nuclear power purchase agreement is a long-term contract under which a buyer pays an agreed price per megawatt-hour for the output of a nuclear plant, typically a share of it, and receives the associated clean energy attributes. The direct answer to a common confusion: in the Amazon and Constellation arrangement, power continues to flow into the PJM regional grid, and Amazon’s facilities draw from that grid under a retail supply arrangement; the PPA is a financial and supply contract, not a private wire.

Figure 3: Flows in the Calvert Cliffs arrangement. Constellation funds the uprate, the plant injects into PJM, Amazon pays under a 20-year contract, and delivery to Amazon sites occurs through the grid.

Figure 3 separates two layers that headlines blur. The physical layer is plant output injected into PJM, the grid operator for 13 states and the District of Columbia. The commercial layer is cash from Amazon to Constellation, matched to energy and attributes. Because the grid is a shared pool, Amazon’s data centers in Virginia or Ohio do not receive Calvert Cliffs electrons specifically; they receive PJM electrons while the PPA hedges price and underwrites the generating capacity.

The additionality argument

The reason hyperscalers prefer PPAs that finance new capacity over simply buying existing output is additionality: the contract causes new clean generation to exist that otherwise would not. Constellation states the agreement adds 190 MW of new generating capacity, with the plant investment above $3 billion also enabling a license renewal for another 20 years, per its release. That framing is economically important. Buying output from an already-running plant shifts who gets the clean attributes; funding an uprate and life extension adds megawatts to the system. Whether the 190 MW is truly incremental depends on counterfactuals that neither party publishes, so treat additionality claims as a company position rather than an audited fact.

Timing mismatch

New capacity arrives in 2030 to 2032 according to Constellation, while AI demand is arriving now. A PPA with a delivery start years out cannot energize a campus today; it is a hedge against power scarcity and price in the 2030s, and a way to secure scarce long-lived capacity before competitors. That is a different problem from the one JERA’s design addresses, which is energizing compute around 2028 rather than waiting out a grid timeline. Operators increasingly stack the two: an interim supply to start the campus, plus a long-dated PPA to stabilize the cost curve.

Contract structure details that matter

Several terms decide whether a nuclear PPA is a good deal. Price structure, fixed versus indexed to market, determines who carries risk if wholesale prices fall. Take-or-pay clauses require payment for contracted energy even if the buyer does not need it, which matters if a campus is delayed. Performance guarantees and force majeure language handle refueling outages, which for a nuclear unit are scheduled every 18 to 24 months. Regulatory change risk, including how a regulator treats large loads and cost allocation, can reshape economics mid-contract. Public releases for these deals disclose headline MW and term but not price; the pricing and contract terms for the Constellation and Amazon agreement were not in the sources I reviewed.

Behind-the-Meter in Practice: The Chiba Pattern

Behind-the-meter means the load connects on the generator’s side of the utility meter, so the energy does not traverse the public transmission system on its way to the load. For a data center, that trades grid dependence for plant dependence, and it removes the interconnection queue from the critical path. JERA, Dell and RHAELM describe the Chiba Project as sited at JERA’s Chiba Thermal Power Station, with up to 400 MW, operations around 2028, and Apollo Global Management involved as a financing and investment partner for RHAELM according to coverage of the announcement. The partners describe an integrated, standardized design spanning power generation, electrical infrastructure, cooling and AI compute, intended for replication across Japan and later elsewhere. Reported ambitions of up to $140 billion across multiple sites in the 2030s come from press reporting on the plan and should be read as aspiration, not commitment.

Figure 4: Two deployment paths. The grid path waits on new lines and transformers; the co-location path builds a private high-voltage link and islanding design, then both converge on energization and phased GPU ramp.

Figure 4 compares the two schedules structurally. Both begin with site selection. The grid path inserts a study and queue step, then waits for utility construction. The co-location path replaces that wait with engineering work the owner controls: a private high-voltage link, protection schemes, and islanding behavior. Both then energize and ramp GPU clusters in phases, which is how a 400 MW campus actually fills: tens of megawatts at a time as halls complete and hardware arrives.

Why a thermal power station is a good host

A thermal plant brings assets that a greenfield site lacks. It has a high-voltage switchyard sized for hundreds of megawatts. It has water intake and discharge rights or sea-water cooling infrastructure, important because heat rejection at this scale is a water-and-permit problem. It has fuel logistics and an operating organization used to running critical, round-the-clock equipment. And it has land inside an industrial zone where a data center is a lower-friction neighbor than in a residential area.

Two caveats apply. First, the project description names the host as a thermal station but, in the sources I reviewed, does not disclose which units supply the campus, the fuel mix, or any carbon-reduction plan. Gas or other thermal supply raises emissions and fuel-price exposure that nuclear PPAs do not, and the clean-energy claims differ accordingly. Second, a plant that serves a campus has to decide what happens to the power it previously exported. Redirecting output to the data center may reduce supply to the grid unless new capacity is added, which is a system-level cost the campus’s own PUE does not show.

Electrical design of a co-located campus

A well-designed behind-the-meter campus keeps three functions distinct. The first is primary supply, the plant’s generators feeding the campus bus through dedicated transformers and breakers. The second is backup supply, either the grid tie kept energized as a standby path or on-site storage and generation sized for ride-through. The third is power quality management, because AI training loads are unusually spiky.

Synchronized training steps can swing load by tens of percent within seconds as thousands of GPUs alternate between compute and communication phases. On the grid this creates ramp and flicker issues that utilities now scrutinize; on a small, islanded system next to a thermal turbine, it can excite torsional or frequency oscillations in the generator. Mitigation includes battery energy storage that absorbs fast transients, software that smooths power draw by throttling or by padding idle phases with dummy work, and protection settings tuned for sharp load steps. Treat these as design requirements from day one, not commissioning surprises.

Frequency and stability limits

A plant connected to a large grid sees a stiff, stable frequency reference. A campus running mostly or fully islanded must hold frequency itself, and a large load step relative to generation capacity can pull frequency outside the tolerance of GPU power supplies. The usual rule of thumb in power engineering is that the largest single load step should be small relative to the online spinning capacity; the exact acceptable ratio depends on governor response and inertia, and has to come from a dynamic study of the specific plant. That study is a real schedule item, and it is part of why “no grid queue” does not mean “no engineering wait.”

Cooling: The Other Half of the Power Problem

Because essentially all electrical input to an accelerator leaves as heat, the cooling plant is the second-largest electrical consumer on campus and often the binding physical constraint. The direct answer: dense AI racks increasingly rely on direct-to-chip liquid cooling, with chillers or dry coolers rejecting heat to the environment, and the choice of heat-rejection method sets both PUE and water use.

Air cooling struggles as rack densities climb past what air can carry away, so liquid loops move heat from cold plates on the chips to a facility water loop. That loop terminates in either mechanical chillers, which use compressors and consume electricity, or in dry coolers and cooling towers, which rely on ambient conditions and, for towers, on evaporated water. The trade-off is direct: chillers cost power but save water, towers save power but consume water, and dry coolers save both but need cool ambient air or large surface area.

LG Electronics announced the same week that it will build a chiller factory in Virginia for AI data center cooling, part of a reported investment of about $110 million across chiller facilities in the U.S. and Korea; the Virginia location sits in the world’s densest data center market. For this post the relevance is supply chain: chillers, like transformers, are a long-lead component that gates site energization, and new U.S. production capacity suggests vendors see sustained demand. I have not verified production volumes or tonnage for the plant, so do not extrapolate capacity from the headline.

Heat reuse and the thermal station advantage

A co-located campus at a thermal plant has an option greenfield sites rarely do: using waste heat or cooling infrastructure in an integrated way. In principle, a plant’s existing cooling-water system, condenser design or district heating connections could absorb campus heat. In practice, low-grade liquid-cooling return temperatures, roughly 30 to 50 degrees Celsius depending on design, are hard to use in high-temperature steam processes, so most reuse opportunities are for space heating or absorption chillers. The Chiba description says cooling is part of the integrated standard design but does not disclose a heat reuse scheme.

Water and the public license

Water use is where the technical and political arguments meet. A 400 MW campus using evaporative cooling can consume water on the order of millions of liters per day under typical assumptions (estimate it from your own heat load and evaporation rate, since design choices change the answer by several times). For a framework to trace claimed per-query water figures back to sources, see our guide to AI query energy and water footprint with sources traced. Sites in water-stressed regions increasingly choose closed-loop or sea-water designs even at a PUE penalty.

How the Options Compare

The three supply strategies differ along five axes that matter to a planner: time to first power, price certainty, reliability posture, emissions profile and regulatory exposure. The matrix below summarizes qualitative differences; it contains no measured data, and the ratings are my own judgment for a generic large AI campus.

Dimension	Grid plus nuclear PPA (Calvert Cliffs type)	Co-located behind-the-meter (Chiba type)	Fully islanded on-site generation
Time to first power	Slowest, set by utility build and queue	Fastest where the host plant has spare capacity and switchyard	Fast to moderate, set by turbine and equipment lead times
Price certainty	High for 20 years if fixed price, but start date is 2030 or later	Depends on plant fuel and internal transfer pricing	Exposed to fuel price unless hedged
Reliability	Grid redundancy plus plant diversity	Plant dependence; needs grid or storage backup	Entirely self-provided; needs N+1 units
Emissions profile	Low carbon at the contract level	Depends on host plant fuel mix, not disclosed for Chiba	Usually gas; can be high without capture
Regulatory exposure	Large-load tariffs, cost allocation, market rules	Co-location rules, grid-impact review, permitting	Air permits, local siting, fuel infrastructure

Choose the PPA route when you already have grid connections or can wait, and your goal is long-term cost stability and a clean-energy claim. Choose co-location when speed dominates and a suitable host plant exists. Choose islanded generation when neither a plant nor a timely grid connection is available and you can staff and permit the generation. In practice, the largest operators combine these across a portfolio, and a single campus often evolves from one mode to another over its life.

What the deals say about the market

Read together, the two announcements show operators spending at both ends of the timeline. Amazon is paying for capacity that arrives in the early 2030s, while JERA, Dell and RHAELM are paying to compress a 2028 schedule. The common signal is that firm, dispatchable, low-marginal-cost power is now the scarcest input in the AI stack, and that the owners of existing generation sites and licenses hold strategic assets. Existing plants gain optionality: relicensing, uprating, hosting loads or selling long-dated contracts.

Deeper Analysis: Modeling Your Own Power Strategy

If you plan capacity rather than read announcements, a few quantities do most of the work. Treat the following as a modeling template, with inputs you must source from your own engineering and utility data.

Step 1: Define the power envelope

Start with the number of accelerators and their all-in power draw, apply a target PUE, and add a design margin of 10 to 20 percent for growth and measurement error. Distinguish between nameplate and expected average draw. Utilities and interconnection agreements specify peak demand, but your energy bill follows average load. A campus with a 400 MW contract but a 280 MW average has a load factor of 70 percent, and that load factor directly affects the effective price per MWh when demand charges apply.

Step 2: Price the delay

Compute the cost of delay as the capital deployed times a required annual return. For a $15 billion program at a 10 percent hurdle rate (an assumption for illustration), each year of delay carries a notional $1.5 billion cost of capital, plus lost revenue from compute not sold. Compare that with the premium for faster power. If behind-the-meter supply costs an extra $30 per MWh on 3 TWh, that is $90 million per year, an order of magnitude below the delay cost. This is the economic logic behind paying more per megawatt-hour to energize earlier, and it explains why co-location is attractive even when the unit price is worse.

Step 3: Stress-test reliability

Model the loss of the largest single source and the loss of the grid tie. What fraction of the compute load stays online? A training job can checkpoint; an inference fleet serving customers needs continuity. If the answer is that an event at the host plant takes the whole campus down, the design needs either storage sized for ride-through until backup starts, a retained grid tie, or workload tiers that accept partial shutdown. Record the checkpoint interval too: if checkpointing takes 10 minutes of cluster time and occurs hourly, a power interruption costs on average about 30 minutes of progress plus restart overhead.

Step 4: Account for flexibility

Large flexible loads can earn value by curtailing during grid stress. Research and early utility programs have explored data centers that shed or shift load for short periods in exchange for faster interconnection. How much of an AI cluster can safely flex depends on workload mix and contracts; training can pause, latency-critical inference cannot. Where it is feasible, flexibility is a bargaining chip with utilities, and it lowers the effective capacity you need to procure at peak.

Step 5: Think in cost per useful token

Convert everything to the unit you sell. Cost per million tokens or per training run combines energy price, PUE, hardware depreciation and utilization. In many analyses energy is a minority of total cost of ownership compared with accelerator capital, yet it controls siting and timing. Our guide to GPU rightsizing and Kubernetes cost optimization covers the software side: raising utilization is the cheapest way to lower the per-token power cost, because an idle GPU still draws substantial power.

Trade-offs, Gotchas, and What Goes Wrong

The announcements are optimistic documents. Here is what tends to go wrong.

Single points of failure in co-location. Sharing a switchyard with a power plant ties your uptime to its protection systems. A fault that trips the plant can drop the campus unless storage rides through. Avoid designs where campus and plant protection relays are not coordinated by a joint study.

Utility and regulator pushback. Co-located and behind-the-meter loads can avoid paying for transmission and capacity that they still benefit from, and regulators in several U.S. jurisdictions have scrutinized such arrangements. A deal that is legal and attractive in one market may face tariff changes later. Build regulatory change scenarios into the economics, especially for contracts running 20 years.

Timeline slippage. Stated operational dates such as 2028 for Chiba and 2030 to 2032 for the Calvert Cliffs uprate are targets. Nuclear uprates and large construction projects have a mixed history on schedule and budget. Price a delay case of at least a year before accepting a headline date.

PPA mismatch with real consumption. Annual matching of purchased clean energy to consumption hides hourly gaps. A campus can claim 100 percent annual clean energy while drawing fossil-heavy power at night. Hourly or 24/7 matching is a stricter standard, and buyers who care about actual emissions should demand it.

Cooling and water risk. Heat rejection design choices lock in water and power consumption for decades. Extreme-weather derating of dry coolers on the hottest days can force GPU throttling just when demand peaks; design for the 99th-percentile ambient, not the average.

Load transient damage. Synchronized training swings can stress turbines and local grids; without mitigation they cause trips or utility penalties.

Supply chain concentration. Transformers, switchgear, turbines and chillers share limited manufacturers, so one delivery delay cascades through the whole schedule. New factories such as LG’s announced Virginia chiller plant help only over the medium term.

Counterparty concentration. A 20-year PPA ties you to one plant’s operating performance, relicensing outcome and the owner’s finances. Relicensing is a regulatory process whose result is not guaranteed by the contract, even though Constellation presents the investment as enabling another 20 years of operation.

Practical Recommendations

If you are a planner, architect or engineering lead evaluating AI infrastructure power, treat the electrical strategy as a product with an owner, a model and a review cadence, rather than a procurement afterthought. Begin from the envelope: how many megawatts at what ramp, for which workloads. Then ask which supply route reaches that ramp, how reliable it is for each workload tier, and what it costs per useful token including the cost of delay.

Do not over-index on announcements. The headline megawatt figure tells you capacity authority, not delivered compute; PUE, cooling design and workload utilization decide how much of it becomes revenue. Where you can, ask for hourly emissions data and commit to hourly matching if clean-energy claims matter to your customers. Where you rely on a single host plant or a single PPA counterparty, design a graceful degradation plan before you need one.

Checklist before committing to a power strategy:

Define nameplate and average load, with a design margin and a phased ramp schedule.
Model PUE at three ambient conditions, including the 99th-percentile hot day.
Quantify the cost of a one-year delay and compare it with any speed premium.
Run a dynamic stability and load-step study for any islanded or co-located design.
Decide the backup path: grid tie, storage ride-through, or on-site generation, per workload tier.
Read the contract for take-or-pay, price indexation, outage treatment and regulatory-change clauses.
Choose annual or hourly matching for clean-energy claims, and say which in public statements.
Lock long-lead items, transformers, switchgear, chillers, early, and track vendor capacity.
Plan for silicon refresh inside a fixed power envelope over the facility’s life.

Frequently Asked Questions

What is behind-the-meter power for an AI data center?

Behind-the-meter means the data center connects to a generator on the plant side of the utility meter, so electricity reaches the load without crossing the public grid. The benefits are speed and independence from interconnection queues. The costs are dependence on the host plant, the need for your own backup and stability engineering, and possible regulatory scrutiny over grid cost sharing. JERA’s Chiba Project, with up to 400 MW at a thermal power station, is a current example.

What did Amazon and Constellation agree at Calvert Cliffs?

Constellation says it signed a 20-year power purchase agreement with Amazon covering 690 MW of power, including 190 MW of new generating capacity at the 1,790 MW Calvert Cliffs plant in Maryland. The plan involves over $3 billion of investment, with new capacity expected in 2030 to 2032. Power flows to the PJM grid, with Amazon supplied through a retail arrangement. Pricing was not in the announcement.

Is the JERA Dell RHAELM data center at Calvert Cliffs?

No. The JERA, Dell Technologies and RHAELM project is the Chiba Project at JERA’s Chiba Thermal Power Station in Japan, with up to 400 MW, more than US$15 billion of planned capital deployment and operations targeted around 2028. Calvert Cliffs is a separate Maryland nuclear plant tied to Amazon’s PPA with Constellation. The two announcements came within days of each other, which is why they are sometimes conflated.

How many GPUs can 400 MW power?

It depends on PUE and per-accelerator power, neither of which has been disclosed for Chiba. Under illustrative assumptions of PUE 1.2 and 1.8 kW all-in per accelerator, 400 MW supports about 185,000 accelerators. At PUE 1.1 the figure rises to roughly 202,000, and at PUE 1.4 it falls to about 159,000. Treat these as a method for estimation, not a forecast for the real project.

Does a nuclear PPA mean the data center runs on nuclear electrons?

Not physically, in the grid-delivered case. Electrons mix on the grid, so a data center in the PJM region draws from the shared pool while the PPA pays a specific plant and assigns its clean attributes to the buyer. The claim is an accounting one, though a meaningful one when the contract funds new capacity. Hourly matching tests how closely actual consumption aligns with clean generation.

Why not just use the grid?

You can, but large new loads face studies, queues and equipment lead times that can take years, and the grid in some regions has little spare capacity. For an operator with billions of dollars of hardware waiting, speed has a large value. Grid supply remains the cheapest and most flexible long-term option when available, so many campuses use it for steady state and bridge with other sources.