The AMICE energy stack behind an AI answer
Follow an AI data center from power plant to utility, grid, fuel, cooling system, water source, contract, and unanswered reporting question.
Published 2026-07-10 · Updated 2026-09-18
Loading content...
Follow an AI data center from power plant to utility, grid, fuel, cooling system, water source, contract, and unanswered reporting question.
A nuclear contract, a renewable certificate, and a grid connection tell different stories about AI electricity. Follow the wires and the accounting separately to see what each claim establishes.
A company can call its AI "powered by nuclear," "100% renewable," or "waterless." I want to know exactly what it is claiming.
Was it describing the electricity physically entering a campus? A regional annual average? A utility's generation portfolio? A long-term contract that helps finance a power plant? Renewable certificates retired somewhere else? A cooling loop that does not evaporate water on site? Those can all be real and valuable. They are not interchangeable.
This is the Energy layer of the AMICE stack: Applications, Models, Infrastructure, Chips, and Energy. It starts where the main stack diagram stops and follows the wires, pipes, contracts, fuel, water, and heat farther down.
The energy primer retains its August 13, 2026 cutoff; the campus evidence includes later, individually dated research batches. The prose and further reading were reviewed on September 17, 2026. Every interactive claim opens its source. Undisclosed allocation stays undisclosed.
The September 18 documentation review adds grid reports, government generation data, and utility and power-contract disclosures. Each record keeps its period and geography. Selecting a cloud provider and a region also exposes separate electricity references for that place, without assigning the region's annual mix to an individual request.
204 additional sources reviewed September 18, 2026. Each record names the workload or geography it supports. Training records describe training; availability and development plans have their own labels.
Your location can suggest a region to research. Endpoint configuration, residency rules, capacity, and cross-region routing determine where a request can run. Nearby data centers are not evidence of your application's serving location.
66 of 204 sources · showing 1-20
Power
An instantaneous rate. A 1 GW interconnection is a ceiling or capability only if the source says so.
Energy
Power over time. A 100 MW load sustained for one hour uses 100 MWh.
PUE
Facility overhead. It does not measure useful computing, carbon, or water.
Site WUE
Operational site water consumption per IT energy within a declared boundary.
A megawatt is a rate. A megawatt-hour is energy over time. A campus announcement measured in gigawatts may describe a future buildout, an interconnection ceiling, gross facility service, or IT load. Those meanings cannot be added together without preserving the basis.
PUE and WUE are ratios with boundaries. A low PUE says the facility uses relatively little energy beyond its IT equipment; it does not say the computing is useful, the electricity is clean, or the water impact is low. Site WUE measures operational water at the data-center boundary per unit of IT energy. Source WUE adds modeled water associated with producing electricity. Neither includes every liter used to manufacture chips or construct the facility unless the methodology explicitly says so.
The IEA's 2026 update on energy and AI estimates global data-center electricity consumption at 485 TWh in 2025 and projects about 950 TWh in 2030, roughly 3% of global electricity demand. That is all data centers, not AI alone, and 2030 is a projection rather than a measured result.
The same report describes efficiency gains alongside growth in usage and more demanding tasks. A cheaper text response does not settle the footprint of a long-running agent or generated video. Nor does a global share tell a local utility whether it has a spare substation. The useful comparison keeps the workload, place, and year attached.
The simplest physical story is familiar: a generator makes electricity, transformers raise the voltage, transmission lines move power over distance, substations lower the voltage, and a local or high-voltage connection reaches the customer. EIA's delivery explainer is a good plain-language tour.
The institutional story is messier. A power-plant owner, plant operator, transmission owner, transmission operator, grid market, balancing authority, distribution utility, retail supplier, and load-serving entity can be different organizations. In other places, a vertically integrated utility performs several of those jobs.
The grid also has to balance supply and demand continuously. In organized U.S. markets, an RTO or ISO dispatches generators and runs markets for energy, capacity, and stability services under different rules. FERC's 2026 market guide is explicit that those are different products. ERCOT is unusually easy to name and unusually easy to misunderstand: for most of Texas, it is the interconnection, balancing authority, and market operator, but it is not the local utility billing every campus.
Follow one data center four ways
Select a lens, then focus or click any step. The people and assets overlap, but the relationships do not. A line can mean physical delivery, dispatch, payment, an attribute, or a water flow—never all five at once.
What physically keeps the GPUs running?
Power is generated, synchronized onto an interconnected grid, transformed across voltage levels, delivered to a campus, conditioned, and turned almost entirely into heat.
Equal width = relationship, not quantity
Mine, well, reservoir, wind, sun, or uranium supply chain
A generator needs an energy source. Wind and sunlight arrive without fuel delivery; gas, coal, uranium, and backup diesel have extraction, processing, transport, and storage chains.
Consumer analogy: The ingredients before the kitchen can cook.
Limits of this evidence
A regional fuel category does not identify the mine, gas basin, uranium origin, or refinery supplying one plant.
The implication for a consumer is that an outage or price increase can begin far from the app. The model provider might run short of compute because a utility cannot energize a substation on time. A heat wave can raise cooling demand while the grid is already stressed. A gas constraint can affect marginal generation. A transformer delay can matter even when there is plenty of annual energy.
To report on who powers a data center, ask:
A power plant converts an energy source into electricity. A plant may contain several generator units. Its owner can hire another company to operate it, buy fuel from several suppliers, and sell the output into a market rather than to one named customer.
A utility is an institution with defined service obligations and assets. Depending on the jurisdiction, it may own generation, transmission, distribution, or all three; buy power from independent generators; procure through a wholesale market; meter the campus; and recover costs through a tariff. Municipal utilities, cooperatives, investor-owned utilities, and federal systems such as TVA do not all have the same structure.
A balancing authority is responsible for matching generation, load, and interchange within its area. An RTO or ISO can operate a wholesale market and transmission system across many utilities. A retail supplier may sell the electricity while a different distribution utility delivers it over local wires.
A diagram should keep those roles separate rather than draw a connection from MLGW to xAI and silently mean generation, market operation, retail billing, water service, and every upstream TVA power plant. Each relationship needs its own label.
Almost all electricity entering the servers and facility eventually becomes heat. The cooling system has to do three separate jobs:
A computer-room air conditioner, or CRAC, contains compressorized refrigeration. A computer-room air handler, or CRAH, moves room air across a chilled-water coil while the chiller sits elsewhere. An air-side economizer uses favorable outdoor conditions to reduce or bypass mechanical refrigeration. "Free cooling" still needs fans, pumps, filters, controls, and sometimes water.
Direct-to-chip cold plates and immersion cooling can move heat efficiently away from dense accelerators. They do not specify the final heat sink. Google's Brazos system demonstrates one surprising path: liquid captures component heat, then a liquid-to-air exchanger sends it into the data hall's hot aisle so the existing air system can finish the job. DOE illustrates the opposite possibility: a closed rack loop transfers heat through a coolant distribution unit and facility loop to an evaporative cooling tower.
Build the heat path
Choose how heat leaves the electronics and where it finally goes. Every combination below is a conceptual architecture, not a performance estimate for a named facility.
Your conceptual system
Heat source
GPU, CPU, memory, power
Capture + transport
Direct-to-chip cold plates
Final destination
Dry outdoor air
chip → cold plate → IT coolant loop → CDU → facility loop → outdoor coil → air
Site-water tendency
Near-zero routine site cooling water
Electricity tendency
Climate and design dependent
Near-zero operational cooling water does not include sanitation, construction, chip fabrication, or water used to generate electricity.
Cooling-tower mass balance
The illustrative balance fixes evaporation at 100 units and ignores drift and leaks. Blowdown removes concentrated minerals. Increasing cycles of concentration reduces blowdown; it does not eliminate evaporation.
Evaporation
100.0
Consumed to atmosphere
Blowdown
50.0
Sent to treatment / discharge
Makeup
150.0
Evaporation + blowdown
Dry cooling rejects heat into outdoor air without routine evaporation. It can use more fan energy and struggle more during hot weather. Evaporative cooling spends water to reduce the compressor work needed to approach a useful temperature. A water-cooled chiller can therefore improve energy efficiency while increasing onsite water consumption. There is no universally best design without climate, water stress, reliability, load, and grid context.
Water withdrawal is the water taken from a river, aquifer, reservoir, public system, or reclaimed-water network. Water consumption is the portion not returned to the immediate water system, commonly because it evaporates.
A once-through power plant can withdraw a very large volume and return most of it warmer. A recirculating cooling tower withdraws far less but consumes a larger share through evaporation. EIA's comparison also points to thermal discharge, entrainment, and impingement. A single consumption number misses those impacts.
The cooling-tower balance is physical. Evaporation leaves dissolved minerals behind. Operators drain concentrated water as blowdown and replace evaporation and blowdown with makeup water. Increasing cycles of concentration can cut blowdown, but it cannot erase the evaporation used to reject heat.
Reclaimed water changes the source; it does not make the system water-free. The EPA's Quincy, Washington case study follows treated cooling-tower blowdown through softening, filtration, reverse osmosis, reuse, and brine ponds. The system reduces potable groundwater demand, but it still needs seasonal makeup water and produces concentrated residuals.
Lawrence Berkeley National Laboratory estimated that all U.S. data centers consumed about 66 billion liters of water directly in 2023. This total includes workloads other than AI. Its model estimated nearly 800 billion liters indirectly through electricity generation, using an average 4.52 liters per kilowatt-hour. The modeled upstream total was roughly twelve times the direct total.
The LBNL report assigned regional grid mixes because facility-level information was unavailable and excluded individual PPAs and behind-the-meter generation from that calculation. It is a national modeled attribution. It cannot supply a meter reading for AI, a provider, a model, or a prompt.
A facility can reduce onsite WUE with dry cooling while using more electricity, shifting some water use to power plants. Another facility can use evaporative cooling to reduce electricity but consume more water in a stressed local basin. A useful comparison needs both site and source boundaries, plus geography and season.
The countries behind AI electricity depend on the fuel and the stage being measured.
Natural gas moves from wells through gathering lines, processing plants, interstate or intrastate pipelines, storage, and plant laterals. Pipeline gas is commingled. National trade statistics cannot tell us which basin or producer supplied one turbine.
Coal is more traceable because plant fuel receipts can identify suppliers, mine locations, quality, and transport. Even then, a national import share is not a plant mix.
Nuclear fuel has the longest passport: mine, mill, conversion, enrichment, fabrication, reactor loading, and inventory. Ore origin and enrichment origin are different facts. EIA's latest 2025 delivery data found U.S. reactor owners obtained most uranium abroad, led by Canada, Kazakhstan, and Australia. A nuclear power contract does not reveal which ore batch or enrichment service produced a specific hour of electricity.
Backup diesel adds a refinery and terminal chain. U.S. refineries produced most of the ultra-low-sulfur diesel used domestically in 2025, while most imported product came from Canada. But refineries blend domestic and imported crude, and products move through shared pipelines and terminals. The supplier truck cannot prove the country of the crude in a campus tank.
Wind, sunlight, and flowing water do not arrive as a delivered fuel. Their turbines, modules, inverters, batteries, transformers, and critical minerals still have international manufacturing chains. Equipment origin belongs in a supply-chain ledger, not in the country-of-origin field for an operating megawatt-hour.
Fuel passports
The country bars are national context at the stated boundary—not a claim about a named plant. Select a fuel to see why uranium, gas, coal, and backup diesel need different provenance questions.
2025 deliveries to owners and operators of U.S. civilian reactors
Fuel for nuclear generators appearing in regional supply and clean-energy contracts.
Denominator: National origin of delivered uranium; enrichment-service origin is a separate dataset
Findings
U.S. reactor owners obtained most uranium abroad. Enrichment was also international: EIA separately reports foreign separative-work origins, including Russia, France, the United Kingdom, and the Netherlands.
Traceability limits
A nuclear PPA or regional nuclear share does not reveal which batch of uranium, conversion plant, enricher, or fuel fabricator supports the contracted output.
The bars intentionally use the strongest comparable national boundary available and then stop. A plant-specific investigation would look for EIA fuel receipts, pipeline interconnections and nominations, utility filings, air permits, contracts, and tank records. Nuclear investigations also need records for each fuel-cycle stage.
It is tempting to join separate press releases into one confident route: an app connected through a model and cloud to a campus, chip, utility, and fuel source. That route only exists when the sources connect those exact scopes.
Abilene has strong public evidence for a campus, an ERCOT interconnection ceiling, onsite natural-gas generation, a partly operational OpenAI workload relationship on Oracle Cloud, and a closed-loop cooling design. It does not have a public hourly allocation among those energy sources.
Memphis has a named utility chain connecting xAI to MLGW and TVA, service requests, and a proposed shared recycled-water project. TVA publishes an annual generation and purchase portfolio. That portfolio is not an hourly Colossus mix, and the shared water system cannot be assigned entirely to xAI.
Meta's Hyperion plan names a utility buildout involving new gas units, batteries, nuclear uprates, and purchased power. It is under construction. Those categories are not current generation for an operating 5 GW campus.
AWS describes outside-air cooling and seasonal water operation at the disclosed Indiana part of Project Rainier. It does not publicly map the whole multi-site cluster, serving utility, hourly generation, or every Claude workload to that campus.
Four real AI campuses
These are evidence profiles, not rankings. One campus has a named interconnection but no hourly source allocation; another has a utility chain but no measured water; a future project has a power plan but no live load.
Abilene, Texas, United States
Workload boundary: OpenAI workloads on Oracle Cloud infrastructure; campus development and operation involve separate parties.
Physical system
Lancium reports an ERCOT-approved 1.2 GW interconnection. The campus also includes onsite natural-gas generation; grid and renewables are cited as parts of the supply approach.
Market + utility
ERCOT is simultaneously the Texas interconnection, balancing authority, and organized market for most of the state. Its rounded 2025 annual generation context was about 41% gas, 23% wind, 14% solar, 13% coal, 8% nuclear, and 1% other—not an Abilene hourly or contractual mix.
Contract boundary
No public source allocates OpenAI workload electricity among ERCOT purchases, onsite gas, or contractual renewable instruments.
Fuel provenance
Onsite natural-gas producer, basin, pipeline, and contract terms are not disclosed in these sources.
Cooling path
Closed-loop, non-evaporative direct-to-chip design with a one-time fill that is recirculated.
Water boundary
The cooling design can reduce ongoing onsite cooling water; sanitation, construction, chip-fab, and power-sector water remain outside the claim.
Strongest documented boundary
Named campus, ERCOT interconnection ceiling, onsite-gas role, operating phase, cooling design, and OpenAI/OCI workload relationship.
Allocation still undisclosed
Current campus draw, IT load, hourly source mix, model share, generator dispatch, fuel origin, and measured water remain undisclosed.
Questions worth reporting
The unknown fields identify the documents needed to fill the gaps: interconnection agreements, tariffs, utility commission dockets, air and water permits, pipeline and fuel contracts, construction milestones, hourly meter data, and workload allocation.
A power-purchase agreement can transfer physical energy within a market, settle financially, support new or existing generation, transfer environmental attributes, or combine several of those functions. Once electricity enters a shared grid, it follows the electrical network regardless of the marketing claim.
A renewable-energy certificate assigns an environmental attribute separately from undifferentiated grid power. Annual matching can pair a year of electricity use with a year of certificate volume while individual hours remain fossil-heavy. Hourly carbon-free matching asks a stricter question, but it is still an accounting and procurement method rather than a dedicated wire.
Onsite generation is different again. A gas turbine or fuel cell can physically serve the campus, export to the grid, provide backup, or operate alongside utility service. A battery stores electricity from some prior source; it does not create primary energy. A solar array needs the controls and protection for intentional islanding if it is supposed to keep running through a grid outage.
For the accounting vocabulary, start with the GHG Protocol's Scope 2 Guidance. It sets rules for purchased-energy emissions and the quality of contractual instruments. A compliant emissions inventory still answers a different question from which generators physically supplied a particular campus hour. Keep the reporting method and the physical supply claim in separate fields.
Claim decoder
Choose a common claim. The decoder keeps what is useful, removes what the evidence does not establish, and turns the gap into a better question.
The claim
“A PPA means the named plant physically powers the data center.”
The agreement can settle financially or transfer attributes through a shared grid. A dedicated physical line requires separate evidence.
Ask this instead
Where are the generator and load, how is the contract settled, and what is matched hourly?
The editorial rule is simple: every line gets one meaning. Physical connection, market dispatch, utility service, money, environmental attribute, ownership, regulation, fuel transport, and water flow use different labels and visual styles.
You do not need to memorize wholesale power markets to understand the implications.
The evidence desk below turns every missing edge into a searchable question and keeps the source ledger downloadable. It favors regulators, utilities, government datasets, technical standards, filings, and company-primary disclosures. Company claims remain labeled as company claims; regional context remains regional context.
Reporter notebook
Reviewed Aug 13, 2026. Public evidence, not a meter behind a model. Physical delivery, market dispatch, utility service, contracts, and environmental accounting are separate lenses.
12 reporting questions
30 primary or authoritative sources
This page will keep changing because the system is changing. New generation will enter service, announced campuses will stall or energize, cooling designs will move from press releases to permits and meters, and contracts will be amended. The stable part should be the method: preserve the boundary, date, status, denominator, and source for every claim.
These reading notes were checked on September 17, 2026. The interactive source ledgers retain their separately dated records.
IEA: the 2026 energy and AI outlook. Historical estimates and forward scenarios, with discussion of grid, equipment, and financing constraints.
GHG Protocol: Scope 2 Guidance. Purchased-electricity emissions accounting and contractual quality criteria. Consult the linked revision process separately from the published guidance.
EIA: electricity data and tools. The Hourly Electric Grid Monitor provides balancing-authority context. It does not identify an individual campus's supply.