What Does a 10-Gigawatt AI Data Center Actually Mean?

Start with the units: power is not energy

The first source of confusion is the difference between power and energy.

A watt measures the rate at which energy is being consumed or produced.

• 1 kilowatt equals 1,000 watts.
• 1 megawatt equals 1 million watts.
• 1 gigawatt equals 1 billion watts.
• 10 gigawatts equals 10 billion watts.

A 10-gigawatt facility operating at full demand is therefore consuming energy at a rate of 10 billion joules every second.

For a continuous 10-gigawatt load, one hour requires 10 gigawatt-hours, one day requires 240 gigawatt-hours, and one year requires 87.6 terawatt-hours.

That annual figure assumes the site operates at the full 10-gigawatt level every minute of the year. Real facilities do not operate so cleanly. Equipment is installed in phases, individual machines fail or undergo maintenance, workloads fluctuate, and electrical capacity is often reserved before it is fully used.

Nevertheless, 87.6 terawatt-hours provides an essential upper-bound reference. The International Energy Agency estimated that all data centers worldwide consumed approximately 415 terawatt-hours of electricity in 2024. A single campus continuously drawing 10 gigawatts would therefore consume an amount equal to roughly one-fifth of that 2024 global total. The comparison is not a forecast—the global industry is growing rapidly, and the proposed campus would take years to develop—but it demonstrates the scale of the number.

The United States Department of Energy reported that American data centers consumed about 176 terawatt-hours in 2023. A fully utilized 10-gigawatt campus would consume approximately half that amount by itself. Again, this is an engineering comparison, not a prediction that the proposed site will immediately or permanently operate at 10 gigawatts.

Ten gigawatts can describe several different things

Before evaluating any data center announcement, it is necessary to ask what the quoted capacity actually represents.

“Ten gigawatts” might refer to:

1. The maximum electricity that could eventually be delivered to the entire property.
2. The combined nameplate capacity of planned generation assets.
3. The maximum power available to data center buildings.
4. The electrical load of the computing equipment alone.
5. The sum of several development phases that may never operate simultaneously.
6. A long-term planning envelope rather than a contracted initial load.
7. A marketing description of a campus that will begin with a much smaller first phase.

These meanings are not interchangeable.

A site with permission to develop up to 10 gigawatts is not the same as a site with 10 gigawatts of completed substations. Completed substations are not the same as 10 gigawatts of energized data halls. Energized data halls are not the same as 10 gigawatts of installed computers. Installed computers are not the same as 10 gigawatts of average workload.

The difference between these stages may represent hundreds of billions of dollars and many years of construction.

A responsible technical description should therefore separate at least five values:

• Planned campus capacity
• Contracted utility capacity
• Energized building capacity
• Installed IT capacity
• Measured average operating demand

Without those distinctions, a large number can create a false impression of immediate scale.

Facility power is not the same as computing power

A data center consumes electricity through more than processors. The total facility load includes servers and accelerator systems, storage equipment, network switches and optical systems, power-conversion equipment, uninterruptible power supplies, cooling pumps, fans, chillers and heat-rejection systems, lighting, security, building controls and electrical losses between the utility connection and the chips.

The standard metric used to express this relationship is power usage effectiveness, or PUE:

A theoretical PUE of 1.0 would mean every watt entering the property reaches computing, storage or networking equipment, with no cooling, conversion or building overhead. That is physically unattainable in a continuously operating production facility, but efficient designs can approach it.

PUE is useful, but it must not be confused with the efficiency of the computing workload itself. A data center can have an excellent PUE while running poorly utilized or inefficient software.

At a PUE of 1.20, approximately 8.33 gigawatts would be delivered to IT equipment and approximately 1.67 gigawatts would be consumed by cooling, electrical distribution and other facility systems. That overhead alone would be larger than many individual data centers.

A 10-gigawatt headline therefore does not mean 10 gigawatts of GPUs. It means a complete electrical and mechanical system whose useful computational output depends on architecture, utilization, software, cooling and workload quality.

Nearly every watt eventually becomes heat

Electrical power entering a computing facility does not disappear. Processors switch transistors. Memory moves electrical charge. Optical modules generate light and heat. Fans move air. Pumps circulate coolant. Power supplies convert voltage. Almost all of the electrical energy eventually becomes thermal energy that must be removed from the equipment and rejected into the environment.

A 10-gigawatt facility at full load produces approximately 10 gigawatts of heat. That is 10 billion joules of thermal energy every second, or approximately 34 billion British thermal units per hour.

This does not mean one enormous room becomes hot. A campus of this size would be divided into buildings, electrical zones, computing clusters and cooling loops. The thermal challenge is distributed, but the heat still has to be transported away from the chips, transferred through one or more coolant systems and ultimately released or reused.

At conventional rack densities, this would require an extraordinary amount of floor space. AI accelerators have therefore pushed the industry toward much denser rack designs. At rack densities of 50 to 100 kilowatts and above, traditional room-level air cooling becomes increasingly difficult because air has limited heat capacity and requires substantial fan energy to move large quantities of heat.

Liquid cooling changes the heat-transport mechanism. Water and engineered coolants can carry far more heat per unit volume than air. Direct-to-chip systems bring a cooled liquid loop close to processors and accelerators. Rear-door heat exchangers remove heat as air exits a rack. Immersion systems place equipment in a dielectric fluid. Hybrid systems may use liquid for the highest-power components while retaining air cooling for storage, networking and auxiliary electronics.

Liquid cooling does not eliminate heat. It moves heat more effectively from the equipment to the facility’s heat-rejection system.

That final heat-rejection stage may use cooling towers, dry coolers, chillers, evaporative systems, refrigerant systems, cooling ponds, ground-source systems, industrial heat reuse or a combination selected according to local climate and water availability.

Water consumption cannot be inferred from power alone

Large data center discussions often produce a misleading shortcut: high power must mean a specific amount of water use. Water demand is not determined by electrical load alone.

It depends on cooling technology, outdoor temperature and humidity, cooling-water temperature, whether cooling towers are used, whether heat is rejected through dry coolers, whether water is recirculated, the number of concentration cycles in cooling towers, wastewater reuse, local water quality, heat-recovery arrangements and the balance between water efficiency and electrical efficiency.

A closed liquid loop inside a building is not the same as water consumption. The coolant may circulate repeatedly. Water is consumed when the external cooling process loses it through evaporation, blowdown, leaks or treatment.

Water usage effectiveness, or WUE, is commonly expressed as liters of site water consumed per kilowatt-hour of IT energy:

A low-WUE facility in a cool climate may use dry cooling or extensive economization. A facility in a hot, dry region may consume more water to reduce electricity used by mechanical chillers. A design that minimizes site water could indirectly increase power consumption, depending on the heat-rejection technology.

This is why claims such as “the data center will use this many gallons per day” should be treated cautiously unless they specify the assumed operating load, cooling architecture, local weather conditions, whether the number refers to withdrawal or consumption, whether recycled or potable water is included, and whether the estimate covers the initial phase or full campus.

The grid connection is a project within the project

A 10-gigawatt load cannot be treated as a normal commercial utility connection. It requires power-system planning at transmission scale.

A simplified connection path would include:

1. Electricity generation
2. High-voltage transmission
3. Regional switching stations
4. Campus substations
5. Transformers that reduce transmission voltage
6. Medium-voltage distribution across the campus
7. Building-level switchgear
8. Uninterruptible power systems
9. Rack-level power distribution
10. Voltage conversion inside servers

Each stage introduces capacity limits, conversion losses, protection requirements and failure modes.

A campus of this size would probably require multiple independent transmission corridors and several large substations rather than one connection point. Engineers would need to study fault currents, voltage stability, reactive power, harmonics, protection coordination, switching events and the effect of large, rapid load changes.

The transmission operator must determine whether the regional network can deliver the requested power during peak summer demand, peak winter demand, generator outages, transmission-line outages, extreme weather, maintenance, fuel-supply disruptions and unexpected data center load changes.

The question is not merely whether sufficient annual energy exists. The question is whether the required power can be delivered at the correct location during every critical operating condition.

Grid development often takes longer than data center construction. Transformers, breakers, transmission equipment, gas turbines, cables and generation equipment also have their own manufacturing lead times. A developer may secure land and servers but remain unable to energize the facility at the intended scale.

For this reason, the most important evidence in a large data center project is often not the architectural rendering. It is the interconnection agreement, transmission plan, generation contract and construction schedule for the electrical infrastructure.

Generation capacity and data center load are not identical

Suppose a project announces 10 gigawatts of associated generation. That does not automatically guarantee 10 gigawatts of continuous data center supply.

Generation technologies have different operating characteristics. Natural-gas plants can be dispatchable but depend on fuel infrastructure. Nuclear plants provide firm output but require long development periods. Solar and wind output vary. Batteries store energy but do not create it. Hydroelectric output depends on water availability and operational constraints. Geothermal can provide firm power where suitable resources exist.

A continuously operating data center therefore needs a portfolio rather than a slogan. That portfolio may include grid electricity, dedicated generation, long-term power purchase agreements, batteries, demand response, backup generation, curtailable workloads, geographic workload shifting and on-site microgrids.

The electricity physically consumed by a facility should also be distinguished from contractual claims based on renewable-energy certificates or power purchased elsewhere on the grid.

A 10-gigawatt campus cannot rely on conventional backup design alone

Data centers are expected to continue operating during power disturbances. Traditional facilities achieve this through layers: utility feeds, automatic transfer switches, uninterruptible power supplies, batteries, flywheels, diesel or gas generators, and redundant distribution paths.

At 10 gigawatts, simply multiplying a smaller data center’s backup design becomes impractical.

Sustaining a full 10-gigawatt load for only five minutes would require approximately 833 megawatt-hours before accounting for conversion losses, reserve margin and battery-aging constraints. One hour at full load would require 10 gigawatt-hours.

The engineering response is usually not to provide one monolithic backup system. It is to divide the campus into independent power blocks. Each block can contain its own substations, UPS systems, batteries, generators and cooling infrastructure. Critical workloads can be replicated across blocks or locations.

This creates failure domains. A fault in one building should not disable the entire campus. A failed transformer should not collapse several gigawatts of computing. A software deployment should not affect every cluster. Maintenance should be possible without shutting down the site.

At very large scale, resilience depends as much on architecture and workload placement as on electrical redundancy.

The campus would be built in phases

No realistic developer installs 10 gigawatts of computing equipment in one operation.

A campus would be divided into phases according to available grid capacity, generation completion, building completion, chip supply, cooling infrastructure, customer demand, financing, network connectivity and regulatory approvals.

The reported Ohio proposal illustrates this distinction. Earlier reporting described a proposed 10-gigawatt campus, while the first phase was expected to be approximately 800 megawatts and potentially available in 2028. The full project schedule remained unclear, and the negotiations had not been finalized.

An 800-megawatt first phase would already be an exceptionally large computing development. It should not be described as a completed 10-gigawatt facility.

A technically accurate project timeline would identify, for every phase, the target electrical capacity, generation source, interconnection status, building count, energization date, installed IT capacity, cooling capacity, customer commitment, financing status and commercial operation date.

The cumulative campus target may remain 10 gigawatts while the operational load stays far below that level for years.

The computing architecture is a network problem as much as a processor problem

A large AI cluster is not simply a warehouse filled with independent accelerators. Training modern models requires processors to exchange data continuously. Individual GPUs or other accelerators work on portions of a larger computation. The performance of the overall system depends on how quickly they can synchronize gradients, exchange activations, retrieve parameters and access storage.

As clusters expand, communication becomes a limiting factor.

A 10-gigawatt computing campus would require multiple layers of networking: links within a server, among servers in a rack, among racks in a cluster, among clusters in a building, among buildings, to external data sources and users, and to geographically separate recovery locations.

Engineers must manage bandwidth, latency, packet loss, congestion, routing, optical power, switch failures, cable density, topology, synchronization and collective communication patterns.

A cluster may possess enormous theoretical processor performance and still deliver poor application throughput if the interconnect is oversubscribed or badly matched to the workload. This is especially important in large training jobs. Thousands of accelerators can spend time waiting for data from other devices. Adding more processors may produce diminishing returns if communication overhead grows faster than useful computation.

The useful performance of the site must therefore be measured through completed work, not only installed hardware.

Relevant measures include training time for a defined model, tokens processed per second, inference requests completed per second, latency at a defined percentile, energy consumed per training run, energy consumed per million useful outputs, cluster utilization, job failure and restart rates, time lost to communication, and time lost to maintenance.

Power capacity is an input. Completed computation is the output.

Storage and checkpointing become infrastructure-scale systems

Large training and inference systems also require extensive storage. Training workloads read datasets, write logs, store checkpoints and preserve multiple versions of model parameters. Checkpoints allow a training job to recover after a failure rather than restarting from the beginning.

At large scale, checkpointing is not a minor background task. If thousands of accelerators attempt to write state simultaneously, the storage system experiences an enormous burst of traffic.

The architecture may therefore include high-speed local storage, distributed parallel file systems, object storage, metadata services, checkpoint tiers, archival storage, data ingestion systems and replication across failure zones.

Storage design affects processor utilization. An accelerator waiting for data still consumes electricity. A failed checkpoint can waste hours or days of computation. A slow data pipeline can reduce the useful output of an otherwise expensive cluster.

Installed capacity is not useful capacity

One of the most important metrics in computing economics is utilization. A processor that is installed but idle still has capital cost. It may also draw substantial power even when it is not doing full work.

Utilization can be reduced by software errors, hardware failures, waiting for data, network congestion, poor scheduling, incompatible job sizes, maintenance, unavailable storage, insufficient cooling, power caps and lack of customer demand.

High utilization does not automatically mean useful work either. A system can run continuously on low-value jobs, repeated experiments or inefficient software.

A serious evaluation therefore needs several utilization measures:

1. Electrical utilization: average power divided by available power.
2. Hardware utilization: active accelerators divided by installed accelerators.
3. Compute utilization: actual arithmetic activity compared with theoretical processor capability.
4. Cluster utilization: time assigned to productive workloads.
5. Economic utilization: revenue-producing or strategically valuable work compared with total available capacity.
6. Energy productivity: useful computational output per unit of electricity.

At 10 gigawatts, a 1% difference represents 100 megawatts. That is why workload scheduling, software optimization and power management are not secondary concerns. They can change the effective capacity of the campus by an amount comparable with a separate large data center.

Some AI workloads can be flexible; others cannot

An online inference service may have strict latency requirements. Users expect responses immediately. The system must preserve enough active capacity to handle traffic spikes and equipment failures.

A long-running training job may be more flexible. It can sometimes pause, slow down or move to another time period, although frequent interruptions can reduce efficiency or increase failure risk. Batch processing, data preparation and nonurgent experiments may be even more flexible.

This creates an opportunity for grid-aware scheduling. A campus could potentially reduce noncritical jobs during regional peak demand, delay batch work until more power is available, shift workloads among buildings, move work to another geographic region, charge batteries during low-demand periods, cap processor power temporarily or coordinate maintenance with grid conditions.

The feasibility depends on contractual obligations and workload design. Engineers should classify loads according to criticality, maximum interruption time, restart cost, geographic mobility, deadline, latency requirement, data-location restrictions and security restrictions.

The electrical system and the computing scheduler should be designed together.

The economics extend far beyond the building

The cost of a large AI campus is often discussed as though it were a construction budget. The actual capital stack may include land, site preparation, roads and drainage, high-voltage transmission, substations, generation, fuel infrastructure, data center buildings, cooling systems, electrical equipment, servers, accelerators, networking, storage, software, financing fees, interest during construction, spare parts and long-term maintenance contracts.

The computing equipment may represent the largest category, but its useful life can be shorter than that of the buildings and electrical systems. A building may operate for decades. Accelerators may become economically outdated much sooner.

This creates an asset-matching problem. Long-lived infrastructure is financed against workloads and hardware that can change quickly. A project may use a 20-year lease to finance buildings while replacing several generations of computing equipment during that period.

What is a financial backstop?

A financial guarantee is not the same as a direct investment.

Suppose a project company borrows money to build a campus. Lenders evaluate whether future lease payments are reliable. If the tenant has a weaker credit profile than the guarantor, a financially stronger company may promise to cover certain obligations if the tenant cannot.

That guarantee can reduce the lender’s risk. Reduced risk may allow lower interest rates, longer maturities, more debt, faster financing or better commercial terms.

The guarantor accepts contingent risk. It may never pay anything if the tenant performs as agreed. If the project fails, however, the guarantor may face substantial obligations.

This arrangement creates several questions: What exactly is guaranteed? Is the guarantee capped? Does it cover lease payments, construction debt or equipment? When does it become effective? What events allow the guarantor to withdraw? Is it secured by assets? What happens if construction is delayed, technology becomes obsolete, electricity costs rise or the tenant uses less capacity than expected?

At 10-gigawatt scale, financing terms are part of the engineering reality. Without finance, equipment cannot be ordered, power plants cannot be built and construction cannot proceed.

A campus can be technically feasible and financially unsuccessful

Engineering feasibility does not prove commercial viability. A project may successfully secure land, power and equipment but still fail to earn an acceptable return.

Its economics depend on construction cost, financing cost, electricity price, hardware price, hardware replacement rate, customer demand, utilization, model revenue, competition, software efficiency, regulatory requirements and tax treatment.

The total cost includes more than electricity. For an AI service, it may include model development, data preparation, human review, safety testing, networking, storage, customer support, security, hardware depreciation, financing and failed experiments.

A campus can have low electricity cost and still be uneconomic if utilization is poor or hardware becomes obsolete quickly. Conversely, a high-cost facility can be valuable if it supports scarce, high-value workloads.

A 10-gigawatt campus would create concentration risk

Concentration produces efficiency and risk at the same time. A large campus can share power infrastructure, cooling systems, security, network backbones, operations teams, spare equipment, procurement and construction resources.

But placing too much capacity in one location creates common-mode failure exposure. Potential events include regional power failure, fuel interruption, flood, heat wave, wildfire smoke, tornado, cyberattack, network cable failure, water shortage, civil disturbance, regulatory action, supply-chain failure and construction defects.

The campus must therefore be divided into zones with independent systems. Logical separation is also essential. An error in identity management, firmware deployment or cluster orchestration must not spread across the entire property.

Geographic redundancy remains necessary. Even a well-designed 10-gigawatt campus should not become the only location capable of running a critical service.

Cybersecurity becomes industrial cybersecurity

A large AI data center combines information technology with operational technology.

Traditional IT security protects user accounts, applications, data, model weights, software supply chains and administrative systems. Operational technology security protects electrical controls, cooling controls, building management, generator systems, batteries, pumps, valves, fire suppression, physical access and industrial networks.

A compromised cooling controller can shut down computing equipment without touching the model software. A malicious power-control command can create instability or force an emergency shutdown. A supply-chain compromise in firmware can affect thousands of machines simultaneously.

Security design must therefore include segmented networks, strong administrative identity controls, hardware roots of trust, signed firmware, controlled maintenance access, continuous monitoring, physical security, emergency operating procedures, recovery testing and independent safety systems.

At this scale, the campus is part of critical infrastructure.

Environmental claims require physical accounting

A company may contract for renewable energy equal to its annual consumption. That does not necessarily mean the facility operates directly on renewable electricity every hour.

Annual matching can hide hourly differences. A solar project may produce surplus electricity during the day while the data center consumes power overnight from the regional grid. Renewable certificates can establish contractual claims without changing the real-time physical source of electricity.

A rigorous assessment should distinguish annual renewable matching, hourly matching, local generation, grid-average emissions, marginal generation, battery-backed supply, firm zero-carbon supply and unbundled certificates.

The relevant engineering question is: what sources physically meet the load during each hour, especially during low-renewable periods and regional peaks?

Community impact is part of system design

A 10-gigawatt campus affects more than its property. It may require new transmission lines, pipelines, power plants, roads, water infrastructure, housing for construction workers, emergency services, telecommunications expansion, noise controls and land-use changes.

The local benefits may include tax revenue, construction activity and infrastructure investment. The local costs may include higher land prices, congestion, water concerns, noise, environmental effects and pressure on electricity rates.

Data centers also create fewer permanent operating jobs per megawatt than many traditional industrial plants. Job claims should distinguish temporary construction employment from long-term positions.

A responsible agreement should address who pays for grid upgrades, who bears cost overruns, whether residential customers are protected, how water use is measured, what happens during drought, which noise limits apply, what tax concessions are offered, what public reporting is required and what happens if the project is abandoned.

These are not public-relations details. They determine whether the development is sustainable for the surrounding region.

What a 10-gigawatt announcement does not prove

A 10-gigawatt announcement does not prove that:

• Ten gigawatts are available today.
• Ten gigawatts have been contracted.
• Ten gigawatts of buildings are funded.
• Ten gigawatts of servers have been ordered.
• Ten gigawatts will operate continuously.
• Ten gigawatts will be devoted only to GPUs.
• The campus will be completed on schedule.
• The project will be profitable.
• The electricity will be carbon-free.
• The facility will not affect local rates.
• The operator has enough customer demand.
• The financing is final.
• Every announced phase will be built.

The number describes an ambition or design envelope until supporting evidence proves otherwise.

How to evaluate a gigawatt-scale data center announcement

A serious review should begin with the following questions.

Power

1. Is the quoted number generation capacity, grid capacity, facility capacity or IT capacity?
2. Has an interconnection agreement been signed?
3. Which transmission upgrades are required?
4. Which party pays for those upgrades?
5. What generation sources will physically supply the load?
6. What capacity is firm?
7. What happens during peak regional demand?
8. Can the workload be curtailed?

Construction

9. How many phases are planned?
10. What is the capacity of the first phase?
11. Has construction started?
12. Are major transformers and switchgear ordered?
13. Are permits final?
14. What is the realistic energization date?
15. What dependencies could delay the project?

Computing

16. What proportion of facility power reaches IT equipment?
17. What PUE is guaranteed?
18. What rack densities are planned?
19. Which cooling technologies will be used?
20. How will heat be rejected?
21. How are clusters divided into failure domains?
22. What utilization is expected?
23. How will useful compute output be measured?

Water and environment

24. What is the expected WUE?
25. Is the water potable, reclaimed or industrial?
26. Is the published figure water withdrawal or consumption?
27. How does water use change in hot weather?
28. What is the hourly electricity source?
29. Are emissions measured physically or contractually?
30. Is waste heat reused?

Finance

31. Who owns the land?
32. Who owns the buildings?
33. Who owns the computing equipment?
34. Who is the tenant?
35. What commitments are guaranteed?
36. Is financing closed?
37. What happens if the tenant uses less capacity?
38. What happens if equipment costs rise?
39. What happens if the project is delayed?
40. What happens if the expected AI revenue does not materialize?

Until these questions have evidence-backed answers, the headline capacity should be treated as a proposal.

The correct mental model

A 10-gigawatt AI data center is best understood as five interconnected systems.

1. An electrical utility-scale system

It needs generation, transmission, substations, transformers, protection, backup power and grid coordination.

2. A thermal system

It converts almost all consumed electricity into heat and must transport that heat from chips to the environment.

3. A distributed computing system

It combines accelerators, processors, storage and high-speed networks into clusters that must continue operating despite hardware and software failures.

4. A financial system

It depends on debt, equity, guarantees, leases, equipment financing, customer demand and long-term operating revenue.

5. A regional infrastructure system

It interacts with communities, water supplies, roads, labor markets, utilities, regulators and environmental constraints.

Failure in any one of these systems can prevent the other four from delivering value.

Final assessment

Ten gigawatts is an extraordinary planning number.

At continuous full load, it implies 87.6 terawatt-hours of annual electricity consumption. Even with excellent facility efficiency, a substantial portion would be associated with cooling, power conversion, networking, storage and other non-accelerator functions. Nearly all of the electricity would become heat. The grid connection would require transmission-scale engineering. The project would need phased construction, multiple failure domains, industrial cooling systems and financing normally associated with major energy infrastructure.

The reported Ohio proposal should therefore not be interpreted as a single data center that will suddenly switch on at 10 gigawatts.

It is better understood as a potential multi-year campus program whose full development would depend on power plants, transmission, buildings, cooling, chip supply, financing, regulatory decisions and sustained demand for AI computing.

The useful question is not whether 10 gigawatts sounds large.

It is whether each phase can be financed, powered, cooled, connected, utilized and operated reliably—and whether the computational and economic value produced by the campus justifies the physical resources required to build it.

That is the engineering meaning of a 10-gigawatt AI data center.