Glossary

annual vs hourly matching

Two approaches to renewable energy accounting. Annual matching means purchasing renewable energy certificates (RECs) equal to total annual consumption—but the renewable generation may occur at different times than consumption. Hourly (or 24/7) matching requires renewable generation to actually occur during each hour of consumption. Annual matching is more common but criticized as "greenwashing" because it doesn't guarantee carbon-free operations in real-time.

attention mechanism

A mathematical process that allows AI models to focus on relevant parts of input when generating output. Attention mechanisms enable transformers to understand context by weighing the importance of different tokens relative to each other. This is why AI can understand that "bank" means different things in "river bank" versus "bank account."

baseline power

Generation that runs continuously at constant output, unable to vary production based on demand. Nuclear plants are classic baseline power: they run at full capacity 24/7 because ramping up and down is technically difficult and economically inefficient. Data centers' constant demand makes them natural customers for baseline power.

behind-the-meter

Power generation equipment located on a customer's property, connected directly to their load rather than through the utility grid. Behind-the-meter generation bypasses interconnection queues because it doesn't feed power back to the grid. Some data centers are deploying natural gas generators or solar arrays behind-the-meter to accelerate operations while grid connections are pending.

brownfield

Previously developed land, often former industrial sites, that may have environmental contamination or complex title issues. Policy often favors brownfield redevelopment to preserve farmland, but data center developers generally avoid brownfields due to remediation costs, timeline uncertainty, and liability concerns.

capacity

The maximum power demand a facility can draw at any moment, measured in watts (typically MW or GW for data centers). Capacity differs from consumption: a 1 GW data center has the *capability* to draw 1 billion watts, even if it doesn't always operate at full capacity. Grid infrastructure must be built to serve peak capacity, making it the critical constraint for data center siting.

capacity factor

The ratio of actual output to maximum possible output over a time period, expressed as a percentage. Nuclear plants achieve 90%+ capacity factors (running nearly constantly). Wind turbines typically achieve 30-40% (generating only when wind blows). Solar panels achieve 20-25% (generating only during daylight). This metric explains why a 1 GW solar farm doesn't replace a 1 GW nuclear plant.

capex / opex (Capital Expenditure / Operating Expenditure)

The two categories of business spending. Capex includes one-time investments in physical assets (land, buildings, equipment); opex includes ongoing costs (electricity, labor, maintenance). Data centers are extremely capital-intensive: a 1 GW facility might cost $5-10 billion in capex but generate decades of revenue. Understanding this split explains why access to capital is crucial for infrastructure development.

carbon intensity

The amount of carbon dioxide emitted per unit of electricity generated, typically measured in grams CO2 per kilowatt-hour. Carbon intensity varies dramatically by generation source: coal (~1,000 g/kWh), natural gas (~400 g/kWh), nuclear and renewables (~0-50 g/kWh). Because marginal generation is often natural gas, new electricity demand frequently has higher carbon intensity than grid averages suggest.

CERCLA (Comprehensive Environmental Response, Compensation, and Liability Act)

The 1980 federal law (commonly called "Superfund") that governs cleanup of contaminated sites and assigns liability. Under CERCLA, anyone who owned or operated a contaminated site—or even transported waste to it—can be held liable for cleanup costs, regardless of fault. This "strict liability" makes brownfield development legally risky.

colocation

A business model where a data center operator provides space, power, and cooling, which customers then fill with their own servers. Multiple tenants share a single facility. Contrasts with "hyperscale" where a single company builds and operates facilities exclusively for its own use. Equinix and Digital Realty are major colocation providers.

conditional use permit

Similar to a special use permit, a conditional use permit (CUP) authorizes a land use that is allowed in a zoning district only when specific conditions are met. Conditions may include setbacks, noise limits, landscaping requirements, traffic mitigation, or community benefit agreements. The discretionary nature of CUPs gives local governments significant power to shape data center projects or deny them entirely.

consumption

Total energy used over time, calculated as capacity multiplied by hours of operation. Measured in watt-hours (Wh), kilowatt-hours (kWh), or megawatt-hours (MWh). A 1 GW data center operating for a year at 90% capacity consumes approximately 7,884 GWh—more electricity than many small countries.

cooling tower

Large structures that dissipate heat from data center cooling systems by evaporating water. Cooling towers are the primary source of data center water consumption, using 3-5 million gallons per day for a large facility. They work by spraying water over fill material while fans draw air through, evaporating water to remove heat. Visible vapor plumes from cooling towers often prompt community concerns about visual impact and water usage.

CUDA (Compute Unified Device Architecture)

NVIDIA's proprietary programming framework that enables developers to harness GPU computing power for general-purpose applications including AI. CUDA's 15+ year head start and extensive software ecosystem create significant switching costs, contributing to NVIDIA's market dominance even when competitors offer comparable hardware.

direct-to-chip cooling

A cooling technology where liquid coolant flows through pipes mounted directly on processors, removing heat at the source before it spreads to surrounding air. This enables higher power densities than traditional air cooling because liquid absorbs heat more efficiently than air. Many new AI data centers use direct-to-chip cooling for GPUs.

dispatchable generation

Power plants that can increase or decrease output on demand to match changing electricity consumption. Natural gas "peaker" plants can ramp from zero to full output in minutes. Solar and wind are generally not dispatchable because output depends on weather conditions rather than operator control.

distribution

Lower-voltage power lines (typically 4kV to 34.5kV) that deliver electricity from substations to homes and businesses. Distribution networks serve the "last mile" but have limited capacity—typically tens of megawatts per feeder. This is why gigawatt-scale data centers can't locate in cities: urban distribution networks can't handle the load.

easement

A legal right to use another person's land for a specific purpose without owning it. Utility easements allow power companies to run transmission lines across private property. Easements "run with the land," meaning they transfer to future owners. Property owners typically receive one-time compensation but retain title and tax liability. Easements can significantly restrict property use and affect resale value.

EBITDA (Earnings Before Interest, Taxes, Depreciation, and Amortization)

A measure of operating profitability that strips out financing decisions and accounting conventions, commonly used to value data center companies. Valuation multiples (purchase price divided by EBITDA) rose from 12-15x in the REIT era to 25x+ during the private equity boom, reflecting confidence in long-term demand growth.

eminent domain

The power of government (or authorized utilities) to acquire private property for public use, with compensation required under the Fifth Amendment. While typically associated with roads and schools, eminent domain can be used for transmission lines serving data centers. "Just compensation" is usually based on fair market value before the project, not the value the property gains from infrastructure proximity. Property owners can challenge the taking or the compensation amount in court.

ERCOT (Electric Reliability Council of Texas)

The grid operator for most of Texas, unique in being almost entirely isolated from the rest of the US grid. This isolation means ERCOT can approve new connections faster (no interstate coordination) but also can't import power during emergencies—as demonstrated during Winter Storm Uri in 2021. Many data centers are choosing Texas despite reliability concerns due to faster interconnection timelines.

export controls

Government restrictions on selling certain technologies to foreign countries, particularly adversaries. The US has imposed increasingly strict export controls on advanced AI chips, blocking sales of NVIDIA's most powerful GPUs to China. These controls are reshaping global AI development and driving domestic infrastructure investment as companies seek secure supply chains.

FERC (Federal Energy Regulatory Commission)

The independent federal agency that regulates interstate electricity transmission, wholesale power markets, and natural gas pipelines. FERC sets rules for grid interconnection, transmission planning, and market operations. Recent FERC orders (2023, 1920) have attempted to address interconnection queue backlogs and improve long-term transmission planning.

GPU (Graphics Processing Unit)

A specialized processor originally designed for rendering graphics but now the dominant hardware for AI workloads. GPUs excel at the parallel matrix multiplication operations that transformers require. NVIDIA controls approximately 80-90% of the AI GPU market through products like the H100 and Blackwell architectures.

greenfield

Undeveloped land, typically agricultural, that has never been used for industrial purposes. Data center developers strongly prefer greenfield sites because they offer clean title, no contamination, predictable timelines, and flexibility in facility design. The conversion of farmland to data centers is one of the book's central tensions.

HBM (High Bandwidth Memory)

A type of computer memory that stacks multiple layers of DRAM chips vertically, enabling extremely high data transfer rates (up to 3 terabytes per second on modern GPUs). HBM is essential for AI inference because model weights must be rapidly accessed during processing. Memory bandwidth is often the limiting factor in AI performance, not raw computation speed.

hyperscale

Massive data center facilities (typically 100+ MW, increasingly 1 GW+) built and operated by a single company to serve millions of users. Amazon, Google, Microsoft, and Meta operate hyperscale facilities. The term reflects the enormous scale of operations—thousands of servers, dedicated power substations, and operations teams.

immersion cooling

The most aggressive cooling technology, where servers are completely submerged in specialized non-conductive (dielectric) fluid. The fluid absorbs heat directly from all components, enabling extreme power densities (300+ kW per rack) while achieving PUE ratios near 1.1. Still emerging for large-scale deployment but increasingly important for AI workloads.

incentives

Tax breaks, subsidies, or other benefits offered by governments to attract business investment. Data center incentives typically include sales tax exemptions on equipment purchases, property tax abatements, and infrastructure investments. Virginia waives over $900 million annually in data center sales taxes. The "race to the bottom" among states competing for investment is a recurring policy debate.

inference

/IN-fer-ens/

The process of using a trained AI model to generate predictions, responses, or outputs based on new input. Unlike training (which happens once), inference happens billions of times daily as users interact with AI systems. A single ChatGPT response requires inference across all 96+ transformer layers, consuming significant computational resources.

interconnection queue

The formal process by which new generators or large electrical loads (like data centers) apply to connect to the grid. Projects must undergo studies to ensure they won't destabilize the system. Interconnection queues are severely backlogged: PJM has 300+ GW of projects waiting, with approval times of 4-8 years. This backlog is a major constraint on data center development.

Jevons Paradox

The observation that efficiency improvements often lead to increased total consumption rather than decreased consumption. Named after 19th-century economist William Stanley Jevons, who noted that more efficient steam engines led to more coal use, not less. For AI, this means algorithmic improvements that reduce per-query energy consumption may simply enable more queries, increasing total energy demand.

latency

The time delay between sending an input to an AI system and receiving a response. For user-facing applications, latency matters enormously—users expect responses in milliseconds, not seconds. This drives data center placement decisions, as network distance affects latency.

load

The amount of electrical power being consumed at any given moment, measured in watts. A "load" also refers to any device or facility that consumes electricity. Data centers are unusual loads because they operate at near-constant power 24/7 (high load factor), unlike most industrial or residential customers whose demand varies throughout the day. This constant demand creates unique challenges and opportunities for grid operators.

marginal generation

The power plant that increases output to meet the next unit of electricity demand. When you turn on a light, somewhere a generator spins slightly faster—that's marginal generation. In most US regions, marginal generation is natural gas, meaning new electricity demand (including from data centers) is often met by fossil fuel generation regardless of renewable energy purchases.

model weight

The collection of all parameter values that define a trained AI model's behavior. Model weights are the "knowledge" encoded through training. When a model is deployed for inference, these weights are loaded into memory (often HBM) and used to process each input.

N+1 redundancy

A design principle where systems have one more unit than strictly necessary, ensuring operation continues if any single component fails. "2N" redundancy doubles everything. Data centers apply redundancy to power feeds, cooling systems, network connections, and other critical infrastructure. Higher redundancy increases costs but improves reliability.

parameter

An individual learned value within an AI model's neural network. Modern large language models contain billions to trillions of parameters. GPT-3 has 175 billion parameters; GPT-4 is estimated to have over 1 trillion. More parameters generally enable more sophisticated capabilities but require proportionally more computing resources for inference.

Phase I/II ESA (Environmental Site Assessment)

Standardized environmental investigations required before property transactions. Phase I reviews historical records and site conditions to identify potential contamination. Phase II involves actual soil and groundwater sampling if Phase I indicates concerns. These assessments can take months and, if contamination is found, trigger remediation requirements.

PILOT (Payment in Lieu of Taxes)

A negotiated agreement where a property owner makes fixed annual payments to local government instead of standard property taxes. PILOT agreements often provide significant tax savings for data centers (sometimes 75-89% reduction) in exchange for guaranteed long-term revenue streams for municipalities. These agreements are controversial because they shift tax burden to other property owners.

PJM Interconnection

The largest RTO in North America, serving 67 million people across 13 states from Illinois to New Jersey. PJM manages 88,000 miles of transmission lines and a $50 billion annual electricity market. "Data Center Alley" in Northern Virginia falls within PJM territory, making it central to AI infrastructure development.

power purchase agreement (PPA)

A long-term contract (typically 10-25 years) between an electricity generator and a buyer, locking in price and quantity. Tech companies use PPAs to secure renewable energy and hedge against electricity price volatility. Corporate PPAs have driven massive renewable energy development, though critics note that "virtual" PPAs (financial instruments) don't always result in new generation or reduce the buyer's actual grid emissions.

private equity

Investment firms that pool capital from institutions (pension funds, endowments, sovereign wealth funds) to acquire and operate companies outside public markets. Blackstone, KKR, and GIP have deployed tens of billions into data center infrastructure, attracted by stable cash flows and long-term growth potential. Private equity's longer time horizons enable larger, riskier projects than public markets typically support.

PUE (Power Usage Effectiveness)

The ratio of total facility power to IT equipment power, measuring data center efficiency. A PUE of 2.0 means half the electricity goes to cooling and overhead; 1.0 would mean perfect efficiency (impossible in practice). Modern air-cooled facilities achieve 1.4-1.6; advanced liquid-cooled facilities can reach 1.1-1.2. Lower PUE means less wasted energy.

rack

A standardized cabinet for housing computer servers, typically 42 "units" (42U) tall, with each unit measuring 1.75 inches in height. Racks provide the fundamental organizing structure for data center equipment. Modern AI racks can consume 60-140+ kW of power—far exceeding the 5-15 kW of traditional server racks.

REIT (Real Estate Investment Trust)

A company that owns and operates income-producing real estate and is required to distribute most profits to shareholders. Data center REITs like Equinix and Digital Realty pioneered the industry's growth through public markets, but their need for predictable cash flows limited their ability to pursue speculative mega-projects. Most have been taken private.

remediation

The process of cleaning up environmental contamination at a site, which may involve soil removal, groundwater treatment, vapor barriers, and decades of monitoring. Remediation costs range from $50-500+ per cubic yard of contaminated soil. Timeline uncertainty (2-4+ years) and the risk of discovering additional contamination during construction make brownfield development particularly challenging for time-sensitive data center projects.

renewable energy certificate (REC)

A tradable certificate representing proof that 1 megawatt-hour of electricity was generated from a renewable source. RECs are separate from the physical electricity and can be bought and sold independently. A company can purchase RECs to claim renewable energy use even if their actual electricity comes from fossil fuels. This accounting mechanism enables corporate sustainability claims but has been criticized for not driving new renewable development or reducing actual emissions.

right-of-way

A legal easement granting utilities the right to build and maintain transmission infrastructure across private property. Right-of-way corridors for high-voltage transmission lines typically span 50 to 600 feet wide and must be cleared of all trees and structures. Property owners retain title but lose most use rights within the corridor. Compensation is typically a one-time payment based on land value, not the value of the infrastructure built.

RTO (Regional Transmission Organization)

An independent organization that manages the high-voltage transmission grid and operates wholesale electricity markets across multiple states. RTOs ensure reliability, coordinate generator dispatch, and manage the interconnection of new power plants and large loads. PJM, MISO, and SPP are major RTOs; ERCOT serves Texas but operates slightly differently.

special use permit

A discretionary approval allowing a specific land use that requires additional review and conditions beyond what is permitted by-right in a zoning district. Unlike variances (which modify dimensional requirements), special use permits allow specific uses deemed compatible with a zone under certain conditions. Data centers often require special use permits in agricultural or residential zones, giving communities significant leverage to impose conditions.

substation

A facility that transforms voltage levels between transmission and distribution systems, or between different transmission voltages. Substations contain transformers, switches, and protective equipment. Data centers requiring 100+ MW typically need dedicated substations connected directly to high-voltage transmission lines, costing $50-200+ million and taking 3-5 years to permit and build.

throughput

The rate at which an AI system can process work, typically measured in tokens per second. Higher throughput enables serving more users simultaneously. Data center design increasingly optimizes for throughput, as inference workloads scale with user demand rather than being bounded like training.

token

/TOH-kuhn/

A numerical representation of a text fragment used as the fundamental unit of AI processing. Tokens are typically 3-4 characters on average, with approximately 1,300 tokens representing 1,000 words of English text. When you interact with an AI system like ChatGPT, your text is first converted into tokens before processing.

TPU (Tensor Processing Unit)

Google's custom-designed AI accelerator chip, purpose-built for machine learning workloads. TPUs represent an alternative to NVIDIA's GPU dominance, though they're primarily available through Google Cloud rather than for general purchase.

training

The process of adjusting an AI model's parameters by exposing it to large amounts of data, allowing it to learn patterns and relationships. Training a frontier model like GPT-4 can cost $100+ million and take months, but only happens once. The resulting model weights are then used for inference.

transformer

The dominant neural network architecture for modern AI systems, introduced in 2017. Transformers use attention mechanisms to process input sequences in parallel, enabling them to understand relationships between words regardless of their position in text. GPT, Claude, and most large language models are built on transformer architecture.

transmission

High-voltage power lines (typically 115kV to 765kV) that carry electricity long distances from power plants to population centers. Transmission infrastructure is the backbone of the grid. Data centers at the gigawatt scale must connect directly to transmission lines because lower-voltage distribution networks can't handle the load. Building new transmission lines takes 5-10+ years.

UPS (Uninterruptible Power Supply)

Battery backup systems that provide immediate power if the main electrical supply fails. UPS systems bridge the gap (typically 10-30 seconds) between power loss and backup generator startup. A hyperscale facility might have thousands of UPS units providing megawatts of backup capacity.

watt / kilowatt / megawatt / gigawatt

Units of power representing capacity to do work. Each step represents a 1,000x increase: Watt (W) – A single light bulb. Kilowatt (kW) – A typical home's peak demand (5-10 kW). Megawatt (MW) – A small town or large building (1,000 kW). Gigawatt (GW) – A major power plant or the entire demand of a million homes (1,000 MW). Modern AI data centers operate at the gigawatt scale, requiring power equivalent to major cities.

zoning variance

A formal exception to standard zoning rules granted by local authorities, allowing a use not normally permitted in a zone. Data centers in agricultural areas often require variances or conditional use permits. The variance process provides opportunities for public input but also creates uncertainty for developers about project timelines and approval.