50 terms that come up when sourcing colocation, AI-ready and GPU-ready capacity, defined plainly.
Efficiency · Power · Cooling · AI Hardware · Networking · Reliability · AI Infrastructure · Compliance · Sustainability
Efficiency
- Power Usage Effectiveness (PUE)
PUE measures data center energy efficiency as the ratio of total facility power to IT equipment power. Calculated as PUE = Total Facility Power / IT Equipment Power. A PUE of 1.0 represents perfect efficiency (impossible in practice). Industry average is 1.6, while leading facilities achieve 1.2 to 1.3. Lower PUE indicates better efficiency and lower operating costs.
Related: DCiE, Carbon Neutral, Renewable Energy
- Data Center Infrastructure Efficiency (DCiE)
DCiE is the inverse of PUE, expressed as a percentage. DCiE = (IT Equipment Power / Total Facility Power) x 100. A DCiE of 100% would represent perfect efficiency. This metric helps data center operators understand what percentage of power actually reaches computing equipment.
Related: PUE
Power
- Power Density
Power density measures the amount of electrical power consumed per unit of floor space, typically expressed in kilowatts per rack (kW/rack). Traditional enterprise workloads require 5 to 10kW per rack, while AI training workloads may require 30 to 100kW+ per rack. Higher power density enables more compute capacity in less physical space.
Related: kW per Rack, High-Density Computing
- kW per Rack
Kilowatts per rack is the standard unit for measuring power density in data centers. It indicates the maximum electrical power that can be delivered to a single equipment rack. AI-ready facilities typically support 20 to 50kW per rack, while cutting-edge GPU deployments may require 80 to 100kW+ per rack.
Related: Power Density, GPU Cluster
Cooling
- Liquid Cooling
Advanced cooling technology using liquid coolant (water, dielectric fluid, or refrigerant) to remove heat directly from IT equipment. Essential for AI workloads exceeding 30kW per rack density where traditional air cooling becomes inefficient. Types include direct-to-chip cooling, rear-door heat exchangers, and immersion cooling.
Related: Immersion Cooling, Direct-to-Chip Cooling, Air Cooling
- Immersion Cooling
A liquid cooling method where IT equipment is submerged in a non-conductive dielectric fluid. The fluid absorbs heat directly from all components, enabling extremely high power densities (100kW+ per tank). Particularly effective for GPU clusters and ASIC mining operations.
Related: Liquid Cooling, Dielectric Fluid
- Direct-to-Chip Cooling
A liquid cooling approach where coolant flows through cold plates mounted directly on processors (CPUs and GPUs). This targeted cooling is more efficient than air cooling and enables power densities of 50 to 80kW per rack while maintaining standard server form factors.
Related: Liquid Cooling, Cold Plate
- Hot Aisle / Cold Aisle
A data center layout strategy that separates hot exhaust air from cold supply air. Equipment racks are arranged so that cold air intakes face one aisle (cold aisle) and hot air exhausts face the adjacent aisle (hot aisle). This prevents hot and cold air mixing, improving cooling efficiency.
Related: Containment, CRAH, CRAC
- CRAC / CRAH Units
Computer Room Air Conditioning (CRAC) and Computer Room Air Handler (CRAH) units are precision cooling systems designed for data centers. CRAC units use compressor-based cooling, while CRAH units use chilled water. Both provide precise temperature and humidity control required for IT equipment.
Related: Hot Aisle / Cold Aisle, Air Cooling
AI Hardware
- GPU (Graphics Processing Unit)
Specialized processors designed for parallel computing tasks, essential for AI training and inference. Unlike CPUs that excel at sequential tasks, GPUs contain thousands of cores optimized for matrix operations. Common data center GPUs include NVIDIA H100 (latest generation, 80GB HBM3), A100 (previous generation, 40/80GB HBM2e), and V100 (legacy but capable).
Related: NVIDIA H100, NVIDIA A100, TPU, GPU Cluster
- NVIDIA H100
NVIDIA's flagship data center GPU (Hopper architecture), released in 2022. Offers up to 80GB of HBM3 memory with 3.35TB/s bandwidth. Designed specifically for AI training and inference, the H100 delivers 3x the AI training performance of A100. Requires liquid cooling for optimal performance at high utilization.
Related: GPU, NVIDIA A100, Tensor Core
- NVIDIA A100
NVIDIA's previous-generation data center GPU (Ampere architecture). Available in 40GB and 80GB HBM2e configurations. Still widely deployed for AI training and inference workloads. Often more readily available than H100s and suitable for many AI applications.
Related: GPU, NVIDIA H100, NVIDIA V100
- TPU (Tensor Processing Unit)
Google's custom-designed AI accelerator, optimized specifically for TensorFlow workloads. TPUs are available through Google Cloud and are particularly efficient for large-scale AI training. TPU v4 pods can scale to thousands of chips for massive distributed training.
Related: GPU, AI Accelerator
- GPU Cluster
A collection of interconnected GPU servers designed to work together on distributed computing tasks. AI training clusters may contain hundreds to thousands of GPUs connected via high-speed networking (InfiniBand or NVLink). Clusters require significant power (multiple MW) and cooling infrastructure.
Related: GPU, InfiniBand, NVLink, Distributed Training
- HBM (High Bandwidth Memory)
A high-performance memory technology used in modern GPUs and AI accelerators. HBM stacks memory chips vertically and connects them via silicon interposers, achieving much higher bandwidth than traditional GDDR memory. HBM3 (used in H100) delivers up to 3.35TB/s bandwidth.
Related: GPU, NVIDIA H100
Networking
- InfiniBand
A high-performance networking technology commonly used in AI clusters and supercomputers. InfiniBand provides extremely low latency (under 1 microsecond) and high bandwidth (up to 400Gb/s with NDR). Essential for distributed AI training where GPUs across multiple servers must communicate frequently.
Related: GPU Cluster, RDMA, NVLink
- NVLink
NVIDIA's proprietary high-speed interconnect for connecting GPUs within a server. NVLink provides 900GB/s bandwidth between GPUs in H100 systems, enabling them to share memory and work as a unified compute resource. Different from InfiniBand which connects servers.
Related: InfiniBand, GPU, NVSwitch
- Cloud On-Ramp
A direct, private network connection between a data center and a cloud provider's network. Cloud on-ramps bypass the public internet, providing lower latency, higher bandwidth, and improved security. Common examples include AWS Direct Connect, Azure ExpressRoute, and Google Cloud Interconnect.
Related: Cross-Connect, Carrier Neutral
- Cross-Connect
A direct physical cable connection between two parties within a data center. Cross-connects enable private, low-latency connectivity to cloud providers, carriers, and other tenants without traversing external networks. Essential for building hybrid cloud architectures.
Related: Cloud On-Ramp, Meet-Me Room
- Meet-Me Room (MMR)
A designated space within a data center where telecommunications carriers and tenants can interconnect. The MMR contains demarcation points for carrier circuits and enables cross-connects between different parties. A well-connected MMR indicates a carrier-neutral facility.
Related: Cross-Connect, Carrier Neutral
- Carrier Neutral
A data center that allows multiple telecommunications carriers to provide services within the facility. Carrier-neutral facilities give customers choice in connectivity providers and typically offer better pricing through competition. Essential for enterprises requiring diverse network paths.
Related: Meet-Me Room, Cross-Connect
Reliability
- Tier Level (Uptime Institute)
A standardized classification system for data center reliability developed by the Uptime Institute. Tier I (99.671% uptime) provides basic infrastructure. Tier II (99.741%) adds redundant components. Tier III (99.982%) allows concurrent maintainability. Tier IV (99.995%) is fault tolerant with 2N+1 redundancy.
Related: N+1 Redundancy, 2N Redundancy, Uptime
- N+1 Redundancy
A redundancy configuration where one additional component backs up N primary components. For example, if a facility needs 4 generators to meet load (N=4), N+1 means having 5 generators. This allows one unit to fail or undergo maintenance without affecting operations.
Related: 2N Redundancy, Tier Level
- 2N Redundancy
A redundancy configuration with fully duplicated infrastructure. If N components are needed to meet load, 2N provides a complete second set. This is the highest level of redundancy, found in Tier IV facilities, ensuring no single point of failure.
Related: N+1 Redundancy, Tier Level, Fault Tolerant
- UPS (Uninterruptible Power Supply)
Battery backup systems that provide instantaneous power during utility outages, protecting IT equipment from power interruptions. UPS systems bridge the gap between utility failure and generator startup (typically 10 to 15 seconds). Modern data centers use rotary or static UPS configurations.
Related: Generator, Power Redundancy
AI Infrastructure
- AI Readiness Score
A composite metric evaluating a data center's capability to support AI workloads. Factors include power density, cooling type, GPU infrastructure, network connectivity, and cloud latency. Scores typically range from 1 to 100, with 75+ indicating full AI readiness.
Related: Power Density, Liquid Cooling, GPU
- AI Training
The computationally intensive process of teaching machine learning models by processing large datasets. AI training requires massive GPU clusters, high-speed networking, and can consume megawatts of power. Training runs may last days to weeks and require specialized cooling.
Related: AI Inference, GPU Cluster, Distributed Training
- AI Inference
The process of using trained AI models to make predictions on new data. Inference is less compute-intensive than training but often requires low latency and high throughput. Edge inference deployments bring AI closer to end users for real-time applications.
Related: AI Training, Edge Computing, Latency
- Distributed Training
A technique for training AI models across multiple GPUs or servers simultaneously. Methods include data parallelism (splitting data across GPUs) and model parallelism (splitting the model). Requires high-speed interconnects like InfiniBand for efficient gradient synchronization.
Related: GPU Cluster, InfiniBand, AI Training
Compliance
- SOC 2
A security framework developed by AICPA that evaluates an organization's controls related to security, availability, processing integrity, confidentiality, and privacy. SOC 2 Type II certification requires demonstrating controls are effective over time (typically 6 to 12 months).
Related: ISO 27001, HIPAA, PCI-DSS
- ISO 27001
An international standard for information security management systems (ISMS). ISO 27001 certification demonstrates that an organization has implemented comprehensive security controls and follows security best practices. Widely recognized globally.
Related: SOC 2, Security
- HIPAA
The Health Insurance Portability and Accountability Act sets standards for protecting sensitive patient health information. Data centers hosting healthcare workloads must implement specific physical, technical, and administrative safeguards to achieve HIPAA compliance.
Related: SOC 2, Healthcare, PHI
- PCI-DSS
Payment Card Industry Data Security Standard defines requirements for organizations handling credit card data. PCI-DSS compliant data centers implement specific controls for network segmentation, encryption, access control, and monitoring.
Related: SOC 2, Security, Payment Processing
- GDPR (EU General Data Protection Regulation)
The EU's omnibus data protection law (Regulation 2016/679), enforced since May 2018. Governs how personal data of EU/EEA residents is processed, stored, and transferred. Requires lawful basis for processing, data subject rights, breach notification within 72 hours, and DPIAs for high-risk processing. Data centers hosting EU personal data must support data residency, sub-processor transparency, and Standard Contractual Clauses for international transfers.
Related: UK GDPR, NIS2, DORA, ISO 27701
- UK GDPR & Data Protection Act 2018
The United Kingdom's post-Brexit data protection regime, mirroring EU GDPR with UK-specific amendments. Supervised by the Information Commissioner's Office (ICO). UK buyers in financial services and public sector also expect alignment with Cyber Essentials Plus and FCA operational resilience requirements.
Related: GDPR, ISO 27001
- NIS2 Directive
The EU's Network and Information Security Directive 2 (Directive 2022/2555), in force from October 2024. Expands cybersecurity obligations to a wider set of essential and important entities, including data center service providers and cloud computing services. Requires risk management, incident reporting within 24/72 hours, and executive accountability.
Related: GDPR, DORA
- DORA (Digital Operational Resilience Act)
EU regulation (2022/2554) effective January 2025, establishing operational resilience requirements for financial entities and their critical ICT third-party providers, including data centers. Mandates ICT risk management, incident reporting, resilience testing, and contractual oversight of third-party providers.
Related: NIS2 Directive, GDPR
- EN 50600
The European standard for data center facilities and infrastructures, covering design, construction, and operation. Defines availability, security, and energy efficiency classes (similar to but distinct from Uptime Institute Tiers). Widely referenced by EU enterprise and public-sector buyers.
Related: Tier Level (Uptime Institute), ISO 27001
- IRAP (Australia)
The Information Security Registered Assessors Program, run by the Australian Signals Directorate (ASD). IRAP-assessed data centers and cloud services can host Australian government workloads, typically certified to PROTECTED level. Required for federal and many state agency deployments in Australia.
Related: ISO 27001, SOC 2
- PDPA (Singapore)
Singapore's Personal Data Protection Act 2012, enforced by the Personal Data Protection Commission (PDPC). Governs the collection, use, and disclosure of personal data and includes the Do Not Call registry. Singapore data centers serving regulated workloads also align with MAS Technology Risk Management (TRM) guidelines.
Related: APPI (Japan), PIPA (South Korea), DPDP Act (India)
- APPI (Japan)
Japan's Act on the Protection of Personal Information, supervised by the Personal Information Protection Commission (PPC). Governs handling of personal data of Japanese residents, including cross-border transfer restrictions. Financial workloads additionally require alignment with FSA-J guidelines.
Related: PDPA (Singapore), PIPA (South Korea)
- PIPA (South Korea)
South Korea's Personal Information Protection Act, one of the strictest data privacy regimes in APAC. Requires explicit consent for collection and tight controls on cross-border transfers. Korean enterprise buyers commonly require domestic data residency.
Related: APPI (Japan), PDPA (Singapore)
- DPDP Act (India)
India's Digital Personal Data Protection Act 2023, the country's first comprehensive data protection law. Establishes consent, purpose limitation, and breach notification requirements, with cross-border transfer rules set by government notification. Drives growing demand for in-country data center capacity.
Related: PDPA (Singapore), GDPR
- Australian Privacy Act (APP)
Australia's Privacy Act 1988, with the 13 Australian Privacy Principles (APPs) governing how personal information is collected, used, and disclosed. Enforced by the Office of the Australian Information Commissioner (OAIC). Financial services workloads additionally require APRA CPS 234 alignment.
Related: IRAP (Australia), GDPR
- LGPD (Brazil)
Brazil's Lei Geral de Protecao de Dados (Law 13.709/2018), enforced by the Autoridade Nacional de Protecao de Dados (ANPD). Modeled closely on EU GDPR, governing personal data of Brazilian residents and applicable to processors located outside Brazil that target the Brazilian market.
Related: GDPR, POPIA (South Africa)
- POPIA (South Africa)
South Africa's Protection of Personal Information Act, in full force since July 2021. Enforced by the Information Regulator. Sets conditions for lawful processing of personal information and underpins growing data residency requirements across Sub-Saharan Africa.
Related: GDPR, LGPD (Brazil)
- PIPEDA (Canada)
Canada's Personal Information Protection and Electronic Documents Act, governing private-sector handling of personal data. Several provinces (Quebec under Law 25, BC, Alberta) have substantially similar regimes. Federal financial workloads also expect OSFI B-13 alignment.
Related: GDPR, HIPAA
Sustainability
- Carbon Neutral
A designation indicating that a facility's net carbon emissions are zero, achieved through a combination of renewable energy, efficiency improvements, and carbon offsets. Carbon neutral data centers help customers meet corporate sustainability goals and reduce scope 3 emissions.
Related: PUE, Renewable Energy, RECs
- RECs (Renewable Energy Certificates)
Tradeable certificates representing the environmental benefits of 1 MWh of renewable energy generation. Data centers purchase RECs to claim renewable energy usage even when directly consuming grid power. RECs enable sustainability claims without on-site generation.
Related: Carbon Neutral, Renewable Energy, PPA
- PPA (Power Purchase Agreement)
A long-term contract to purchase electricity directly from a renewable energy generator at a fixed price. PPAs enable data centers to claim 100% renewable energy and provide price certainty. More impactful than RECs as they directly fund new renewable capacity.
Related: RECs, Renewable Energy, Carbon Neutral