Data center readiness checklist
A site readiness checklist for AI data centers: power and interconnection, pad and structural loads, network, water, permitting, logistics, and security. Open the readiness checklist.
R-01Resources
AI infrastructure glossary
This glossary defines 40 terms used across AI data-center engineering — power, cooling, compute, and grid interconnection — in two to four plain sentences each. It is written for people evaluating megawatt-scale AI infrastructure, not for people selling it.
- 40
- Terms defined
- 11
- Cited sources
PUBLISHED LAST VERIFIED BY JOSEF ELIMELECHREVIEWED PODOS AI ENGINEERING
How to read this glossary
01
Four vocabularies, one building
Power, cooling, compute, and grid interconnection — the four languages that meet inside an AI data center, defined in the terms they are actually used in.
02
Sourced inline
External figures cite the source register at the point of use: Uptime Institute, LBNL, the IEA, ASHRAE, NREL, the Open Compute Project, and vendor documentation.
03
Company figures marked
Where PODOS figures appear they are design targets, not measured results from operating deployments.
Resources
Resources and next steps
Every entry below sits inside a larger decision. These pages carry the decisions.
Engineering overview
How the cooling, power, and enclosure decisions fit together. Engineering overview.
Platform overview
What a standardized modular AI data-center unit is. Platform overview.
Modular vs traditional data centers
The decision framework behind the vocabulary. Modular vs traditional data centers.
Metrics
Efficiency metrics, side by side
All three share a denominator — IT energy — but answer different questions. Compare them only with measurement boundary and season stated.
| Metric | Formula | What it tells you | Published reference points | Common distortions |
|---|---|---|---|---|
| PUE | Total facility energy ÷ IT energy | Overhead spent on cooling, conversion, and building systems | Industry average roughly flat for ~6 years[1]; Google fleet TTM 1.09[2] | Snapshot vs annualized figures; partially loaded facilities read worse |
| WUE | Site water (L) ÷ IT energy (kWh) | Water consumed on site, mostly by evaporative cooling | Microsoft design WUE 0.30 L/kWh, down from 0.49 in 2021[3] | Excludes water embedded in off-site electricity generation |
| ERE | (Total energy − reused energy) ÷ IT energy | Efficiency after crediting heat exported to a real customer | Can fall below 1.0 with heat export; NREL’s ESIF pairs heat reuse with PUE ~1.04[8] | Meaningless without an actual off-taker for the heat |
A PUE or WUE quoted without methodology, load level, and averaging period is not comparable to one quoted with them.
Index
Term index — 40 entries
A to Z
Terms, A to Z
- 800 VDC power architecture
- A rack power-distribution approach that replaces multiple AC conversion stages with a single high-voltage direct-current bus feeding the rack. It is promoted for multi-hundred-kilowatt AI racks to cut conversion losses and copper mass; standards are still forming.
- ASHRAE thermal guidelines
- The de facto environmental envelope for IT equipment. The fifth edition defines air-cooled classes A1–A4 — A4 permitting inlet air up to 40 °C — plus liquid-cooling classes named by maximum facility water temperature.[6]
- Battery energy storage system (BESS)
- Utility-scale batteries co-located with a facility to ride through disturbances, shave peaks, or sell flexibility to the grid. An NREL demonstration dispatched a 35 MW battery at a 70 MW grid-interactive data center in under five seconds without breaking service commitments.[11]
- Behind-the-meter (BTM)
- Generation or storage connected on the customer side of the utility meter, serving local load without waiting on grid-connection studies. BTM power is one route around long interconnection queues; the deployment overview covers how it pairs with modular capacity.
- CDU (coolant distribution unit)
- A pump-and-heat-exchanger skid that isolates the facility water loop from the technology cooling loop serving cold plates. It regulates flow, pressure, and supply temperature — typically holding coolant above dew point so condensation never forms on hardware.
- Closed-loop liquid cooling
- A cooling architecture in which a fixed charge of coolant circulates continuously between heat sources and heat rejection, with no evaporation and no continuous make-up water. Closed loops are what make zero-water operation and practical heat recovery possible.
- Cold plate
- A metal plate with internal fluid channels mounted directly on a processor package, moving heat into liquid at the die instead of into room air. The Open Compute Project publishes cold-plate requirements so plates, manifolds, and CDUs from different vendors interoperate.[9]
- Colocation
- Renting space, power, and cooling for your own hardware inside a shared third-party facility. Colocation transfers the facility problem to a landlord; available density and liquid-cooling readiness vary widely between sites.
- Demand response
- Utility or market programs that compensate large electricity consumers for reducing or shifting load during grid stress. A flexible data center can treat demand response as a revenue stream rather than an interruption.
- Direct-to-chip (D2C) liquid cooling
- A liquid-cooling method that pipes coolant through cold plates mounted on GPUs and CPUs, removing most heat at the silicon. Loop architecture and trade-offs are covered in the direct-to-chip liquid cooling explainer.
- Dry cooler
- A closed-circuit, air-to-liquid heat exchanger that rejects loop heat to ambient air without evaporating water. Dry coolers enable zero-water heat rejection at the cost of efficiency in peak heat.
- Edge data center
- A small facility placed near where data is generated or consumed, cutting latency and backhaul cost. AI inference is pushing edge sites from closet scale toward megawatt-class units; see the designed-for use cases.
- Energy reuse effectiveness (ERE)
- (Total facility energy − energy reused elsewhere) ÷ IT energy. Unlike PUE, ERE credits heat exported to buildings or processes, so it can fall below 1.0 for a facility with a real heat customer.
- Evaporative cooling
- Heat rejection that cools air or water by evaporating water, trading water consumption for electrical efficiency. It is the main driver of data-center water use and of WUE differences between designs.
- Facility water system (FWS)
- The building-side hydronic loop carrying heat from CDUs out to heat rejection — dry coolers, cooling towers, or a heat-reuse customer. Its supply temperature largely determines which rejection options are viable.
- GPU cluster
- A set of GPU servers linked by high-bandwidth fabric so thousands of chips behave as one machine. Density is rising fast: NVIDIA’s GB200 NVL72 places 72 GPUs and 36 CPUs in a single liquid-cooled rack acting as one NVLink domain.[7]
- Heat reuse (waste-heat recovery)
- Capturing heat a facility would otherwise reject and delivering it to a productive use — district heating, offices, industrial processes. NREL’s ESIF facility heats its office space with HPC waste heat and reports a PUE near 1.04.[8]
- Hyperscale data center
- A facility class operated by the largest cloud and AI platforms, typically tens to hundreds of megawatts with custom hardware and leading efficiency. Google reports a fleet-wide trailing-twelve-month PUE of 1.09 per its latest reporting.[2]
- Immersion cooling
- Liquid cooling that submerges entire servers in a non-conductive dielectric fluid, in single-phase or two-phase variants. It removes nearly all heat to liquid but changes serviceability and hardware qualification.
- Inference
- Running a trained model to produce output. Inference is latency-sensitive and often memory-bound, and over a model’s life it usually consumes more total compute than the training run did.
- Interconnection queue
- The ordered backlog of projects waiting on utility studies and network upgrades before they may connect to the grid. Queue timelines run to years in many regions, making grid access a defining constraint on where new AI capacity can be built.
- IT load
- The electrical power consumed by computing, storage, and network equipment alone — excluding cooling, conversion losses, and building systems. IT load is the denominator in PUE, WUE, and ERE.
- KV cache
- In transformer inference, the per-token key and value tensors held in GPU memory so earlier context is not recomputed. KV caches grow with context length and batch size; published research shows quantization to roughly 3.5 bits per channel can be quality-neutral.[10] The Syntropic overview covers the software side of memory efficiency.
- Medium-voltage (MV) switchgear
- Switching, protection, and metering equipment operating at medium voltage — roughly 1 kV to 35 kV, with 13.8 kV a common North American distribution class. How a site accepts and steps down MV power is covered in data-center power architecture.
- Microgrid
- A local energy system — generation, storage, and controls — that can run connected to the utility grid or islanded from it. Microgrids let compute operate where grid service is weak, delayed, or absent.
- Modular data center
- Data-center capacity produced as a factory-built, transportable unit — structure, power, cooling, and racks integrated and tested before shipment — rather than constructed in place. PODOS applies the model at fixed scale: each PODOS Pod is designed as a standardized 1 MW building block. For the full decision framework, see modular vs traditional data centers.
- N+1 redundancy
- A redundancy scheme providing one spare unit beyond the number required to carry the load — five pumps where four suffice. Contrast 2N, which duplicates the entire system, and N, which carries no spare at all.
- Open Compute Project (OCP)
- An industry body that open-sources data-center hardware and cooling specifications, including cold-plate, immersion, and CDU requirements. Its cooling-environments work is what keeps the liquid-cooling ecosystem multi-vendor.[9]
- Organic Rankine cycle (ORC)
- A heat engine that generates electricity from low-temperature heat by boiling an organic working fluid instead of water. ORC systems are one route to converting data-center waste heat into usable power rather than venting it.
- Power distribution unit (PDU)
- Equipment that takes conditioned facility power and distributes it to racks. PDUs are usually where circuit-level power monitoring lives.
- Power usage effectiveness (PUE)
- Total facility energy ÷ IT energy — the industry’s standard overhead metric, where 1.0 would mean zero overhead. Industry-average PUE has stayed roughly flat for about six years in Uptime Institute’s surveys,[1] while leading hyperscale fleets report figures near 1.09.[2]
- Rack density
- The power drawn by a single rack, in kW. Uptime’s 2025 survey shows densities climbing into the 10–30 kW band for many operators;[1] current AI racks sit far above that, which is why liquid cooling is displacing air.
- Rear-door heat exchanger (RDHx)
- A liquid-cooled coil that replaces a rack’s rear door, absorbing server exhaust heat into water before it enters the room. A retrofit-friendly middle step between air cooling and direct-to-chip.
- Substation
- The transformers and switchgear that step transmission or distribution voltage down for site use. For large campuses a new substation is often the longest-lead physical asset on the project.
- Technology cooling system (TCS)
- The server-side loop between the CDU and the cold plates, running treated coolant at controlled temperature and pressure — kept separate from the facility loop to protect hardware from water-quality excursions.
- Time-to-power
- The elapsed time from committing to a site until it can draw its contracted power — driven by interconnection studies, substation work, and utility construction. PODOS targets a 90-day window from order to commissioning for a standard unit, a target that presumes power is already available at the site; see how deployment is sequenced.
- Training
- Building or updating a model’s weights by processing large datasets across many synchronized GPUs. Training is throughput-bound rather than latency-bound, and it stresses interconnect bandwidth and sustained power draw.
- Uninterruptible power supply (UPS)
- Batteries or flywheels plus power electronics that carry the IT load for the seconds-to-minutes between a grid disturbance and generator pickup. UPS topology is a core decision in power architecture.
- Water usage effectiveness (WUE)
- Liters of water consumed on site per kWh of IT energy. Microsoft reports a design WUE of 0.30 L/kWh for new builds, down from 0.49 in 2021;[3] closed-loop designs aim for zero operational water consumption.
- Zero-water cooling
- Heat rejection designed to consume no water in operation — typically a closed coolant loop rejecting heat through dry coolers. It removes water permits and drought exposure from siting decisions.
Honest limits
What this glossary does not settle
- Definitions reflect common industry usage; vendors attach narrower or looser meanings to terms like “pod” and “AI-ready.” Contracts should define terms explicitly.
- Every metric here depends on measurement boundary and season. A PUE or WUE quoted without methodology, load level, and averaging period is not comparable to one quoted with them.
- ASHRAE, NFPA, and IEEE standards are paywalled; these entries summarize scope and do not substitute for the current editions.
- Where PODOS figures appear they are design targets, not measured results from operating deployments. Third-party numbers carry their source and as-of year inline.
- The vocabulary is moving — 800 VDC distribution, for one, is still being standardized. These definitions were last verified on 2026-08-31.
Sources
- [1] Global Data Center Survey 2025 — Uptime Institute, Jul 2025
- [2] Data center efficiency (fleet trailing-twelve-month PUE) — Google, updated annually
- [3] Measuring energy and water efficiency for Microsoft datacenters — Microsoft, ongoing
- [4] 2024 United States Data Center Energy Usage Report (LBNL-2001637) — Lawrence Berkeley National Laboratory, Dec 2024
- [5] Energy and AI — Executive Summary — IEA, Apr 2025
- [6] Thermal Guidelines for Data Processing Environments, 5th ed. (TC 9.9) — ASHRAE, 2021
- [7] GB200 NVL72 product page — NVIDIA, spec page
- [8] HPC Data Center Waste Heat Reuse (ESIF) — NREL (DOE), ongoing
- [9] ACS Liquid Cooling Cold Plate Requirements, Rev 1.0 — Open Compute Project
- [10] TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate — arXiv (2504.19874), Apr 2025
- [11] Demonstrating the Data Center as a Flexible Grid Asset — NREL (DOE), FY2025
Know the terms. Now check the site
PODOS AI builds factory-built modular AI compute infrastructure.