Skip to content
7Hive

Reference

AI data center glossary

Shared language for the people on both sides of a renewable-to-compute table — power owners, hyperscalers, operators, and capital. Terms are defined as they show up in a transaction, not as a textbook. See also the diligence checklist and markets desk.

Dominant players

Operators, owners, grid bodies, and vendors that commonly appear in AI data-center offtake, power, colo, silicon, and interconnect. Each name links to the organization’s site.

Hyperscale

Neocloud

Colocation

Silicon

Power

Grid

Cooling & electrical

EPC

Standards

0–9

Term In the transaction
2N / N+1 Redundancy models. N+1 means one spare of each critical system; 2N means a fully duplicated independent path. AI halls often mix both — 2N on power, N+1 on cooling — because GPU clusters cannot ride through a long outage.
415 / 480 V Common facility distribution voltages. North American halls typically land at 480 V; many high-density AI designs also use 415 V or higher rack-level DC to cut conversion loss as rack power climbs past 50–100 kW.
8760 Hours in a non-leap year. “8760 power” means the site is expected to run continuously. Training usually requires that; inference and mining can sometimes take a shaped or interruptible product, which should be contracted as such.

A

Term In the transaction
Air-side economizer Using outside air to cool the hall when climate allows. Cheap at moderate densities; usually insufficient alone for liquid-cooled GPU clusters, which dump heat into a liquid loop instead of the room air.
Ampacity How much current a feeder, bus, or transformer can carry without overheating. In AI deals, ampacity at the substation and the service entrance often caps campus MW before land or GPUs do.
Ancillary services Grid products beyond energy — frequency response, reserves, voltage support. A large AI campus can buy these, sell them (with batteries), or be penalized for not providing them, depending on the ISO tariff.
Anchor tenant The first creditworthy offtaker that makes a campus financeable. Hyperscalers, neoclouds, and well-capitalized training operators are typical anchors. Mining can be a real offtake, but it is a different credit, SLA, and community product — it should be named, not implied.
ASIC Application-specific chip (TPU, Trainium, custom inference silicon). Changes rack density, cooling, and power shape versus a GPU-only hall, which is why “AI data center” is not one product.
Availability Share of time the contracted power or hall is actually up. SLAs are written in nines (99.9% ≈ 8.8 hours down/year). GPU training is highly sensitive to unplanned interruption, so the power PPA and the hall SLA need to match.

B

Term In the transaction
Baseload Power that is on around the clock. Training wants baseload. Wind and solar are not baseload unless paired with storage, hydro, firm gas, or a hybrid BTM+FTM interconnect that can fill the gaps.
Basis risk Price difference between where power is generated and where it is delivered (nodes, hubs, congestion). A cheap renewable PPA can still lose money if congestion sits between the plant and the AI campus.
BESS Battery energy storage system On-site batteries. Used to ride through ramps, shave peaks, firm solar/wind, and sometimes sell ancillary services. For AI, batteries are usually a bridge — not a substitute for firm interconnect.
Behind-the-meter BTM The compute load takes power from on-site or adjacent generation before it hits the bulk grid. Can lower delivered cost and ease some transmission constraints. Rarely financeable alone for 24/7 AI without a firm FTM backup.
Black start Ability to restore power without the grid. Relevant for islanded or highly BTM campuses. Most AI sites still assume the utility can re-energize; true black start is a premium, negotiated capability.
Brownfield Reusing an existing industrial, generation, or data-center site. Often faster than greenfield because interconnect, water, and zoning already exist — the usual reason stranded or underused power plants show up in AI deal flow.
Build-to-suit A hall or campus designed and built for one tenant’s density, cooling, and security spec. Standard for hyperscale AI; less common in multi-tenant colo unless the tenant underwrites the delta.
Busbar The metal conductor that collects and distributes power inside a substation or switchgear lineup. “Busbar PPA” or “busbar delivery” means energy is delivered at the generator, not at the data-center meter — the buyer then takes basis and transmission risk.

C

Term In the transaction
Campus A multi-building AI site sharing substations, water, fiber, and often a phased MW build. Deals are usually optioned in phases (e.g. 100 MW then 300 MW) so interconnect and capital can catch up to GPU supply.
Capacity factor Actual energy produced divided by nameplate over time. A 200 MW solar plant at 25% capacity factor is not 200 MW of AI power. Matching capacity factor and shape to workload is the core of renewable-to-compute underwriting.
Capex Upfront capital — land, substation, hall, cooling, transformers, GPUs. AI data-center deals often split power-side capex from IT capex. Origination and close are commonly paid as a mix of promote and success fee on the power and site work.
Carbon-free energy CFE Energy matched to carbon-free generation, sometimes hourly (24/7 CFE). Hyperscalers increasingly require this in RFPs. RECs alone are no longer enough for many AI offtakers.
COD Commercial operation date The date the plant, interconnect, or hall is deemed operational under the contract. Missed CODs trigger delay damages, GPU idle cost, and sometimes termination. Align generation COD, interconnect COD, and hall COD or the deal breaks.
Colocation Colo A third party owns the hall; the tenant brings (or leases) the IT. AI colo is denser and more power-constrained than enterprise colo. Many “AI colo” deals are closer to powered-shell or build-to-suit than classic retail cabinets.
Commissioning Cx Testing that power, cooling, BMS, and failover actually work at design load. AI halls should be commissioned at the kW/rack they will run, not at a legacy 8 kW test — otherwise first GPU turn-up is the real commissioning.
Curtailment Forced reduction of generation or load. Renewables get curtailed when the grid is full; AI loads can be asked to drop during emergencies. Contracts must say who eats the lost MWh and whether the workload can actually shed.
Cluster A tightly networked group of GPUs/ASICs that train or serve as one machine. Cluster size drives hall design: east-west fabric, liquid cooling, and a power block that can be energized as a unit.

D

Term In the transaction
Data hall The white-space room that holds racks. AI halls are planned around cooling loops and busway, not just raised floor. MW per hall is now a primary spec, not cabinets.
Demand charge A utility charge based on peak kW, not just kWh. A training ramp that spikes the meter can dominate the bill. Rate design and BTM generation are how deals manage this.
Demand response Being paid (or required) to drop load when the grid is tight. Compatible with some inference and mining; usually incompatible with large training runs unless the cluster is designed to checkpoint and yield.
Direct liquid cooling DLC Coolant to the cold plate on the GPU, not just to the room. Required at high rack densities. Adds CDU plant, fluid chemistry, leak detection, and a different EPC package than air-cooled colo.
DSCR Debt service coverage ratio Cash flow available for debt divided by debt service. Lenders underwrite AI campuses on contracted offtake, not on GPU residual value. Thin counterparty credit or merchant power typically compresses DSCR.
Duck curve Net-load shape in high-solar regions: inexpensive midday, steep evening ramp. An AI campus on solar-only BTM is long at noon and short in the evening. Hybrid FTM interconnect or storage is how training load is typically firmed.
Due diligence The site, title, interconnect, water, permitting, environmental, and counterparty review before a term sheet becomes a close. For AI, add fiber paths, transformer lead times, and whether the utility will actually energize the requested MW.

E

Term In the transaction
EPC Engineering, procurement, and construction The contractor wrapping design and build. AI campuses slip when power EPC, hall EPC, and liquid-cooling vendors are uncoordinated. Delivery oversight is a workstream of its own.
ERCOT The Texas grid operator. Energy-only market, fast load growth, relatively open interconnect compared with some eastern RTOs — and weather/scarcity price risk. A frequent region for large AI and other flexible compute load.
Exclusive option A paid right to control a site or MW block for a defined period while diligence runs. Standard in land and stranded-power deals so the buyer can diligence without losing the position.

F

Term In the transaction
FERC U.S. federal regulator of wholesale power and interstate transmission. Interconnect rules, wholesale market design, and many large-load disputes sit here or at the ISO under FERC-jurisdictional tariffs.
Fiber lateral The last-mile fiber into the campus. AI training clusters need diverse, high-count paths, not a single ISP drop. A power-perfect site with one fiber route is not a training site.
FID Final investment decision The board/lender moment when capex is committed. For AI campuses, FID usually waits on interconnect certainty, offtake, and long-lead equipment (transformers, switchgear, CDUs, sometimes turbines).
Firm power Energy the seller is obligated to deliver, not just when the resource blows or shines. Training offtakers ask for firm; as-available renewable is a different product and should be priced and contracted as such.
Floor loading Structural live load (psf). Liquid-cooled AI racks, CDUs, and busway are heavier than legacy IT. Brownfield conversions often stop here if the slab cannot take the density.
Front-of-the-meter FTM Interconnecting as a grid customer or generator on the utility/ISO side of the meter. Provides firm draw (and sometimes export) when on-site generation dips. Hybrid BTM+FTM is how many renewable-to-AI sites become financeable.

G

Term In the transaction
Generation shape When the MWh actually show up — hourly, seasonal. Hydro, wind, and solar have different shapes. The deal is matching that shape to inference, training, or a more flexible load, then filling residual with storage or FTM.
GPU The default AI accelerator. A rack of current-gen GPUs can draw tens to 100+ kW and wants liquid cooling, fat east-west fabric, and stable power. GPU delivery schedules now drive hall phasing as much as interconnect does.
Greenfield Building on undeveloped land. Full control of layout; full exposure to zoning, water, interconnect queue, and community process. Longer than brownfield unless the region is actively recruiting AI load.
Grid interconnection The studies, upgrades, and agreements that let a large load or generator touch the grid. For AI, the load interconnect is often the long pole — years in some ISO queues — which is why existing or near-ready positions trade at a premium.
GWh Gigawatt-hour. Energy, not capacity. A 100 MW hall running 8760 hours uses 876 GWh/year. PPAs, RECs, and carbon matching are GWh problems; substations and transformers are MW problems. Deals need both.

H

Term In the transaction
HBM High-bandwidth memory The memory stacked on AI accelerators. It is a chip constraint, not a site constraint — but HBM supply can slip GPU shipments, which slips hall energization and offtake start dates. Build the contract around that.
High-density rack Racks well above classic 5–10 kW — now commonly 40–120+ kW for AI. Changes cooling, busway, fire, structural, and how many MW fit in a hall. kW/rack belongs next to campus MW in any term sheet.
HVAC Air-side plant. Still needed for offices, electrical rooms, and residual heat, but no longer the primary cooling story for GPU halls. An AI campus should not be specified as if it were a legacy enterprise colo.
Hydro Water-driven generation. Often a strong renewable shape for AI: high capacity factor, dispatchable within water constraints, and frequently paired with existing substations. A core resource class, with wind and solar, in the 100 MW–3 GW renewable-to-compute range.
Hyperscale Operators who build and run at cloud scale (the large public clouds and their peers). They buy power in hundreds of MW, require CFE matching, and usually want control of design. Different counterparty than colo or neocloud.

I

Term In the transaction
Inference Running a trained model to produce answers. More geographically distributable and sometimes more interruptible than training. Can fit a wider set of power shapes — including some renewable-heavy BTM sites.
Interconnection agreement The contract with the utility or ISO that sets MW, voltage, upgrades, cost allocation, and energization date. For large AI load this is as important as the land PSA. Study results should be read before an LOI is treated as real.
Interconnection queue The backlog of generation and, increasingly, large-load requests waiting for studies. Queue position is an asset. A strong site with no position can still be years from energization.
IRR Internal rate of return The yield a sponsor underwrites. In these deals IRR is driven by offtake price, capex (especially substations and cooling), COD timing, and whether the structure is promote, fee, JV, or merchant residual.
ISO / RTO Independent system operator / regional transmission organization — PJM, MISO, CAISO, SPP, NYISO, ISO-NE, ERCOT (ERCOT is the Texas analogue). They run the wholesale market and the interconnect process. Region is a commercial term, not just a map.
IT load Power that reaches the IT equipment, versus facility load (cooling, UPS losses, lighting). PUE = total facility power / IT load. Interconnect and offtake should be explicit about which number the utility will meter.

J

Term In the transaction
Joint venture JV Shared ownership of the site, hall, or generation. Common when a power owner brings MW and a compute operator brings offtake and capex. Governance, promote, and exit rights are the commercial core.

K

Term In the transaction
kW / kWh Kilowatt is capacity (how large the connection is); kilowatt-hour is energy (what was used). Racks are specified in kW; bills and PPAs mix kW demand charges with kWh energy. Mixing them in a term sheet is a common source of restatement.
kVA Apparent power (volts × amps). Transformers and generators are rated in kVA/MVA, not kW. Power factor on GPU loads is usually good, but the nameplate that matters for interconnect is often MVA.
kW per rack The density number that decides cooling and electrical design. If this is unspecified, campus MW is incomplete. Parties typically ask for design kW/rack and max kW/rack, not an average that assumes empty white space.
KYB Know your business Counterparty diligence on the company — ownership, sanctions, financials, litigation. Institutional offtakers and power owners typically require it before sharing CIM-level site data. It is a standard onboarding step, not a slight.

L

Term In the transaction
Land option A contract to buy or lease land later at agreed terms, usually while interconnect and permitting run. Standard path from a target region to a controlled parcel without purchasing the land on day one.
LCOE Levelized cost of energy Lifetime cost per MWh of a generator. Useful for comparing hydro/wind/solar/gas. Incomplete for AI without delivered cost at the bus, shape, curtailment, and the cost of making it firm.
Liquid cooling Removing heat with fluid — rear-door, DLC, or immersion. Default for modern AI halls. Adds water or glycol plant, CDUs, and a different failure mode than CRAC units. Spec it at term sheet, not as a change order.
LMP Locational marginal price Nodal wholesale price: energy + congestion + losses. Two sites 30 miles apart can have very different LMPs. Siting an AI load at a cheap node (or behind a constrained generator) is a real source of deal value.
Load letter The formal request to the utility stating MW, voltage, ramp schedule, and in-service date. Starts the study clock. A phased, supportable ramp is what utilities and counterparties can take seriously; an unsupported peak number is not.
LOI Letter of intent Non-binding (usually) outline of a deal. Useful to start exclusivity and diligence. Not a substitute for a term sheet that actually allocates interconnect cost, COD risk, and offtake volume.
LNTP Limited notice to proceed Authorization for the EPC to spend a capped amount on long-lead items before full NTP. Used to order transformers and switchgear before every land covenant is complete.

M

Term In the transaction
Merchant power Selling (or buying) at spot LMP with no long-term PPA. Can look inexpensive in a model and move sharply in a scarcity week. Most training loads cannot rely on pure merchant; some mining and flexible inference can, if the contract says so.
Mining Proof-of-work compute. Highly interruptible, density-flexible, and sometimes a bridge load while an AI hall is built. Different offtake credit, community process, and SLA than training — those differences belong in the term sheet.
MOU Memorandum of understanding A lighter cousin of the LOI. Useful for alignment; it does not by itself control a site. Exclusivity has to be written as exclusivity.
MW / MWh Megawatt is capacity; megawatt-hour is energy through that capacity. Large renewable-to-compute opportunities are typically discussed first in MW (often 100 MW to 3 GW), then in GWh and shape so the power is actually usable for the workload.

N

Term In the transaction
NDA Non-disclosure agreement The start of every real conversation. Site files, queue positions, and offtake names leak value. Mutual NDAs are standard before a CIM, load letter strategy, or KYB pack moves.
NERC North American reliability organization. Large loads and generation above certain thresholds pick up reliability standards, registration, and cyber (CIP) obligations. Not just a utility problem.
Net-zero A corporate carbon target, often by a dated year. Distinct from 24/7 CFE. Offtakers may require a path (RECs, PPAs, on-site renewables) as a condition of the AI campus deal.
NTP Notice to proceed The instruction that lets EPC start full construction. Usually gated on financing close, permits, and interconnect. Splitting LNTP and NTP protects schedule without uncapped capex.

O

Term In the transaction
O&M Operations and maintenance Running the plant or campus after COD. Who pays, who staffs, and who is on the hook for availability should be in the offtake or JV — especially when BTM generation and the hall are different owners.
Offtake The agreement to take power, capacity, or hall MW. The offtaker’s credit is what lenders underwrite. Without contracted offtake, most campuses remain equity-only.
On-site generation Generation on the parcel or immediately adjacent — gas, solar, hydro, battery, sometimes nuclear later. The BTM half of a hybrid structure. It should be sized to contracted load, not to an uncommitted nameplate.
Opex Ongoing cost: power, water, staff, leases, maintenance. For AI, power opex dominates. A “cheap” site with a bad rate design or water constraint is not cheap.

P

Term In the transaction
PDU Power distribution unit The last electrical hop to the rack. High-density AI PDUs are a different SKU than office IT. Lead times and connector standards belong on the procurement list next to transformers.
Permitting Land use, environmental, air, water, and building approvals. AI campuses now draw the same political heat as power plants. Start community and water work in parallel with interconnect, not after.
PJM The large eastern RTO (Mid-Atlantic / Midwest). Deep market, heavy data-center demand, and a constrained interconnect/transmission story. A core region for AI load — and for queue and basis risk.
POD Two common uses: (1) point of delivery for power; (2) a repeatable hall “pod” of MW and cooling. Term sheets should say which one is meant. Phased pods are how campuses grow without restudying the entire load at once.
Powered shell A building with power and cooling to the hall, tenant finishes the white space. Faster than full turnkey for some AI tenants who want to spec liquid cooling themselves; still requires the MW to be real at the bus.
PPA Power purchase agreement Long-term contract to buy energy (and sometimes capacity and RECs) at a defined price and shape. Physical PPAs deliver electrons; virtual PPAs (VPPAs) are financial. AI deals often need a physical path to the meter plus a CFE story.
Promote A sponsor’s extra economics after investors hit a return hurdle — common in real estate and project JVs. Often paired with a capex success fee so origination and closing are both compensated.
PUE Power usage effectiveness Total facility power divided by IT power. 1.0 is perfect; legacy colo sits higher; well-designed liquid-cooled AI can be very efficient on paper. Still: interconnect and bills are on total MW, not IT MW. Never size the substation on IT load alone.

Q

Term In the transaction
Queue position Place in the interconnect study line. Transferable in some markets, tightly controlled in others. It should be confirmed like title — against the ISO queue, not only a marketing slide.

R

Term In the transaction
Rack density kW per rack, and increasingly kW per square foot. The input to cooling, fire, and how much white space a 100 MW block actually needs. Low-density assumptions on high-density GPUs are how halls get redesigned mid-build.
REC Renewable energy certificate The environmental attribute of a MWh. Can be bundled with a PPA or sold separately. Necessary but often not sufficient for hyperscale CFE claims, which are moving toward hourly matching.
Renewable mandate A corporate or offtaker requirement that a stated share of energy be renewable or carbon-free. A standard diligence question: it changes which sites and which PPA shapes will clear procurement.
RFP Request for proposal How hyperscalers and many utilities buy. Power RFPs ask for MW, COD, CFE, and credit. Site RFPs ask for land, water, and interconnect. Responses are won in the data room as much as in the presentation.
Right-of-way ROW Legal access for transmission, distribution, fiber, or water lines. A site without ROW to the substation or the long-haul fiber is not a site. Option the easements with the land.

S

Term In the transaction
SLA Service level agreement Uptime, power quality, and remedy (credits, termination). The hall SLA and the power PPA should not contradict each other. Training clusters typically need a tighter product than a legacy enterprise colo SLA.
SOC 2 A security attestation many AI tenants require of colo and cloud operators. Not a power term, but it shows up in the same MSA and can gate occupancy even when the MW is ready.
Solar Day-shaped, high-capex, zero-fuel generation. Strong midday energy; rarely sufficient alone as firm supply for training. Usually paired with storage, hydro, wind, and FTM backup.
Stranded power Generation or interconnect capacity that cannot reach a willing offtaker — curtailed plants, industrial sites with unused service, remote hydro. The classic origination thesis for behind-the-meter compute, if the rest of the site (fiber, water, community) works.
Substation Where voltage is stepped and switched. Often the longest-lead, highest-cost item on an AI campus. Ownership (utility vs customer), spare transformers, and spare feeders are term-sheet issues, not construction details.
Success fee Compensation paid when a defined close happens — offtake signed, capex committed, land closed. Often paired with promote so origination and the work that reaches close are both compensated.
Switchgear The breakers and gear that turn MW on and off safely. Medium-voltage gear for large campuses is long-lead and globally tight. LNTP exists largely because of switchgear and transformers.

T

Term In the transaction
Take-or-pay The buyer pays for contracted volume whether or not it is used. How generators and hall owners get financeable offtake. Offtakers typically negotiate ramp-in and curtailment exceptions; the agreed middle is what gets financed.
Term sheet The binding-intent commercial outline: parties, MW, price, COD, security, interconnect cost, and structure (PPA, JV, lease, promote). If it is not in the term sheet, assume it is not agreed.
Training Building the model. Wants dense, continuous, low-latency clusters and firm power. Least tolerant of curtailment. The workload that makes hybrid BTM+FTM and existing queue positions so valuable.
Transformer Steps voltage up or down. Large GSU and station transformers are multi-year procurement items. A signed PPA still needs a transformer strategy. Spare units are an availability decision.
Transmission High-voltage bulk movement of power. Congestion, upgrade cost, and timeline live here. BTM is partly a way to reduce exposure to that; FTM is how the site stays connected when on-site generation is short.
TIA-942 Telecom/data-center infrastructure standard (Rated 1–4). Useful language; not a substitute for a power and cooling design that matches actual kW/rack. “Rated 3” is not, by itself, an AI density spec.

U

Term In the transaction
Underwriting How capital decides the deal is real: counterparty credit, contracted cash flows, COD, interconnect, EPC wrap, insurance. AI residual GPU value is not an underwriting input for the power side.
UPS Uninterruptible power supply Bridges the gap until generators start. At AI densities, UPS runtime and battery chemistry are a major capex line. Some designs shrink UPS in favor of generator + flywheel or rack-level hold-up — a deliberate risk choice.
Uptime Institute The organization behind Tier I–IV certification. Tiers describe redundant infrastructure, not GPU performance. A Tier III hall designed for 10 kW racks is a different product than a high-density AI campus.
Utility The load-serving entity that meters the campus, and often owns the wires. Tariff, political will, and study bandwidth vary by territory. The utility relationship is a workstream from first load letter through energization.

V

Term In the transaction
VPPA Virtual power purchase agreement A financial contract on a generator’s output with no physical delivery to the campus meter. Helps CFE accounting; does not energize the hall. Physical power still has to come from the local utility or a BTM resource.
Voltage class The kV of the interconnect (e.g. 13.8, 69, 115, 230, 345). Higher voltage usually means more MW and more upgrade cost. The right class is an engineering and queue question, not a preference.

W

Term In the transaction
Water usage effectiveness WUE Water used per IT energy. Liquid cooling and evaporative plant can use substantial water. In the West and in drought basins, water rights can stop a site as quickly as interconnect. It belongs in diligence with the load letter.
Wheeling Moving third-party power across someone else’s wires, for a charge. How a remote hydro or wind plant can theoretically serve an AI campus. In practice: transmission availability, pancaked rates, and basis risk decide if it is real.
White space The raised-floor or slab area for IT racks, as opposed to gray space (electrical, mechanical). AI shrinks white space per MW because racks are denser — the mechanical yard grows instead.
Wholesale market ISO/RTO markets for energy, capacity, and ancillaries. The backdrop for LMP, congestion, and many PPAs. Retail tariff is what the campus may actually pay; wholesale is what the generator earns. The spread is the deal.
Wind Variable, often night-stronger than solar, high capacity in the plains. Pairs well with solar in a portfolio and with FTM backup. Curtailment and transmission (not turbines) are usually the AI-siting issues.
Workload What the chips are doing — training, inference, mining, or mixed. It determines firmness, latency, geography, and whether BTM shape is acceptable. Workload comes before site selection.

X

Term In the transaction
XLPE Cross-linked polyethylene cable, the usual medium-voltage feeder insulation. Shows up in EPC specs and ROW plans. Not exotic — but cable route and duct bank are how MW actually get from the substation to the hall.

Y

Term In the transaction
Year-1 availability Uptime in the first year after COD, when infant-mortality on plant and hall is highest. Lenders and offtakers often haircut Year-1. Ramp-in and liquidated damages should reflect that, not nameplate 99.99% from day one.
Yield Return on the power or real-estate vehicle. Distinct from GPU cloud margin. Take-or-pay offtake at a contracted $/MWh is typically underwritable; residual or merchant compute margin is a different risk product.

Z

Term In the transaction
Zero-carbon Generation with no operating CO₂ — hydro, wind, solar, nuclear. The supply side of CFE and net-zero claims. Still needs firming, transmission, and often hourly matching to satisfy offtaker procurement.
Zoning Local land-use permission for a data center or generation use. Rezoning can take longer than a queue study. Industrial or generation zoning on a brownfield is one of the quiet reasons those sites clear.