Glossary

These are the working meanings used in this book. Where a provider or technical discipline defines a term more narrowly, keep its definition beside this one and say which you mean.

Accepted task. A task that met the business’s acceptance criteria for its task class, not merely a completed request or a response that arrived on time. Cost per accepted task divides the workload’s total cost, including failed attempts, attributable subscriptions, and committed capacity, by the number of accepted tasks.

Attribution. Connecting spend or consumption to the workload, team, customer, or decision that caused it. Shared costs need an agreed rule. Attribution can support accountability without giving the recipient control over the cost.

Baseline forecast. A forecast made with an explicitly defined starting method and information set, such as a seasonal-naive forecast, which repeats the latest observation from the matching season. Keep it separate from later business-event adjustments so their contribution can be assessed.

Basis risk. The gap between what a hedge’s price index measures and what we pay for the hardware, region, terms, and dates we actually need. A position on an index offsets a bill only to the extent that the bill moves with the index.

Budget. An authorized financial boundary or plan. It says what spending is permitted or expected, not whether that amount will support the service’s obligations.

Capacity. The work a system can support in a stated interval under stated workload, quality, configuration, and operating conditions. A measured result applies within the evidence behind it; an estimate must carry its assumptions.

Capability. What the business and its technology need to be able to do: features, reliability and resilience, customer experience, and operational outcomes. Capacity and cost are judged against it; see the Introduction.

Cohort. Workloads grouped for a stated planning reason, with maintained membership, an owner, and a way to detect exceptions. A cohort is more than a residual line called “Other.”

Commitment. A commercial or operational promise whose parties, scope, and term must be stated. A spending commitment, such as a savings plan or minimum spend, can lower the price without holding any supply; a capacity hold can secure supply without a discount; a service commitment promises behavior to a customer. Each can exist without the others, so name the kind.

Coordinated omission. A load-testing error in which a generator that waits for each response stops sending the requests that would have suffered most during a stall, so measured latency understates what users arriving on their own schedule would see.

Cost. The money and obligations a system creates while it runs, whether or not anyone has attributed them yet. Cost behavior is how a cost changes over time and responds to demand.

CPU-bound. Limited materially by available processing capacity under the observed conditions. A high CPU reading alone does not establish the limiting mechanism.

Data-bound. In this book, work whose limiting behavior depends materially on data access, movement, layout, state, or coordination. It is broader than disk-bound, and the limiting resource can change as the data or its use changes.

Demand intelligence. Information about future business activity, kept with its source, evidence date, assumptions, timing, dependencies, and a condition for review; also the practice of gathering it. It is how a capacity team comes to understand the systems that drive consumption, and much of it arrives through relationships (Volume II, Chapter 17).

Double bubble. The temporary overlap of old and replacement environments during a transition. It does not require exact doubling of cost. Count useful business outcomes once even when both systems process them.

Elasticity. The ability to add and release resources as demand changes. Whether the bill falls on release depends on the purchase arrangement, and at large scale elastic supply is an inventory promise somebody has to fund.

Evidence date. The date information became available to the planners, as distinct from the date of the activity it describes. Judge a forecast only against the evidence available on its as-of date.

Failure group. Resources assumed to become unavailable together in a specified scenario. Independence between groups is an assumption requiring evidence, not a property conferred by naming them differently. Degraded capacity is what remains after a group is lost: in the book’s service, 48 workers at 250 requests per second each, or 12,000 requests per second.

Fixed cost. Cost that does not change with the relevant activity over a specified range and period. A cost can be fixed this month and avoidable next year.

Forecast. An estimate of future demand under stated assumptions, with an as-of date showing what was known when it was made. A revision after plans change is a new forecast, not evidence that the earlier one was accurate.

Headroom. The deliberately unused portion of capacity, bought on purpose to absorb demand and failure. Calculate it as usable capacity minus demand under the same stated conditions, and say which state you mean: in Chapter 2, 16,000 minus 10,000 leaves 6,000 requests per second normally, but 12,000 minus 10,000 leaves only 2,000 after one failure group is lost.

Lead time. The interval between deciding to obtain supply and its readiness date, when it is qualified and usable for the stated work. Approving the expenditure does not shorten it.

Little’s law. In a stable system, the average number of items present equals their arrival rate times the average time each spends there, $L = \lambda W$. It relates long-run averages, not percentiles or peaks.

Long tail. The many smaller consumers in a ranked portfolio. This differs from the statistical upper tail of unusually large observations over time. Neither concentration nor an 80-percent share proves a power-law distribution.

Marginal cost. The change in cost associated with an additional unit or a specified change in work. It may differ substantially from average allocated cost.

Metered cost. Cost calculated from counted consumption, such as requests, worker-hours, GB-months, or tokens, and a pricing rule. Metering describes the billing mechanism; it does not exclude a minimum payment or commitment.

Obligation. What the organization has promised to pay, including promises that outlast the reason for making them. The contract, not current use, determines which kind of obligation exists and what can stop if the workload disappears.

Opportunity cost. The value of the best alternative displaced by a choice. An unpriced consequence is unknown, not necessarily zero.

Prediction interval. A range derived from a forecast distribution under specified assumptions. Assess its coverage at the relevant horizon; do not label uncalibrated business scenarios as percentiles.

Provisioned throughput. Model-serving capacity held for a deployment and billed by the hour, busy or idle. The work it holds varies with the model and the prompt mix; quota permits it without holding it, and a reservation discounts it without guaranteeing it.

Qualified capacity. Resources demonstrated to perform the intended work under the relevant conditions. Requested, purchased, delivered, and qualified are different states, each with its own date.

Quota. A limit on permitted resource allocation or consumption within a defined scope. Approved quota permits a deployment or reservation but holds no supply; the provider’s rules determine what a particular quota means.

Reconciliation. Making related views add up under consistent boundaries. Agreement between totals does not demonstrate forecast accuracy or resolve different business intentions.

Request. A statement of what someone wants, which begins investigation or negotiation. Submitting it does not establish that the resources exist, that funding is approved, or that another customer will give up an allocation.

Reservation. A commercial arrangement whose effect depends on its terms: a price discount, a hold on specified capacity, or both. Establish which benefit is included before relying on it. The name alone does not guarantee location, configuration, timing, or failure coverage. Inside an organization the word may instead mean an internal hold or a place in a planning queue; ask what is reserved, for whom, where, for which dates, and under what release rule.

Reserve. Capacity held for an identified purpose, with stated readiness, authority to use or release it, and review conditions. Shared reserve needs explicit assumptions about simultaneous calls on it.

Rolling-origin evaluation. Evaluating forecasts made from successive past starting points, each a forecast origin, using only information available at that origin. Match evaluation horizons to the decisions and supply lead times.

Saving. In this book, a reduced expenditure: an actual obligation or charge that fell within the stated boundary. Chapter 13 keeps it apart from an avoided purchase, which changes a forecast rather than a bill; released capacity, available for another use; shifted cost, now carried by another workload or organization; and improved service, a useful change in quality, capability, or timing without a lower bill.

Savings plan. A commitment to spend a fixed amount per hour for one or three years in exchange for discounted rates on eligible usage. The commitment applies hour by hour; an unused hour expires but is still paid for, and the plan holds no capacity.

Scenario. A stated combination of conditions used to examine consequences and choices. A scenario need not have an estimated probability.

Shadow price. In a linear model, the local value of relaxing a constraint by one unit. An observed market price is not automatically that value.

SKU. Stock-keeping unit, borrowed from retail inventory: one sellable configuration of a service in a provider’s catalog, with its size, features, and price. The name encodes the configuration — Microsoft’s Standard_D4s_v5 is a general-purpose family, four vCPUs, premium-storage capable, fifth version — and is worth reading before pricing or reserving. Commitments, quotas, shortages, and supply reservations are SKU-specific, and the choice among SKUs is an engineering decision with a price attached.

Showback and chargeback. Showback makes attributed cost visible to the groups responsible for it; chargeback assigns those costs to budgets or accounting structures. Both require an understandable basis, correction routes, and a recipient with an appropriate role.

Spillover. Sending requests that exceed provisioned capacity to shared pay-as-you-go capacity, billed at that rate. It supports reserving the steady base and buying the peak on a shared tier.

Take-or-pay. A commitment charged for the committed quantity whether or not it is used. In Volume II, Chapter 20’s illustration, $7 per unit committed against $10 bought as needed breaks even at 70 percent use.

Systems that drive consumption. The customers, product decisions, architecture, recovery requirements, migrations, and AI behavior that produce the consumption a bill and a capacity chart report. In this book they, rather than the bill or the server count, are the object of management.

Target. Something the organization intends to achieve. Capacity planning examines a target as a scenario; an ambitious sales target is not a prediction of realized demand.

Token budget. An allowance of specified token categories over a stated period. Name what happens at the limit, because an alert, a cutoff, and an advisory budget make different promises. It is distinct from a dollar budget, an enforced rate limit, and measured serving throughput, and it does not convert into promised work.

Token economics (Tokenomics). The relationship among AI workload design, consumption, supplied capacity, cost, and accepted work. Tokens are accounting and technical units, not a universal measure of business value.

Unit economics. The cost of one completed unit of useful work, rather than of a machine-hour, with a stated cost boundary and quality condition. It informs decisions alongside capacity, timing, risk, and obligations.

Usable supply. Capacity qualified and available at the required place and time, after the constraints relevant to the decision are applied.