Reserved Capacity vs On-Demand LLM Spend

Reserved capacity commits you to a minimum spend or usage level over a defined term in exchange for a more favorable rate or a stronger availability guarantee; on-demand spend carries no commitment and a less favorable rate, priced purely on consumption. The choice between them should be driven by how predictable and confident your volume forecast is, not by which option has the better headline number.

The Structural Tradeoff

Both purchasing models exist because the provider is pricing in a version of the same risk you're evaluating: how confident is anyone that a given volume will actually be consumed. When you commit to reserved capacity, you're accepting the risk that your forecast is wrong in exchange for a rate or guarantee that reflects the provider's reduced uncertainty about your usage. When you stay on-demand, the provider prices in their own uncertainty about whether you'll show up at all, and passes that uncertainty back to you as a higher marginal rate.

Neither model is a discount independent of your usage matching the assumption behind it. Reserved capacity only pays off if you actually consume close to what you committed to. On-demand only makes sense if the flexibility is worth more to you than the rate difference — which depends heavily on how much your workload actually varies.

The Variables That Should Drive the Decision

Workload Predictability

A workload with stable, forecastable volume — a production feature with established usage patterns and low month-to-month variance — is the clearest case for reserved capacity, because the risk of overcommitting is low. A workload that's early-stage, still finding its usage pattern, seasonal, or subject to frequent changes in scope or architecture is a weaker case, because the forecast underlying any commitment decision carries more uncertainty than the commitment structure can absorb safely.

Volume Confidence Interval

Don't reduce this to a single forecast number — build a range. If your low-confidence and high-confidence volume estimates are far apart, that width itself is information: it tells you that a large commitment carries meaningful risk of being wrong in either direction, and that risk should push you toward on-demand or a smaller reserved baseline until the range narrows. See Token Usage Forecasting and Capacity Planning for how to build that forecast and its confidence interval from your own usage data rather than a single-point guess.

Switching and Architecture Risk

If you're likely to change models, providers, or the underlying architecture of the workload within the commitment term, a reservation locks you into paying for capacity that may not match what you actually need by the time you'd use it. Reserved capacity is a better fit for a stable, mature part of your workload than for one still undergoing significant iteration.

Overcommitment vs Undercommitment Asymmetry

These two mistakes are not symmetric in cost. Undercommitting on reserved capacity typically means paying the on-demand rate for the excess above your commitment — a known, bounded cost. Overcommitting typically means paying for capacity you never use, with no straightforward recovery path in most commercial structures. This asymmetry argues for sizing any commitment conservatively relative to your central forecast, rather than to your optimistic case.

A Decision Table

The Blended Approach

For most workloads past the early exploratory stage, a blend outperforms committing fully to either extreme: reserve a conservative baseline sized to your low-confidence volume estimate, and cover demand above that baseline on-demand. This captures the rate or availability benefit of reservation for the volume you're confident about, while avoiding the overcommitment risk on the portion of your forecast that's genuinely uncertain.

The baseline sizing decision should be revisited on a defined cadence, not set once and left static. As your volume forecast confidence improves — because you've accumulated more production history — the case for increasing the reserved baseline strengthens. As your workload changes materially — a new model version, a shift in task mix — the existing baseline may no longer reflect the right level of commitment, even if the volume number itself hasn't moved much.

How This Interacts with the Broader Contract

Reserved capacity is one term inside a larger provider contract, and it shouldn't be negotiated in isolation from the other commercial levers. Overage treatment in particular determines what happens the moment your usage exceeds the reserved baseline — a contract with a hard cutoff on overage makes undercommitment materially riskier than one with automatic on-demand billing above the baseline. See FinOps for AI Agents: Controlling Token and Tool Costs for the internal budget enforcement that should sit on top of whichever purchasing model you choose, so a forecast miss doesn't translate directly into an uncontrolled bill, and Securing LLM Provider API Keys for the operational controls that apply regardless of which commitment structure you negotiate.

Revisiting the Decision Over Time

Neither reserved capacity nor on-demand spend should be treated as a one-time decision locked in at the start of a workload's life. The right answer for a workload changes as it matures: a workload that starts on-demand because its volume is uncertain often becomes a strong candidate for a reserved baseline once it accumulates enough production history to narrow the forecast range. Conversely, a workload that was reserved during a stable period can become a poor fit for its existing commitment if a redesign changes its usage pattern significantly.

Build a review checkpoint into your planning calendar — tied to contract renewal dates if you're already committed, or to a fixed interval (quarterly is a reasonable default) if you're evaluating fresh. At each checkpoint, ask three questions: has our volume confidence interval narrowed or widened since the last review, has anything changed about the workload's architecture or model dependency that affects switching risk, and does the current commitment level still reflect a conservative read of our low-confidence estimate. Treating this as a recurring review rather than a set-and-forget decision is what keeps the purchasing model aligned with the workload as both evolve.

Common Questions

Is there a rule of thumb for what percentage of expected volume to reserve?

No defensible rule of thumb exists independent of your own forecast confidence — any fixed percentage borrowed from another organization's practice doesn't account for how wide or narrow your specific volume range is. The more useful discipline is sizing the reserved baseline explicitly to your low-confidence estimate, and treating that sizing decision as something to revisit as your forecast confidence changes, rather than adopting a static ratio.

Does reserved capacity make sense for a workload still in pilot?

Generally not, unless the pilot has run long enough to produce a volume history with low variance, or the reservation term is short enough that the overcommitment downside is small. Pilots exist specifically because volume and usage pattern are still uncertain — committing capacity against that uncertainty inverts the purpose of running a pilot in the first place.

How does multi-provider routing change this decision?

If your architecture routes across multiple model providers — see Model Routing: Cost and Quality Tradeoffs — you may need to make this reserved-versus-on-demand decision independently for each provider, since your volume confidence and switching risk can differ significantly by provider. A workload that's stable in aggregate but routed dynamically across providers may still look volatile from any single provider's perspective, which argues for on-demand with that specific provider even if your total workload volume is predictable.

What Good Looks Like

• The commitment decision is driven by a volume confidence interval built from your own data, not a vendor's suggested tier.
• Reserved capacity is sized to a conservative, low-confidence estimate rather than an optimistic one, given the asymmetric cost of overcommitment.
• A blended approach is considered before committing fully to either reserved or on-demand.
• Overage treatment in the underlying contract is understood and factored into the sizing decision.
• The baseline is revisited on a defined cadence as forecast confidence and workload characteristics change.