Committed-Use Pricing: Examples & Companies

What is it

It is the enterprise counterpart to pure-usage pricing. Where pure usage lets a buyer start at zero and pay only for what they consume, committed-use asks the buyer to put a floor under their spend — a dollar minimum, a reserved machine, or a prepaid volume — and rewards that floor with a lower per-unit rate. The trade is symmetric: the vendor converts a volatile usage stream into predictable, bookable revenue, and the buyer converts an unpredictable bill into a known rate.

The discount is almost always the headline. CoreWeave lists 8-GPU HGX H100 nodes at $49.24/hr on-demand but quotes “up to 60% discounts over our On-Demand prices for committed usage” — and those reserved contracts, not the public card, carry most of its $5.13B in 2025 revenue. Vast.ai advertises up to 50% off its marketplace rate for reserved, pre-paid capacity; Nebius cuts up to 35% for multi-month and large-cluster commitments. The mechanism is identical across very different products: spend certainty buys rate.

Committed-use rarely stands alone. It is almost always the top rung of a ladder that starts with self-serve usage — a free trial or pay-as-you-go entry, then volume tiers, then an annual commit reserved for the enterprise buyer. That structure is so consistent among infrastructure vendors that it has become a documented cross-corpus pattern: the infrastructure-layer commitment-pricing trend tracks how GPU-compute and data vendors converged on offering identical “annual commit” enterprise tiers alongside on-demand rates.

How it works

The unit math is what makes the model work for both sides. A commit is a floor, not a cap: the buyer pays at least the committed amount, draws down usage against it, and — at most vendors — pays standard on-demand rates on anything above the commitment. LanceDB’s AWS Marketplace listing is unusually explicit: a $60,000/year committed contract with $0.01 per LCU billed on overage above the committed quantity. Fireworks AI, Baseten, and Groq similarly bill over-commitment usage at list price, so the discount applies only inside the committed band.

Worked example. Suppose a team runs steady inference on CoreWeave H100 nodes, one 8-GPU HGX H100 node continuously for a year:

• On-demand: $49.24/hr × 730 hr/mo × 12 = roughly $431,000/year, fully variable and cancellable any hour.
• Reserved commit at up to 60% off: roughly $172,000/year at the deepest quoted discount — but the team is now contractually on the hook for that node whether or not it runs a single job.
• If the team needs to burst above one node, the extra capacity bills at the on-demand rate, not the reserved rate — the commit is a floor, not a ceiling.

The break-even logic is the buyer’s whole decision: a commit only pays off if forecasted usage reliably clears the floor. That is why committed-use clusters in vendors whose customers have predictable, always-on workloads — GPU inference, proxy bandwidth, vector search — and is far rarer in spiky, experimental usage. Use the pricing calculator hub to model commit-versus-on-demand break-even, and see the prepaid-credits models guide for how prepaid commitments draw down against usage.

Companies using this

Patterns observed

GPU compute is the center of gravity. The largest concentration is GPU-hour vendors, where the underlying capital cost makes utilization guarantees valuable to both sides. Anyscale layers committed contracts on its $1-per-ACU rate card (from $0.0135/hr CPU up to $10.68/hr for an H200); RunPod and Together AI both drop committed cluster rates below their published on-demand token and GPU cards. Lambda is a telling variant — its 1-Click Clusters publish hard per-GPU-hour numbers ($5.54–$6.16/hr for committed H100 clusters) while on-demand H100 SXM sits at $3.99/hr, inverting the usual logic because scarce frontier capacity is a sellers’ market.

Discounts scale with commitment depth, and the deepest ones are gated. The published-ladder version is legible: Nebius prints both on-demand ($3.85 H100) and preemptible ($2.15 H100) rates against its commit line, and Snowflake Cortex sets credit price by edition (Standard $2, Enterprise $3, Business Critical $4 per credit) so the platform tier itself is the commitment. But most vendors quote the deepest tier privately — Baseten, Fireworks, and Replicate all say “annual commit, contact sales” and reveal the reserved rate only in a quote. The published-ladder approach is the exception, not the rule.

The commitment is increasingly disguised as a tier minimum or a seat-plus-commit hybrid. turbopuffer sets a monthly minimum spend per tier that functions as a soft commitment floor: pick a tier, and you’ve effectively committed to that minimum even on pure usage. Lorikeet does the same with annual credit pools billed monthly. Nomic fuses commitment into a seat model entirely — its Platform Business plan is $40/user/month on an annual commit, with a 25-seat minimum and a $1,000/month platform commitment, so the floor is baked into both the seat count and the platform fee. The floor is the commitment, dressed as packaging.

Counterexamples & variants

LanceDB and m3ter: the platform-fee floor. Not every commitment is a discount mechanism. LanceDB’s public /pricing page is a contact form, not a price table — its committed AWS Marketplace contract is a fixed annual quantity with a per-LCU overage rate, not a published discount off a list rate. m3ter’s “commitment” is a custom core platform fee bundling allowances for usage data ingested and bills calculated — a committed platform spend to access the product at all. Both are reminders that a commit can be a minimum-to-enter rather than a reward for scale.

Reserved-capacity commits carry lock-in risk the dollar-floor variant doesn’t. Vast.ai is explicit that reserved credits are locked to a single machine — pre-pay for a host and your capacity is tied to it, not portable across the marketplace. DeepInfra’s 3-year DeepCluster delivers among the corpus’s lowest GPU rates precisely because the buyer absorbs three years of obsolescence risk on a fast-moving hardware curve. Lambda shows the sharpest edge of this: its on-demand H100 SXM rate rose from $2.99 to $3.99 through 2025–2026 as capacity tightened, so a buyer who committed early locked in scarcity protection while a buyer who waited paid more. The deeper the discount and the longer the term, the more the buyer is betting on a hardware and demand curve that can move against them.

The mid-tier gap is where committed-use silently fails. The most common structural weakness is not a bad commit but a missing one. Fireworks AI’s own teardown flags that self-serve customers grow into the $10K–$50K/month band with no published mid-tier discount before the Enterprise commit — so a growing account sees on-demand rates all the way up until a sales-gated jump. That gap is exactly where accounts shop around, and it is a variant failure mode distinct from the gate-only problem: the ladder exists, but it skips a rung.

What this means for buyers vs vendors

For buyers

Commit only against the floor you are confident you will clear in your worst month, not your average month. The worked example above turns negative the moment forecasted usage dips below the committed capacity, because you still pay the floor whether the node runs or not. Favor dollar-floor commits (an annual quantity, a platform fee, a tier minimum) over reserved-capacity commits (machine-locked credits, a multi-year cluster) when your workload mix is uncertain — a dollar floor is fungible across products; a reserved machine is not.

Watch the lock-in window closely on GPUs. A multi-year reservation delivers the deepest rate but spans several hardware generations; the H100 you commit to today may be the previous generation before your term ends. Always extract the overage rate before signing: a pre-committed pool sold without a published overage rate, as Lorikeet does, leaves you unable to model the downside — insist on a stated rate for everything above the floor.

Use the visible rate cards as leverage. Where a vendor prints its on-demand, preemptible, and committed numbers, you can anchor a negotiation for the sales-quoted tiers against those published figures. Model the break-even before you talk to sales — the pricing calculator hub and the prepaid-credits models guide both help translate committed spend into an effective per-unit rate.

For vendors

Committed-use converts volatile usage into bookable revenue, which is why every infrastructure vendor in this corpus offers it — public-market investors reward a contracted backlog far more than a self-serve rate card. The commit is not a discount you give away; it is the mechanism that turns spiky demand into a predictable, financeable revenue stream.

Publish the ladder rather than gating it. The visible-discount approach outconverts the gate-only approach, because the buyer can see the saving and self-qualify: a growing account facing a printed commit line models the upgrade itself, while an invisible Enterprise gate forces every conversion through a sales call and loses the ones who won’t book a demo. Fill the missing middle so accounts don’t shop around exactly when they are most valuable, and state the overage rate up front. The usage-invoicing & billing-cycles guide covers how to reconcile commits, drawdowns, and overage cleanly on the invoice.

Explore this theme in the knowledge graph →