Reserved vs on-demand GPU pricing, sourced OCT 13, 2026 · 5 MIN READ We pulled the public numbers on CoreWeave, AWS, and Lambda — then show why utilization, not sticker price, decides whether committing saves you money. Reserved GPU capacity runs 20–60% cheaper than on-demand across every major provider we checked — but only if you actually keep the hardware busy. Here’s what the market is charging, sourced, and the one number that decides whether a commitment is the right call for you. 01 / WHAT THE MARKET IS CHARGING, RIGHT NOW Third-party pricing trackers put an 8-GPU CoreWeave B200 node at roughly $68.80/hr — about $8.60 per GPU-hour on demand — with reserved contracts reported to discount up to 60% depending on term. AWS’s P5 instances (8×H100) list around $55/hr on demand; aggregator snapshots show 1-year reserved pricing near $41–43/hr and 3-year terms closer to $24/hr, roughly 55–57% off sticker. Lambda Labs’ on-demand H100 pricing is the least consistent of the three — different trackers show anywhere from $2.49 to $4.29 per GPU-hour for the same instance class — with reserved discounts estimated, not officially published, at around 20% for one year and 26% for three. 20–60% Reported discount range for reserved vs. on-demand GPU pricing across the providers we checked (CoreWeave, AWS, Lambda), per third-party aggregator data, October 2026.* None of this comes from the vendors’ own pricing pages — it’s aggregator data, and the sources frequently disagree with each other by double digits. That disagreement is itself useful information: sticker price is a moving, contested target, and nobody — us included — publishes a single rate that survives six months unchanged. If a number matters to your budget, verify it directly with the provider before you model it. 02 / THE NUMBER THAT ACTUALLY DECIDES THE DEAL Sticker price isn’t what decides reserved vs. on-demand — utilization is. One widely-cited 2026 analysis found that at 5% utilization, an AWS H100 billed at its on-demand rate effectively costs over $245 per used GPU-hour once idle time is counted against it. Reserved capacity only wins once you’re running it most of the time, which is exactly the baseload math we’ve written about separately: the committed rate is a bet on your floor, not your ceiling. 03 / WHERE THAT LEAVES OUR ~$5/GPU-HOUR TARGET We’re not claiming to be the cheapest line item on anyone’s spreadsheet, and we don’t publish a single sticker price — see what “qualifying deployments” actually means for why. What we can say: our ~$5/GPU-hour target for qualifying long-term commitments lands in roughly the same band as the deeper multi-year reserved tiers reported for H100-class capacity elsewhere in the market, and below most on-demand sticker rates for B200-class hardware. The number that applies to you depends on GPU type, node configuration, region, quantity and term — same as it does everywhere else. 04 / WHEN NONE OF THIS SHOULD MATTER TO YOU If your workload is still spiky, seasonal, or you’re not yet sure it’s staying at this size, none of the reserved rates above are the right comparison — stay on-demand or spot, and revisit the commitment question once your utilization chart settles down. The honest version of this comparison isn’t “we’re cheaper than CoreWeave” or “we’re cheaper than AWS” — none of us publish numbers stable enough to make that claim responsibly. It’s that committed pricing, from any provider, only beats on-demand once your utilization crosses the line where flexibility stops being worth its premium. * Competitor figures are third-party aggregator data (Holori, Devzero, Spheron Network), not official vendor price pages, captured October 2026. Sources disagree with each other, sometimes by double digits, and rates change by region, configuration and term. Confirm current pricing directly with each provider before budgeting. $5/hr is our own target/starting rate for qualifying long-term commitments, not a universal price.