According to GridStackHub.ai data, the cheapest H100 reserved instance in 2026 is $1.79/hr at CoreWeave (1-year commitment, single GPU SXM5) — a 20% discount versus their $2.23/hr on-demand rate. For 8-GPU clusters, AWS 1yr Reserved Instance on p5.48xlarge costs $19.22/hr ($2.40/GPU) — 41% below on-demand at $32.77/hr. GCP Committed Use Discounts on a3-highgpu-8g are $19.63/hr (37% savings). Azure Reserved VM Instances on ND H100 v5 cost $20.49/hr (37% savings). GridStackHub tracks both on-demand and reserved pricing across all providers daily.
Maximum savings on H100 via AWS 1-year Reserved Instance (p5.48xlarge: $19.22/hr vs $32.77/hr on-demand). CoreWeave single-GPU: $1.79/hr reserved vs $2.23/hr on-demand. Independent clouds negotiate 15–25% off for 6-month+ commitments.
AWS 1yr RI
$19.22/hr vs $32.77/hr
8x H100 SXM (p5.48xl)
GCP CUD 1yr
$19.63/hr vs $31.21/hr
8x H100 SXM (a3-high)
Azure 1yr RI
$20.49/hr vs $32.78/hr
8x H100 (ND H100 v5)
CoreWeave 1yr
$1.79/hr vs $2.23/hr
1x H100 SXM5 (per GPU)
Complete GPU Reserved Pricing Table — All Providers (May 2026)
GridStackHub tracks every confirmed reserved/committed pricing tier across hyperscalers and independent GPU clouds. Below is the complete comparison of on-demand vs reserved rates for H100, including per-GPU cost normalization for easy comparison:
| Provider | Instance / Config | GPU Count | Term | On-Demand /hr | Reserved /hr | Savings | Per GPU /hr |
|---|---|---|---|---|---|---|---|
| AWS | p5.48xlarge (H100 SXM) | 8 | 1 year RI | $32.77/hr | $19.22/hr | 41% | $2.40 |
| GCP | a3-highgpu-8g (H100 SXM) | 8 | 1 year CUD | $31.21/hr | $19.63/hr | 37% | $2.45 |
| Azure | ND H100 v5 (8x H100) | 8 | 1 year RI | $32.78/hr | $20.49/hr | 37% | $2.56 |
| CoreWeave | H100 SXM5 (per GPU) | 1 | 1 year | $2.23/hr | $1.79/hr | 20% | $1.79 |
| Lambda | H100 SXM (8x node) | 8 | 6mo+ contract | $15.92/hr | ~$13.50/hr EST | ~15% | ~$1.69 |
| FluidStack | H100 SXM5 80GB | 1 | 3mo+ contract | $2.15/hr | ~$1.83/hr EST | ~15% | ~$1.83 |
| DataCrunch | H100 SXM5 80GB | 1 | 6mo+ contract | $2.20/hr | ~$1.87/hr EST | ~15% | ~$1.87 |
VERIFIED = confirmed from published provider pricing pages. EST = estimated based on market norms for committed contracts (typically 15–25% below on-demand for independent GPU clouds). Hyperscaler reserved data from AWS EC2 Reserved Instances, GCP Committed Use Discounts, and Azure Reserved VM Instances pages, April–May 2026. Prices subject to change.
Independent cloud per-GPU reserved rates beat hyperscaler 8-GPU rates. CoreWeave's $1.79/hr per H100 GPU on a 1-year reserve is cheaper per GPU than any hyperscaler reserved cluster ($2.40–$2.56/GPU at AWS/GCP/Azure). The catch: CoreWeave single-GPU and small-node configs don't offer hyperscaler NVSwitch interconnect at scale. For massive clusters (64+ GPUs), hyperscalers still win on networking.
AWS GPU Reserved Instances: How They Work
AWS Reserved Instances for GPU workloads apply to the p5 (H100), p4 (A100), and G6e (L40S) instance families. AWS offers three payment options for 1-year terms:
- All Upfront: Pay the full 1-year cost upfront. Maximum discount — typically 2–3% more savings than Monthly. Best for teams with CapEx budget.
- Partial Upfront: Pay ~50% upfront, rest monthly. Moderate discount between All Upfront and No Upfront.
- No Upfront: Monthly payments for the commitment period. Lowest discount (~37% on H100) but preserves cash flow.
For the p5.48xlarge (8x H100 SXM5), the 1-year No Upfront Reserved Instance is $19.22/hr versus $32.77/hr on-demand — saving $13.55/hr or approximately $119,000/year per node. At 80% utilization, the annual savings per node exceed $95,000.
| AWS Instance | GPU | On-Demand /hr | 1yr RI /hr | Savings | Annual Savings/Node |
|---|---|---|---|---|---|
| p5.48xlarge | 8x H100 SXM | $32.77/hr | $19.22/hr | 41% | ~$119k/yr |
| p4de.24xlarge | 8x A100 80GB | $40.97/hr | ~$24.50/hr EST | ~40% | ~$144k/yr |
| g6e.48xlarge | 8x L40S 48GB | $13.74/hr | ~$8.24/hr EST | ~40% | ~$48k/yr |
Google Cloud GPU Committed Use Discounts (CUD)
GCP's equivalent of reserved instances is the Committed Use Discount (CUD). Key differences from AWS:
- Billed as a resource commitment, not an upfront purchase. You commit to a specific machine type and region for 1 or 3 years. GCP bills monthly; no large upfront payment required.
- Machine-type specific. An a3-highgpu-8g CUD applies only to that machine type in the committed region. You cannot apply it to H200 instances later.
- 1yr CUD saves ~37%. 3-year CUDs save up to 55% on eligible GPU machine types — the most aggressive long-term discount of any hyperscaler.
- Sustained use discounts do NOT stack with CUDs. GCP's automatic sustained use discounts are replaced, not supplemented, by CUD pricing.
| GCP Instance | GPU | On-Demand /hr | 1yr CUD /hr | 3yr CUD /hr | Max Savings |
|---|---|---|---|---|---|
| a3-highgpu-8g | 8x H100 SXM | $31.21/hr | $19.63/hr | ~$14.05/hr EST | ~55% |
| a3-megagpu-8g | 8x H200 SXM | $40.32/hr | ~$25.40/hr EST | ~$18.14/hr EST | ~55% |
| a2-ultragpu-8g | 8x A100 80GB | $29.39/hr | ~$18.50/hr EST | ~$13.22/hr EST | ~55% |
GCP's 3-year CUD is the most aggressive long-term GPU discount from any hyperscaler. At an estimated 55% off on-demand for the a3-highgpu-8g, a 3-year GCP CUD on H100 costs roughly $14/hr for 8 GPUs — less than $1.75/GPU/hr, comparable to CoreWeave's 1-year reserved single-GPU pricing. If you can forecast 3 years of stable H100 demand, GCP 3yr CUD is worth modeling carefully.
Azure Reserved VM Instances for GPU
Azure Reserved VM Instances for GPU work similarly to AWS Reserved Instances — you commit to a VM size in a region for 1 or 3 years in exchange for ~37% discounts (1yr) or ~55% discounts (3yr). Key Azure specifics:
- Scope flexibility: Azure Reserved VMs can be applied at the subscription or resource group level. Shared scope allows the reservation to apply to any VM of the reserved type in the subscription.
- Instance size flexibility: Azure offers instance size flexibility groups — an ND H100 v5 reservation may apply to smaller related VMs if the reserved instance isn't fully utilized, reducing wastage risk.
- Exchange and cancel: Azure allows exchanging reservations for different VM types (within the same family) — more flexible than AWS or GCP.
| Azure VM | GPU | On-Demand /hr | 1yr RI /hr | Savings |
|---|---|---|---|---|
| ND H100 v5 | 8x H100 SXM | $32.78/hr | $20.49/hr | 37% |
| ND H200 v5 | 8x H200 | $44.52/hr | ~$28.00/hr EST | ~37% |
| ND A100 v4 | 8x A100 80GB | $32.77/hr | ~$20.60/hr EST | ~37% |
CoreWeave Reserved GPU Pricing
CoreWeave is the leading independent GPU cloud with published reserved pricing — a significant differentiator from peers like Lambda and FluidStack who only offer committed contracts on request.
CoreWeave's 1-year reserved H100 at $1.79/hr per GPU is notable for two reasons: it's the cheapest published per-GPU H100 reserved price of any provider, and it applies to single-GPU commitments (not requiring an 8-GPU node purchase). This makes it accessible for teams running 1–4 GPU workloads who want commitment discounts without hyperscaler overhead.
CoreWeave reserved ≠ hyperscaler reserved. CoreWeave reserved contracts typically require direct negotiation with their sales team for multi-node commitments. The $1.79/hr published rate is for single-GPU H100 SXM5 on a 1-year term. For 8+ GPU node reservations, CoreWeave pricing is custom and generally includes infrastructure SLAs, InfiniBand fabric commitments, and support tiers not included in per-GPU rates.
When Reserved GPU Instances Are Worth It
The math is straightforward: reserved pricing pays off when your utilization exceeds the break-even point. Here is the break-even analysis for the most common reservation:
| Provider / Term | On-Demand /hr | Reserved /hr | Break-Even Utilization | Annual Savings at 80% Util |
|---|---|---|---|---|
| AWS 1yr RI (H100 8-GPU) | $32.77 | $19.22 | ~59% | ~$94,700/yr |
| GCP 1yr CUD (H100 8-GPU) | $31.21 | $19.63 | ~63% | ~$80,900/yr |
| Azure 1yr RI (H100 8-GPU) | $32.78 | $20.49 | ~63% | ~$85,800/yr |
| CoreWeave 1yr (H100 1-GPU) | $2.23 | $1.79 | ~80% | ~$3,075/yr |
Break-even utilization = (Reserved Rate / On-Demand Rate). Below break-even, you save more with on-demand. Above it, reserved wins. Key insight: AWS and GCP reservations break even below 65% utilization — very achievable for production workloads. CoreWeave's lower absolute discount means you need 80%+ utilization to benefit.
Model reserved vs on-demand for your specific workload
Enter your GPU model, hours per month, and utilization rate — see exact break-even and annual savings for each provider's reserved pricing.
Open GPU Cost Calculator →H200 Reserved Pricing: What's Available in 2026
H200 reserved pricing is less standardized than H100 in 2026. The H200 launched in late 2024 and most cloud providers are still in the process of adding formal reservation tiers. Current status:
- AWS: H200 reserved instances (p5e) are available in limited regions. Pricing not yet published at scale — inquire via AWS sales for committed rates.
- GCP: a3-megagpu-8g (8x H200) CUDs are available. Estimated ~37% discount for 1yr ($25.40/hr vs $40.32/hr on-demand).
- CoreWeave: H200 reserved contracts are available on request. No published list price; expect similar 15–25% discounts vs $3.49/hr on-demand.
- Lambda: H200 ($2.99/hr on-demand) — long-term committed contracts available on request. No published reserved rate.
A100 Reserved Pricing: Still Relevant in 2026
Despite A100 being two generations behind H100, it remains relevant for mid-scale training and fine-tuning workloads where 80GB VRAM is sufficient and cost efficiency matters more than raw throughput. Reserved A100 pricing in 2026:
| Provider | Instance | GPUs | On-Demand /hr | Reserved /hr | Per GPU /hr |
|---|---|---|---|---|---|
| Lambda | 1x A100 SXM 80GB | 1 | $1.29/hr | ~$1.10/hr EST | ~$1.10 |
| FluidStack | A100 SXM 80GB | 1 | $1.21/hr | ~$1.03/hr EST | ~$1.03 |
| CoreWeave | A100 SXM 80GB | 1 | $1.62/hr | ~$1.38/hr EST | ~$1.38 |
| AWS | p4de.24xlarge (8x A100) | 8 | $40.97/hr | ~$24.50/hr EST | ~$3.06 |
Get reserved pricing alerts
New reserved pricing tiers, new providers launching commitment contracts, and rate changes — we'll notify you. Free, no credit card required.
Frequently Asked Questions
Compare reserved vs on-demand vs spot for your stack
GridStackHub tracks reserved, on-demand, and spot GPU pricing across 32+ providers daily. Model the full cost for your specific GPU, utilization, and term.
View Full Pricing Database →