Provider record · updated 2026-09-10
AWS — GPU pricing & deliverability
15 tracked offerings · grade mix A:10 · B:4 · C:0 · D:0 · U:1. AWS P5 family: 3200 Gbps EFA per instance (p5=EFA, p5en=EFAv3 on Nitro v5), 900 GB/s NVSwitch intra-node; EC2 UltraClusters published claim of scaling up to 20,000 H100/H200 GPUs with up to 20 exaflops aggregate. P6: EFAv4 at 3.2 Tbps (B200) and 6.4 Tbps (B300) per instance; P6e-GB200 UltraServers publish 72 GPUs per NVLink domain (u-p6e-gb200x72) with 28,800 Gbps EFA per UltraServer and 'up to 130 TB/s' NVLink; UltraCluster scale for P6 stated only as 'tens of thousands of GPUs within a petabit-scale nonblocking network'.
Tracked offerings
| GPU | SKU | $/GPU-hr | Terms | Grade | Conf | Fabric (as published) | Src |
|---|---|---|---|---|---|---|---|
| H100-SXM | p5.48xlarge (3yr) | $2.97 | savings-plan · min: 3yr | A | publ | EFA, 3200 Gbps per instance (AWS P5 page) | ↗API |
| H200-SXM | p5en.48xlarge (3yr) | $3.42 | savings-plan · min: 3yr | A | publ | EFAv3 (Nitro v5), 3200 Gbps per instance, 'up to 35% improvement in latency' vs earlier EFA (AWS P5 page) | ↗API |
| H100-SXM | p5.48xlarge (1yr) | $4.33 | savings-plan · min: 1yr | A | publ | EFA, 3200 Gbps per instance (AWS P5 page) | ↗API |
| H200-SXM | p5en.48xlarge (1yr) | $4.98 | savings-plan · min: 1yr | A | publ | EFAv3 (Nitro v5), 3200 Gbps per instance, 'up to 35% improvement in latency' vs earlier EFA (AWS P5 page) | ↗API |
| H100-SXM | p5.48xlarge (1-182 day block) | $5.19 | capacity-block · min: 1-182 day block | A | publ | EFA, 3200 Gbps per instance (AWS P5 page) | ↗vendor |
| H200-SXM | p5e.48xlarge (1-182 day block) | $5.97 | capacity-block · min: 1-182 day block | A | publ | EFAv3 (Nitro v5), 3200 Gbps per instance, 'up to 35% improvement in latency' vs earlier EFA (AWS P5 page) | ↗vendor |
| B200 | p6-b200.48xlarge (3yr) | $6.15 | savings-plan · min: 3yr | B | publ | EFAv4, 3.2 Tbps per instance (AWS P6 page) | ↗API |
| H200-SXM | p5en.48xlarge (1-182 day block) | $6.87 | capacity-block · min: 1-182 day block | A | publ | EFAv3 (Nitro v5), 3200 Gbps per instance, 'up to 35% improvement in latency' vs earlier EFA (AWS P5 page) | ↗vendor |
| H100-SXM | p5.48xlarge us-east-1 | $6.88 | on-demand | A | publ | EFA, 3200 Gbps per instance (AWS P5 page) | ↗API |
| H200-SXM | p5en.48xlarge us-east-1 | $7.91 | on-demand | A | publ | EFAv3 (Nitro v5), 3200 Gbps per instance, 'up to 35% improvement in latency' vs earlier EFA (AWS P5 page) | ↗API |
| GB200-NVL72 | u-p6e-gb200x72 (1-182 day block) | $10.58 | capacity-block · min: 1-182 day block | U | unkn | not published | ↗vendor |
| B200 | p6-b200.48xlarge (1yr) | $11.22 | savings-plan · min: 1yr | B | publ | EFAv4, 3.2 Tbps per instance (AWS P6 page) | ↗API |
| B200 | p6-b200.48xlarge (1-182 day block) | $12.36 | capacity-block · min: 1-182 day block | B | publ | EFAv4, 3.2 Tbps per instance (AWS P6 page) | ↗vendor |
| B200 | p6-b200.48xlarge us-east-1 | $14.24 | on-demand | A | publ | EFAv4, 3.2 Tbps per instance (AWS P6 page) | ↗API |
| B300 | p6-b300.48xlarge us-east-1 | $17.80 | on-demand | B | publ | EFAv4, 6.4 Tbps per instance (AWS P6 page) | ↗API |
Grades measure what sellers publish — deliverability of the written claim, not measured performance. No fleet holds a measured (Verified) badge yet; that tier arrives with the verification harness. Source classes: API = provider's live pricing API · vendor = provider's published page · mkt API = marketplace API · 3rd-party = aggregator (provider publishes no price on its own pages) · quote = quote-gated or "starting at", not a walk-up price. Quote-class prices are shown in tables but excluded from headline ranges, spreads, and fair-price bands.
Terms & trust (as published, 2026-09-14)
| Uptime SLA | 99.99% region-level; 99.5% instance-level — measured on: Amazon Compute SLA: 99.99% applies when instances run concurrently across 2+ Availability Zones in a region; 99.5% applies to each individual EC2 instance. One claim type per instance, credits cannot be stacked. SLA text does not carve out GPU instances — applies to EC2 generally. |
| SLA remedy | credits — 10% credit for uptime 99.0-99.99%, 30% for 95.0-98.99%, 100% below 95.0% (same tiers for both SLAs); instance-level SLA adds no charge for an instance unavailable >6 min in an hour. Claim-based, within two billing cycles, credits apply only to future charges. |
| Capacity guarantee | No published 'guarantee' wording. EC2 Capacity Blocks for ML lets you 'reserve' GPU instances (P6e-GB200/P6-B300/P6-B200/P5en/P5e/P5/P4d/Trn2/Trn1) up to 8 weeks ahead, up to 6 months, 1-64 instances (512 GPUs), dynamic supply/demand pricing. On-Demand Capacity Reservations 'help ensure that you always have access' — 'help ensure', not 'guarantee'; future-dated Capacity Reservations exclude the P family (Capacity Blocks is the P-family path). |
| Egress fees | 100 GB/month data transfer out to internet free (aggregated across AWS services/regions, excl. China/GovCloud); beyond that tiered per-GB, first tier $0.09/GB confirmed on AWS-controlled pages. Full tier table is JS-rendered on the pricing page and could not be captured directly. |
| Billing granularity | per-second with 60-second minimum for Linux (incl. GPU P-family); Capacity Blocks priced separately (dynamic, reservation-based) |
| Self-serve signup | partial — credit-card signup yes, but published default quota for 'Running On-Demand P instances' is 0 vCPUs (also 0 for G/VT, Inf, Trn, DL, F); GPU launch requires a Service Quotas increase request first |
| Security certifications | SOC 2, ISO 27001, HIPAA (via BAA), FedRAMP — all listed on aws.amazon.com/compliance/programs/ |
Notes: AWS never publishes the word 'guarantee' for GPU capacity. Egress tier table ($0.09/$0.085/$0.07/$0.05 per GB commonly cited) only partially verifiable on provider-controlled pages due to JS rendering; 100 GB free tier and $0.09/GB first tier confirmed.
Sources: ↗ ↗ ↗ ↗ ↗ ↗ ↗ · Fields marked "not published" are exactly that — the seller does not state them anywhere we could find. Absence is a finding.
Reliability record
AWS publishes no machine-readable incident history (status page: https://health.aws.amazon.com/health/status). We track every provider's public status feed weekly; a seller you cannot audit for downtime is a seller whose SLA you cannot price. Absence is a finding.
What AWS does not publish
- p5e.48xlarge on-demand price (no entry in the us-east-1 on-demand pricing feed; only p5/p5en/p6-b200/p6-b300 listed)
- P6e-GB200 (GB200 NVL72) on-demand list price — UltraServers not in the us-east-1 on-demand feed; sold via Capacity Blocks/UltraServer channels
- Numeric UltraCluster maximum for P6 family (page says 'tens of thousands of GPUs')
Unpublished fields are recorded as unknown and grade conservatively. Sellers improve their grades by disclosing — that is the point.
Weekly context on every provider
The Delivered Compute Report tracks how these prices and grades move.
Subscribe free