Provider record · updated 2026-09-10
Google Cloud — GPU pricing & deliverability
12 tracked offerings · grade mix A:0 · B:0 · C:12 · D:0 · U:0. Google publishes per-VM network bandwidth in the accelerator-optimized machines doc: a3-highgpu-8g 1,000 Gbps (5 NICs, gVNIC/GPUDirect-TCPX), a3-megagpu-8g 1,800 Gbps (9 NICs, GPUDirect-TCPXO), a3-ultragpu-8g 3,600 Gbps (10 NICs, RoCE CX-7), a4-highgpu-8g 3,600 Gbps (10 NICs, RoCE CX-7, four-way rail-aligned network), a4x-highgpu-4g 2,000 Gbps (6 NICs, RoCE CX-7) and A4X Max (GB300) up to 3,200 Gbps via CX-8 SuperNICs with eight-way rail-alignment. No numeric max-cluster GPU count published on the fetched pages.
Tracked offerings
| GPU | SKU | $/GPU-hr | Terms | Grade | Conf | Fabric (as published) | Src |
|---|---|---|---|---|---|---|---|
| H200-SXM | a3-ultragpu-8g (3yr) | $4.65 | committed · min: 3yr | C | publ | RDMA over Converged Ethernet (RoCE) via ConnectX-7, 3,600 Gbps max per VM, 10 NICs (Google doc table) | ↗vendor |
| H100-SXM | a3-highgpu-8g (3yr) | $4.86 | committed · min: 3yr | C | seco | gVNIC with GPUDirect-TCPX, 1,000 Gbps max per VM, 5 NICs (Google accelerator-optimized machines doc table) | ↗vendor |
| H100-SXM | a3-megagpu-8g (3yr) | $5.08 | committed · min: 3yr | C | seco | gVNIC with GPUDirect-TCPXO, 1,800 Gbps max per VM, 9 NICs — 'doubles the network bandwidth when compared to the A3 High' (Google doc) | ↗vendor |
| B200 | a4-highgpu-8g (3yr) | $7.09 | committed · min: 3yr | C | publ | RoCE via NVIDIA ConnectX-7 NICs on Google's four-way rail-aligned network, 3,600 Gbps max per VM, 10 NICs (Google doc) | ↗vendor |
| H200-SXM | a3-ultragpu-8g (1yr) | $7.31 | committed · min: 1yr | C | publ | RDMA over Converged Ethernet (RoCE) via ConnectX-7, 3,600 Gbps max per VM, 10 NICs (Google doc table) | ↗vendor |
| H100-SXM | a3-highgpu-8g (1yr) | $7.67 | committed · min: 1yr | C | seco | gVNIC with GPUDirect-TCPX, 1,000 Gbps max per VM, 5 NICs (Google accelerator-optimized machines doc table) | ↗vendor |
| H100-SXM | a3-megagpu-8g (1yr) | $8.03 | committed · min: 1yr | C | seco | gVNIC with GPUDirect-TCPXO, 1,800 Gbps max per VM, 9 NICs — 'doubles the network bandwidth when compared to the A3 High' (Google doc) | ↗vendor |
| H100-SXM | a3-highgpu-8g us-central1, us-east4 | $11.06 | on-demand | C | seco | gVNIC with GPUDirect-TCPX, 1,000 Gbps max per VM, 5 NICs (Google accelerator-optimized machines doc table) | ↗3rd-party |
| B200 | a4-highgpu-8g (1yr) | $11.12 | committed · min: 1yr | C | publ | RoCE via NVIDIA ConnectX-7 NICs on Google's four-way rail-aligned network, 3,600 Gbps max per VM, 10 NICs (Google doc) | ↗vendor |
| H100-SXM | a3-megagpu-8g us-east4 | $11.62 | on-demand | C | seco | gVNIC with GPUDirect-TCPXO, 1,800 Gbps max per VM, 9 NICs — 'doubles the network bandwidth when compared to the A3 High' (Google doc) | ↗3rd-party |
| H200-SXM | a3-ultragpu-8g asia-south1, asia-south2, europe-west1 +5 more | — | — | C | publ | RDMA over Converged Ethernet (RoCE) via ConnectX-7, 3,600 Gbps max per VM, 10 NICs (Google doc table) | ↗quote |
| B200 | a4-highgpu-8g asia-northeast1, asia-southeast1, europe-north1 +8 more | — | — | C | publ | RoCE via NVIDIA ConnectX-7 NICs on Google's four-way rail-aligned network, 3,600 Gbps max per VM, 10 NICs (Google doc) | ↗quote |
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% multi-zone; >=99.95% single instance Memory Optimized; >=99.9% single instance all other machine families (GPU A3/A2/G2 fall here); >=99.9% multi-zone on Standard network tier — measured on: Monthly Uptime Percentage per calendar month, per Project per Region (multi-zone) or per instance. Downtime = loss of external connectivity or persistent disk access in periods of >=1 consecutive minute. Spot VMs (the discounted GPU path) are explicitly excluded from the Compute Engine SLA. |
| SLA remedy | credits — Financial Credits on future bills, sole and exclusive remedy, request within 60 days. Multi-zone: 99.00-<99.99% = 10%, 95.00-<99.00% = 25%, <95% = 100%. Single instance (other families): 95.00-<99.90% = 10%, 90.00-<95.00% = 25%, <90% = 100%. |
| Capacity guarantee | Published AI Hypercomputer consumption options: standard reservations ('very high level of assurance for capacity'); future reservations for capacity blocks (1-year+ commitment for A4X Max/A4X/A4/A3-class, quota auto-increased); Dynamic Workload Scheduler calendar mode = reserved exclusive-access dense GPU capacity up to 90 days, no commitment, up to 53% discount; DWS flex-start = best-effort, not guaranteed. No financial remedy published for reservation non-delivery. |
| Egress fees | Premium Tier internet egress from US: ~1 GiB free, then $0.12/GiB (to 1 TiB), $0.11 (1-10 TiB), $0.08-0.085 above 10 TiB destination-dependent. Standard Tier: first 200 GiB/month free, then $0.085 (to 10 TiB), $0.065 (10-150 TiB), $0.045 (>150 TiB). Inbound free. |
| Billing granularity | per-second with 1-minute minimum — billing model page states this applies to 'all vCPUs, GPUs, and GB of memory' |
| Self-serve signup | partial — credit-card signup works, but GPU quota requires an approved quota increase request (per-model/per-region quotas); Free Trial billing accounts cannot request quota changes at all, so no GPUs on free trial |
| Security certifications | SOC 2 (SSAE 18 reports via Compliance Reports Manager), ISO 27001 (Compute Engine in scope), HIPAA (BAA supported), FedRAMP High P-ATO (requires Assured Workloads for FedRAMP High) |
Notes: SLA text has no GPU-specific exclusion but defers to per-product documentation exclusions; Spot/preemptible GPUs carry no SLA. DWS/capacity-block reservations are capacity commitments without a published uptime-SLA-style credit remedy. Mexico and Stockholm regions have lower multi-zone SLO (99.95%).
Sources: ↗ ↗ ↗ ↗ ↗ ↗ ↗ ↗ ↗ ↗ ↗ · Fields marked "not published" are exactly that — the seller does not state them anywhere we could find. Absence is a finding.
Reliability record (public status feed, last 90 days)
4 incidents on Google Cloud's own status page in the trailing 90 days, collected 2026-09-15. Self-reported by the provider — a floor on true incident count, not a ceiling; it still beats the war-story thread you were going to read instead. High counts often reflect granular per-rack reporting, not worse infrastructure: read the incidents, not just the number.
| Date | Incident | Impact | Duration |
|---|---|---|---|
| 2026-09-01 | Multiple products in us-central1-b are experiencing network service degradation. ↗ | medium | 4.1 h |
| 2026-08-20 | We are investigating an issue where customers may experience timeouts, service degradations, errors, and elevated latencies across multiple products in the us-w ↗ | high | 3.7 h |
| 2026-07-15 | Google Cloud VMware Engine (GCVE), Google Cloud NetApp Volumes, and Bare Metal Solutions (BMS) services are experiencing a service outage in europe-west4-a due ↗ | medium | 12.5 h |
| 2026-07-14 | Google Cloud VMware Engine (GCVE) Stretched Cluster customers are experiencing zonal outages impacting network connectivity across multiple regions. ↗ | medium | 10.7 h |
What Google Cloud does not publish
- On-demand prices for a3-ultragpu-8g (H200) and a4-highgpu-8g (B200) — Google's docs state these have no on-demand consumption option (Spot/Flex-start/reservations only)
- Official Google pricing-page dollar figures — cloud.google.com pricing pages are client-side rendered and returned no prices to fetch; a3-high/a3-mega prices are from the gcloud-compute.com Billing-data mirror (secondary)
- Maximum contiguous GPU cluster scale per machine series (not stated on fetched pages)
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