GPU Instance SLAs: What AWS, Azure, GCP and DigitalOcean Pay When Your Training Run Dies
October 19, 2026
Your $4.41-per-hour H100 has a service level agreement. When the node dies mid-training, that SLA pays you a credit. GPU SLAs measure connectivity on the node that failed, not the success of your job, and the payout is a percentage of that node's monthly bill. A dead NCCL ring stalls all 64 GPUs in your cluster, but only the failed node accrues downtime. Here are the commitments, tiers, deadlines and money math for all four providers.
The mismatch: SLAs count connectivity, not checkpoints
Every provider defines downtime for GPU instances the same way they do for any VM: loss of external connectivity. Google counts "loss of external connectivity or persistent disk access," AWS counts minutes with "no external connectivity," and DigitalOcean counts 5-minute intervals where a node has no external connectivity.
None of them count a hung NCCL collective, a corrupted checkpoint, or a job that stalls because one GPU in the ring dropped. If your instance still answers ping, the SLA clock does not run.
The credit is a percentage of the affected node's monthly bill, not your total spend and not your lost compute. A 10% credit on a $3,000 node is $300. The same outage on a 64-GPU job costs you 64 nodes of idle time plus a rerun.
What each provider actually promises
GPU instances inherit the compute SLA of the platform they run on, with one exception: DigitalOcean wrote a dedicated GPU Droplet SLA in October 2025.
| Provider | Covered object | Monthly commitment | Credit tiers |
|---|---|---|---|
| AWS EC2 | Region-level (all instances across 2+ AZs) | 99.99% | 10% below 99.99% · 30% below 99.0% · 100% below 95.0% |
| AWS EC2 | Single instance | 99.5% | 10% below 99.5% · 30% below 99.0% · 100% below 95.0% |
| Azure VM | 2+ instances across 2+ availability zones | 99.99% | 10% below 99.99% · 25% below 99.0% · 100% below 95.0% |
| Azure VM | 2+ instances in an availability set | 99.95% | 10% below 99.95% · 25% below 99.0% · 100% below 95.0% |
| Azure VM | Single instance, premium disk | 99.9% | 10% below 99.9% · 25% below 99.0% · 100% below 95.0% |
| GCP Compute Engine | Instances in multiple zones | 99.99% | 10% below 99.99% · 25% below 99.0% · 100% below 95.0% |
| GCP Compute Engine | Single instance | 99.9% | 10% below 99.9% · 25% below 95.0% · 100% below 90.0% |
| DigitalOcean GPU Droplet | Individual server node | 99% | 10% below 99% · 25% below 95% · 50% below 90% |
Three things stand out. AWS's single-instance commitment of 99.5% is the loosest of the big three: it allows 3.6 hours of downtime per 30-day month before a credit triggers at all. DigitalOcean's 99% node-level commitment is looser still, and it is the only one of the four that measures at the node level rather than the cluster or fleet level.
The single most important fact about GPU SLAs: the credit pays on the node that failed, not the job that died. A dead NCCL ring stalls 64 GPUs, but only the failed node accrues downtime, and the payout is a percentage of that node's monthly bill.
The money math: what an outage actually pays
DigitalOcean's numbers are published and simple. A single H100 GPU Droplet on-demand costs $4.41 per GPU-hour, so a node run flat for a 672-hour month bills $2,963.52. Run the outage math against the 43,200 minutes in a 30-day month:
| Downtime | Monthly uptime | DO credit tier | Credit on 1x H100 node |
|---|---|---|---|
| 4 hours | 99.44% | 10% | $296.35 |
| 8 hours | 98.89% | 10% | $296.35 |
| 24 hours | 96.67% | 10% | $296.35 |
| 48 hours | 93.33% | 25% | $740.88 |
| 72 hours | 90.00% | 25% | $740.88 |
| 90 hours | 87.50% | 50% | $1,481.76 |
Notice what the table hides: a 4-hour outage and a 24-hour outage pay the same $296.35, because DigitalOcean's 10% band runs from 95% to 99% uptime. You need nearly four days of node downtime to reach the 25% tier. On AWS, a single instance needs 3.6 hours of downtime to breach 99.5% at all, then 36 hours in a 30-day month to cross 95% and jump to the 100% tier. On GCP, a single instance pays 10% anywhere between 90% and 99.9% uptime.
The percentages apply to whatever you actually pay for the affected node. A $20,000-per-month p5-class node on AWS earns $2,000 at the 10% tier; a $3,000 H100 on DigitalOcean earns $296. The credit is always a fraction of what the dead run cost you in idle fleet time and rerun hours.
The traps that kill GPU claims
The single-node rule. DigitalOcean's SLA measures uptime "at the server node level, rather than at the cluster level." AWS's instance-level SLA counts only the instance that lost connectivity. If one node in an 8-node job fails and the other seven sit idle at a barrier, only the failed node earns downtime.
The 24-hour notice trap. DigitalOcean requires you to notify it in writing within 24 hours of downtime starting, or you forfeit the credit entirely. The request can come within three months of the billing cycle, but the notice clock is 24 hours. If your monitoring only tells you about the outage on Monday, the claim is already dead.
The 99.5% floor on AWS. A single GPU instance on AWS needs 3.6 hours of downtime in a month to breach the instance-level SLA. A 2-hour blip pays nothing. The region-level 99.99% claim only triggers when all of your instances across two or more AZs lose connectivity at once, and you cannot stack region and instance claims.
The disk tier on Azure. A single GPU VM only carries the 99.9% commitment if it runs on premium storage. Standard SSD drops the single-VM commitment to 99.5%, and standard HDD to 95%. GPU VMs usually ship with premium disks, but check the disk SKU before you assume the tier.
Capacity is not downtime. You cannot claim for a GPU you could not launch. Google's Compute SLA excludes errors "that resulted from quotas applied by the system," and the newest Blackwell machines are not sold on-demand at all: A4X, A4X Max, A4 and A3 Ultra are reservation-only through AI Hypercomputer. AWS documents InsufficientInstanceCapacity as a standard launch failure. Out of stock is not an outage.
Upfront payments shrink the base. AWS calculates credits on the monthly bill "excluding one-time payments such as upfront payments made for Reserved Instances." GPU fleets are commonly bought with 1- or 3-year reservations and large upfront payments, so a reserved node earns credits only on the recurring portion of its bill.
The claim clocks
The deadlines are short. AWS wants the credit request by the end of the second billing cycle after the incident, with "Amazon Compute SLA Credit Request - Instance-Level Claim" in the subject line, your instance IDs, and request logs. Azure wants the claim within 60 days of the incident, filed as a billing support request. Google wants notification within 60 days of eligibility, with log files showing the downtime periods. DigitalOcean wants written notice within 24 hours of the outage and the credit request within three months of the billing cycle, and the credit expires 90 days after it is issued. AWS also waives the charge for any single instance unavailable more than six minutes in a clock hour, automatically.
Before your next training run, look up which SLA tier each GPU node sits in: single instance or multi-AZ, premium disk or not, on-demand or reserved. Write the claim window into your calendar the day the run starts. The credit is small, but it only exists if someone files inside the window.