Your outage, priced by the contract

Every figure on this page was read from the provider's own service level agreement on 24 August 2026, or, where no such document exists, from the terms of service, subscription agreement or pricing page that stands in its place. Nothing here comes from a comparison article or a summary. Where a provider publishes several documents for one product family, each was treated as its own row, which is why sixteen providers produce twenty-six service levels.

The calculator does one piece of arithmetic: it turns your outage length into a monthly uptime percentage as one minus the outage divided by the minutes in the month, then matches that percentage against the rungs each provider publishes. That is deliberately the simplest possible reading of the contract, and it is more generous than most of these documents actually are. Nine specific things the calculator cannot model, all of which push the real credit down rather than up:

1. Minimum durations. Scaleway counts unavailability only from four continuous minutes. OVHcloud requires more than three consecutive minutes on VPS. Google Cloud discards intermittent downtime shorter than one minute entirely, so a flapping outage can register as zero.

2. When the clock starts. OVHcloud counts from the moment the incident ticket is opened, and DigitalOcean Spaces from the moment you give written notice. A 67-minute outage noticed after 40 minutes is formally 27 minutes at one and 27 minutes at the other.

3. What counts as an error. Cloudflare R2 counts only 500, 502, 503 and 504 responses and excludes throttled or rate-limited requests. AWS S3 averages per-five-minute error rates and treats an interval in which you made no requests as a zero percent error rate, which dilutes the average in the provider's favour. Google Cloud Storage discards repeated identical requests unless they follow a mandatory exponential backoff from one to thirty-two seconds.

4. The measurement window. Cloudflare R2 uses a rolling thirty days and bases the credit on fees in the thirty days preceding the claim, so the credit base depends on when you file. Azure measures pay-as-you-go services over the thirty days prior to and including the first day of the incident. Neither is a calendar month.

5. What the percentage is taken from. AWS excludes upfront Reserved Instance payments. Azure credits only the actual downtime rather than the outage duration, and only for the affected tier. Fastly multiplies the monthly usage charge, not the commitment.

6. Exclusions. Hetzner excludes required live migrations, DDoS and hacker attacks, and judges availability at the host, so an instance that booted and shows CPU activity counts as available even if your traffic cannot reach it. Vercel's excused downtime is subtracted from the denominator rather than counted as downtime, and includes the failure of the Internet in general. Google Cloud excludes errors caused by quotas its own system applied.

7. Conditions on the guarantee. OVHcloud monitors by ARP or ICMP ping, and a configuration that blocks it voids the SLA. The 99.99% on the Scale and High Grade dedicated ranges is conditional on the customer having enabled link aggregation. Fastly requires that the outage forced you to redirect traffic off the service, so an outage you waited out formally never happened.

8. One claim per period. Azure pays one credit per service per applicable period, so a month that breached several service levels forces a choice. AWS forbids combining a region-level and an instance-level EC2 claim for the same incident. Cloudflare forbids claiming under the R2 SLA and the enterprise support SLA for the same incident.

9. Currency and expiry. Credits are almost never money. Hetzner and Scaleway exclude cash explicitly, OVHcloud credits expire at the end of the following month, DigitalOcean credits expire after 90 days, and both AWS and DigitalOcean decline to issue credits under one dollar. OVHcloud's general terms also allow it to invoice you for diagnosis time if the fault turns out not to be theirs.

Two further limits are worth stating plainly. Enterprise customers routinely negotiate terms that differ from every document here, and none of those contracts are public, so this page describes the published floor rather than what a large account signed. And the arithmetic ignores the practical question of whether anyone at your company will notice the outage, gather the log files and file the claim inside the window. On the shorter deadlines in this set, that question decides the payout more often than the ladder does.