Two cloud outage stories landed on the same day this week. On Wednesday, November 11, Amazon Web Services worked through a five-and-a-half-hour incident in one of its Tokyo availability zones. The same day, Google published the root cause of the October 20 outage that took down a zone of its London region for about six hours, hitting banks, courts and couriers. Both are reminders that even the largest clouds fail. They also raise a question most teams only ask afterward: when it happens, what does your provider actually owe you?

Tokyo: launches failing, running instances “not affected”

According to AWS status updates summarized by UptimeRobot, the incident began at 15:40 UTC and lasted 5 hours 33 minutes. It started with elevated API error rates and launch failures in a single zone, apne1-az1, and spread to ECS and Fargate task launches, load balancer provisioning, and NAT Gateway workflows. CloudWatch health data for EC2 and EBS in that zone moved to “Insufficient Data.”

AWS told customers that running instances were not affected, and that they should rely on their own health telemetry rather than the DescribeInstanceStatus signal during the window. Later, a second related problem affected connectivity for newly created or modified instances in the same zone, and AWS advised launching new instances in the two healthy zones instead. AWS had not published a public root cause at the time of writing.

In other words: if you were already running, you may have been fine. If you needed to launch, scale, fail over or provision anything in that zone, you were stuck, and the status signal you would normally check was itself unreliable.

London: one sensing error, six hours dark

Computer Weekly reports that Google traced its October 20 London outage to a programmable logic controller that measured mains voltage on the wrong side of a circuit breaker. The site hosting europe-west2-b was running on a single utility feed during maintenance when that feed failed. Because the controller still saw voltage on an isolated cable, it never told the diesel generators to start. The UPS batteries, designed to bridge about five minutes, ran flat and the data halls lost power.

Google lists more than 35 Cloud services affected in europe-west2. Reports at the time named Lloyds, HSBC, Halifax and Revolut, alongside HM Courts & Tribunals Service video hearings and Royal Mail. Failover traffic then saturated the two surviving zones, so even some customers running multi-zone architectures felt the impact. Google has since reconfigured the controller logic and is auditing similar facilities.

What the SLA actually says

Every major cloud publishes a service level agreement, and they are real commitments. But it is worth reading what the remedy is. Under the Amazon Compute SLA, if EC2 misses its uptime target, you become eligible for a service credit of 10%, 30% or 100% of that month’s EC2 bill in the affected region (or for the single affected instance), depending on how far uptime fell. Credits are applied against future EC2 payments, and the SLA states plainly that they do not entitle you to a refund. AWS also automatically does not charge for a single instance that is unavailable for more than six minutes of a clock hour.

Google’s Compute Engine SLA works the same way: credits of 10%, 25% or 100% of the monthly bill for the affected service in the region, applied to future bills, capped at what you owed for that service that month.

Both put the work on the customer. AWS requires a claim through its Support Center by the end of the second billing cycle after the incident, with your own request logs documenting the errors. Google requires you to notify support within 60 days, with log files showing the downtime. Miss the deadline or the evidence, and you forfeit the credit.

Both SLAs describe the credit as your sole and exclusive remedy. The AWS Customer Agreement goes further and excludes liability for lost profits, revenues, customers, opportunities or goodwill, and for unavailability of the services, apart from SLA credits. And definitions matter. The EC2 SLA counts an instance as unavailable when it has no external connectivity. In an incident where running instances stay connected but you cannot launch or scale, the outage may not count toward the SLA at all.

Put simply, the hyperscaler SLA amounts to: we won’t bill you for the downtime. It does not cover what the downtime cost your business.

What you want during an outage

Outages can happen to any provider. No honest host will tell you otherwise. The question that matters when it happens is who picks up, what they can see, and whether they own the machines your business runs on.

At InnoScale we own our hardware and run our own datacenters in the United States and in Europe. When something breaks, you reach real engineers, any hour, not a status page and a claim form. We have been independent for twenty-five years.

Plan for the bad day, not the credit

This week’s two stories are a useful exercise. Read your provider’s SLA, then write down what an hour of downtime actually costs you, and compare. If the gap is large, the answer is not a better credit. It is knowing who you can call, and having an environment you understand, on hardware someone is accountable for.

If you want to talk about running workloads on hardware we own, with engineers who answer, contact the InnoScale team. We will keep the conversation practical.