Vivian Voss

Your SLA Is a Refund Policy

it philosophy cloud sla resilience

The credit note arrives in the middle of June, a single line on the cloud invoice of an ordinary online shop. The shop ran its servers in one availability zone of Amazon's region us-east-1, and on 7 May that zone lost power in a thermal event; the servers "automatically shut down when the temperatures exceeded the operating thresholds", and the cooling was back to full capacity twenty-one and a half hours later. The compute bill for May came to 8,000 euros. The credit is 30 per cent of it: 2,400 euros, applied to a future bill, provided somebody asked in time. On a normal day the shop takes about 40,000 euros. The finance team read the line twice, which is about once more than it deserved.

Nobody in that company did anything foolish. They bought from the largest provider in the world, they read that its compute service is built for 99.99 per cent, and they took the number for a description of their own uptime. It describes something else, and the contract says so in plain English to anyone who reads the small print.

What the four nines buy

Amazon's service level agreement for EC2 promises a monthly uptime of 99.99 per cent for a region and 99.5 per cent for a single instance. Below that line it pays in three steps: 10 per cent of the monthly bill under 99.99, 30 per cent under 99.0, and the whole month only below 95. The customer has to claim it through the support centre, "by the end of the second billing cycle after which the incident occurred", and a credit under one dollar is not paid at all, which spares both sides the paperwork. Then comes the sentence that matters: "this SLA sets forth your sole and exclusive remedies".

Four nines sound like reliability until somebody converts them into minutes. 99.99 per cent of a month leaves 4.32 minutes of downtime. The full refund begins below 95 per cent, which is thirty-six hours. Anything between those two marks, from a tea break to a day and a half, is settled at 10 or 30 per cent of one month's bill.

Google's contract for Compute Engine is built the same way and was last revised on 14 October 2026. Instances in several zones get 99.99 per cent, the credits follow the same steps (with 25 per cent in the middle, where Amazon pays 30) and go "to Customer's future monthly bills", the claim must come within 60 days, and the total in a month can never exceed what the customer paid. It, too, "states Customer's sole and exclusive remedy". Cloudflare goes furthest on paper. Its business plan promises to serve content "100% of the time without qualification", and then pays back the outage minutes in proportion to the customers affected, capped at one month's fee in any twelve months, provided support hears about it "within five business days".

Read side by side, the three documents have one subject: how much the provider owes when it fails, and how that amount is kept small. The customer's revenue appears in none of them, and neither do its data, its staff or its own customers. That is no oversight. A contract protects the party that drafted it, and these were drafted with some care.

What three published contracts promise, and what they pay Amazon EC2 Google Compute Cloudflare Promise 99.99 % per region 99.99 % multi-zone 100 % Credit 10 / 30 / 100 % of a month 10 / 25 / 100 % of a month outage minutes Claim by second billing cycle within 60 days notice in 5 days Limit nothing under 1 dollar at most the month's bill 1 month in 12 Remedy sole and exclusive sole and exclusive sole and exclusive Lost revenue not mentioned not mentioned not mentioned Published SLAs of Amazon EC2, Google Compute Engine and Cloudflare Business, read 19 November 2026

The year so far

2026 has given those clauses plenty of work. On 7 February an onsite transformer in Microsoft's West US region failed; the generators started, but "a cascading failure within a control system prevented the automated transfer", the batteries ran down, and customers in the region lost their compute and their storage for 20 hours and 26 minutes. In May the thermal event in us-east-1 followed. On 29 May "a severe thunderstorm produced lightning strikes" near West US 2, and the cooling plant, sensing unstable power, went into a protective "lockout" by design. As the halls warmed up, the servers switched themselves off. Two zones were hit, and the last services came back after 22 hours and 6 minutes. In July a voltage drop of three milliseconds in the Netherlands took out a Google zone for 14 hours and 55 minutes, after the backup supply "failed to take over the facility load" and the data halls reached 44 °C. And last week, from 16:20 UTC on 12 November until 16:10 the next day, Microsoft 365 was down, after a latent defect from a July update met a failover from Teams and a gateway fleet with nearly 14 per cent of its servers already out of service. Microsoft's preliminary report calls it a "global outage for the networking gateway".

In March drone strikes hit Amazon's facilities in the United Arab Emirates and damaged one in Bahrain. Amazon advised its customers to "consider taking action now to backup data and potentially migrate your workloads to alternate AWS Regions". On 15 September it said it was "unable to restore access to the resources and data hosted exclusively" in one zone of the UAE region and in the whole of the Bahrain region. Eight months on, both regions are still listed as disrupted.

A tracker kept by Control Plane, which counts the outages the four large providers document themselves, has 19 major ones for 2026 so far against 16 for the whole of 2025, and five of its twelve longest since 2020 belong to this year. The Uptime Institute, looking at data centres in general, finds that outages per site have become rarer for five years running, whilst the bill keeps rising: 57 per cent of operators put their last major outage above 100,000 dollars, one in five above a million, and power remains the leading cause.

Put the year into the contracts and the arithmetic is short. Twenty hours in February leave 96.96 per cent of the month, twenty-two in May 97.03, and Microsoft's twenty-four hours, near enough, leave November at 96.69. Each lands in the 30 per cent band of a contract of Amazon's type and well clear of the 95 per cent that would return the whole month. There is a further wrinkle. Google's contract carries the same exclusion as Amazon's, which leaves out outages "caused by factors outside of our reasonable control, including any force majeure event". Lightning, a failing transformer, a three-millisecond dip in the grid and a war are the very causes of this year's longest outages, and they are exactly the sort of thing that clause was written for. Whether a provider invokes it is another matter, of course. The contract leaves it the option.

Credit against downtime, Amazon EC2, region level, 30-day month 0 % 10 % 30 % 100 % 0 7.2 h 20 h 36 h hours of downtime in the month 2026's long outages, 20 to 24 hours, on this scale 99.99 % = 4.32 minutes full month Own calculation from the published EC2 SLA and the providers' incident reports, 19 November 2026

The attacker's arithmetic

Professional attackers keep accounts as well, and theirs run the other way round. A first-rate way in is expensive. In 2024 the exploit broker Crowdfense offered up to seven million dollars for a single zero-click chain into the iPhone, and whoever buys at that price expects a return. The return comes where one entry opens many doors, which is to say on a platform. In 2024 a group walked into Snowflake customer accounts with passwords stolen elsewhere. Mandiant and Snowflake ended up notifying "approximately 165 potentially exposed organizations". On 11 November this year attackers got into the ScreenConnect remote-access environments of managed service providers, and by that evening Huntress was tracking "6 compromised MSPs and 286+ endpoints", spread over 47 organisations. Each of the 47 came as part of a bulk purchase.

One way in, and what it reaches attacker platform ScreenConnect, 11 November 2026: 6 MSPs, 47 organisations Snowflake, 2024: about 165 organisations notified attacker own servers the same effort reaches one firm Huntress via BleepingComputer, Mandiant; schematic

A firm that administers its own servers, whether in its own rack or on plain rented machines, spoils that calculation. Getting in costs the same effort as anywhere else and yields one firm, which puts it outside the budget of every group that plans in millions. The hyperscaler's customer, meanwhile, pays with every invoice for a defence built against exactly those groups, and receives in return a place in the queue on the day the defence fails. A large provider adds its own attack surface to yours, and its contract leaves the cost of the day it fails with you. The SLA has nothing to say about any of this, because it measures uptime, and a mailbox read by a stranger is, technically, up.

The strongest case for the contract

The providers have a good answer, and it deserves its full weight. No company can insure the turnover of millions of customers for the price of a server, and none pretends to. Amazon's own guidance divides the work cleanly: the provider is responsible for the "resiliency of the cloud", the customer for resiliency in it, and "customers that deploy Amazon EC2 instances are responsible for deploying Amazon EC2 instances across multiple locations". The same page names the threat: spreading a workload across zones protects it from "power outages, lightning strikes, tornadoes". The shop in the opening scene, on this view, built its own outage, and two zones would have carried it through May.

All of that is true, and it settles the matter in an unexpected direction. If availability is the customer's job, then the SLA was never evidence of it, and a company that puts four nines into its risk register or a tender has filed a refund policy under resilience. The year also shows where the customer's job ends. The lightning in West US 2 reached two zones at once, the damage in Bahrain spread across several, and the only safe place for data in the Gulf turned out to be another region entirely. Real availability means a second region or a second provider, with copies of the data that somebody has actually tested, and a switch-over that has been rehearsed. The customer pays for all of it, out of pocket, and the SLA contributes a share of a monthly bill.

Owning that is less exotic than it sounds. Two machines at two providers in two countries, a ZFS snapshot sent from one to the other every few minutes, and a DNS change practised once a quarter: on FreeBSD that is a periodic job and a page of notes, and it survives a thunderstorm in Washington State without anyone filling in a claim form.

The limit

The large clouds are good at this, on average. Most server rooms in most companies would not reach the uptime the big providers achieve in an ordinary month, and the Uptime Institute's falling outage rate per site includes their work. The argument here concerns what the contract promises, which is a separate question from how often the service runs.

Cheap attacks reach everyone. The same organised groups also run automated campaigns that cost them next to nothing per victim, and Verizon's breach report for 2025 found ransomware in 88 per cent of the breaches at small and medium businesses, against 39 per cent at large organisations. Those campaigns live on known holes and stolen passwords, on a rented server as much as in a cloud account, so the self-administered firm leaves the expensive attacks behind and keeps the ordinary care: patches on time, a key for every login, a backup out of the machine's reach and somebody who reads the nightly mail.

Large customers negotiate. An enterprise agreement can carry better terms than the published ones, and the figures above are the published ones. The shape of the remedy, a credit against future bills, tends to survive the negotiation.

Insurance can fill part of the gap. A business interruption policy may pay for a lost day; whether it pays for somebody else's lost region is a question for its own small print.

Two regions cost money. A small business may reasonably decide that a day of downtime every few years is cheaper than a second site. That is a fair decision, as long as it is taken knowingly, with the credit note priced in at 2,400 euros and the outage at what it actually costs.

An SLA tells you what the outage will cost the provider, which is the one figure nobody in your company asked for.