Every hosting sales page says something like "tier III facility, N+1 power, generator backup". Buyers nod and move on, because almost nobody knows what those words commit to. When we place customer servers in UK datacentres, the power chain is one of the first things we interrogate, because power incidents are quiet, rare and catastrophic when they land. Here is what the terminology actually means, layer by layer, and what it does not cover.
Power reaches your machine through a chain: the utility grid feed, a UPS stage built from batteries and inverters, one or more diesel generators, and finally the power distribution inside the hall, which means PDUs feeding rack-level strips. A facility can be strong at one layer and weak at another. A site with an excellent generator but a single utility feed and a half-empty battery string is one substation fault away from an outage that the generator cannot bridge. So evaluate the whole chain, not the badge.
N+1 means: the facility needs N units of something to run at full load, and it has one spare. If a datacentre needs two UPS modules to carry the load, N+1 means three, so any single module can fail or be pulled for maintenance without dropping customers. The same arithmetic applies to generators, cooling chillers and distribution paths. The catch is capacity discipline. An N+1 design only stays N+1 while the load stays below what the remaining units can carry. A facility that slowly loads its UPS to 92 per cent has quietly converted N+1 into N, because a unit failure now trips the rest onto overload. When you tour a facility or read its uptime claims, the honest question is not "is it N+1?" but "what is the actual utilisation today?". Ask for the number. Good operators know it.
The UPS exists for exactly one job: hold the load while the generators start. A healthy battery string gives roughly five to fifteen minutes at load, and a diesel generator under an automatic transfer switch typically reaches stable output in ten to thirty seconds. That gap is what the batteries are sized for, usually with margin. What the UPS does not do is protect you from battery neglect. VRLA batteries age faster in hot halls, a weak cell drags its whole string, and a UPS that has not been discharge-tested in two years is a hope rather than a system. Facilities that publish quarterly battery test results are telling you something; facilities that go quiet on the question are telling you something too.
Diesel generators will happily run for days, so the binding constraints are fuel on site and the refill contract. A typical tank holds 24 to 72 hours at full load. After that the facility depends on a fuel supply agreement, and in a regional event, like the storms that have knocked out chunks of the UK grid in past winters, every datacentre in the county is calling the same handful of tanker companies at once. Ask two things: how many hours of fuel at current load, and whether the refill contract is a firm commitment or best-efforts. The second question is where the marketing usually goes soft.
The other factor is exercise versus load. Many facilities run generators weekly with no building load attached, which proves the engine starts and almost nothing else. A monthly or quarterly load-bank test, where the generator actually carries the facility or a dummy bank at rated capacity, is what finds failing injectors and clogged fuel systems before a real event does. Prefer facilities that do it.
All the upstream redundancy in the world ends at your power supply. An A+B design means each rack is fed by two independent distribution paths, and a server with dual PSUs, one on each feed, survives a whole path failing, maintenance included. A single-PSU server on the same floor survives nothing, because its entire chain, however redundant upstream, converges into one cord. If your workload justifies dedicated hardware, dual PSU is one of the cheapest uptime purchases available, and it pairs naturally with the out-of-band access we have argued for before: IPMI and KVM over IP let you act on a power event from home instead of driving to the facility.
One subtlety: two feeds only count as A+B if they are genuinely independent, sourced from separate PDUs, separate UPS systems and separate generator inputs. Some budget offerings sell "dual feed" that converges into the same UPS upstream. Ask for a one-line diagram. Serious operators have one ready.
The Uptime Institute tier vocabulary gets borrowed loosely in marketing. In its strict form, Tier III means concurrent maintainability: every element can be taken offline for maintenance without dropping load, which in practice means dual paths and N+1 capacity. Tier IV adds fault tolerance. But many sales pages use "tier III-style" or quote the building's original design rather than its current operated state, and the certification belongs to specific rooms and contracts, not to the whole business park. Treat the label as a starting question, not an answer. The same scepticism belongs in your provider shortlist generally; our guide to choosing a UK dedicated server provider covers the wider set of claims worth pressure-testing.
Planned work causes more power incidents than equipment failure. UPS battery replacements, PDU work and generator transfers all put the facility into states where redundancy is temporarily reduced, and that is exactly when the second surprise arrives. Ask any operator for their maintenance notification policy: how much notice, whether work is scheduled inside low-risk windows, and whether your server's feed is affected. Facilities that publish planned maintenance calendars are easier to trust than those that discover their own outages alongside you.
First, what is the current utilisation of the UPS and generator capacity relative to the N+1 design? Second, how many hours of fuel at present load, and is refill contracted firm? Third, are generators load-bank tested, and how often? Fourth, are the A+B feeds in a rack electrically independent end to end, and can I see the one-line diagram? Fifth, how much notice do customers get before power-chain maintenance? A facility that answers all five with numbers is rare and worth paying for. One that answers with adjectives is telling you where its investment went, and it was not the power room.
Power redundancy at the facility layer is bought in bulk and shared by everyone in the hall. Your personal exposure is narrower and cheaper to fix: dual power supplies on independent feeds, out-of-band access so you can see and reboot through a brownout, and backups that leave the building entirely, because no power design, however good, substitutes for a real off-machine backup. Get those three right and the datacentre's power chain becomes someone else's well-managed risk instead of your single point of failure.