Before you order: the checklist that prevents regrets
An infrastructure order commits you for years, and the most expensive problems are the ones that only show after signature: a rack that cannot carry the promised power, a cross-connect billed for life, a recovery plan never tested. This checklist gathers the questions to ask, item by item, before signing anything.
Estimate my cost →Power: the question that eliminates sites #
What maximum power per rack does this site actually sell? That is the deciding question, especially once GPUs are involved. Require two separate feeds (A and B), each able to carry the full load, and clarify billing: subscribed power or metered energy, and is cooling included in the rate. Finally, subscribe with a margin: at the exact measurement, the first peak trips the breaker.
Hardware: the requirements you cannot negotiate later #
For a virtualization cluster: server-grade drives with power-loss write protection (the marking to look for is PLP), ECC memory, and never a hardware RAID controller under storage that manages redundancy itself (ZFS or Ceph). Exact part numbers and compatibility are confirmed on the manufacturer's configurator at quote time: catalogues change every quarter.
Network: the hidden recurring cost #
The rack rent is the most visible and most negotiated item; cross-connects, the cables linking you to carriers, are billed monthly and add up year after year. Before signing: which carriers are actually present on site, at what cross-connect price, and is a single link enough or do you need redundancy? Our interconnection guide details this item.
Resilience: RPO/RTO understood, not recited #
Two numbers summarise your recovery plan: RPO (how much data you accept to lose) and RTO (how long you accept to wait for recovery). Choose them knowingly, from a simple backup set to a dual active site, and remember that an announced RTO is only an intention: only a failover test, scheduled before going live, turns it into a fact. Our BCP/DRP guide lays out the definitions.
Backup: outside the cluster, outside the site #
Storage replication copies mistakes as faithfully as data. Backup lives on a separate machine, with an off-site copy (the 3-2-1 rule: three copies, two media, one off-site), and a dated restore test. A backup never restored is a hope, not a protection.
Contract: read the exit before the entrance #
Renewal and exit terms are negotiated before signature, never after: rent indexation, notice periods, what happens to hardware and data at the end, reversibility. Our renewal invoice guide shows where the increases hide. And for every number in an offer, the same requirement: its source and its date, otherwise it is a verbal promise.
Generate the checklist for your project #
The Digital Architect applies these checks to your project: it sizes the cluster, computes the requirement sheet (space, power, cooling) and outputs the pre-filled checklist, each point marked covered, to demand, or to schedule.
FAQ #
The maximum power actually sold per rack, with two A and B feeds each able to carry the full load. That constraint eliminates sites long before floor space does.
Three copies of your data, on two different media, one of them off-site. It protects against what replication does not cover: deletion, ransomware, human error.
RPO is the amount of data you accept to lose (the time since the last copy); RTO is the time you accept to wait before recovery. Both are chosen before the incident, and proven by a test.
Because a number without a source or date can be neither challenged nor compared: it is a verbal promise with a layout. Require the provenance of every value in an offer, price, lead time or performance.
Written on 10 September 2026.
From reading to comparing: relevant data centers
Estimate my cost → Compare data centers