Cloud repatriation: data centre colocation sourcing guide

Workload selection is the foundation when evaluating an exit from public cloud infrastructure. According to AWS Prescriptive Guidance (2025), applications remain on premises “to take advantage of the low latency and local data processing of the on-premises environment, to avoid high data transfer costs, or for regulatory compliance”. Infrastructure planners must separate predictable baseline systems from variable compute needs.

Predictable systems offer a clear case for dedicated colocation. Core databases, distributed storage, and analytics engines run with continuous duty cycles. In cloud environments, these workloads incur steady compute charges, egress fees, and storage surcharges. Dedicated colocation operates with flat equipment amortisation and predictable power costs.

Conversely, irregular workloads rarely suit dedicated hosting. Moving short-term testing or seasonal jobs to owned hardware creates idle capacity. Sizing colocation requires isolating steady baselines while keeping hybrid links for workloads that need cloud autoscaling.

Questions for workload evaluation:

• Which applications run on continuous duty cycles rather than irregular spikes?
• What are current monthly egress charges across public cloud accounts?
• Do specific workloads require direct physical custody for compliance?
• Which applications require sub-millisecond local latency to storage?
• Can burst capacity stay in cloud instances while baseline workloads move?

Sizing physical infrastructure from cloud dashboards requires translating virtual metrics into bare-metal specifications. Cloud metrics can overstate hardware needs due to hypervisor overhead, overprovisioning, and fixed instance tiers.

Teams should audit 95th-percentile resource utilisation over several months. High core counts per socket allow large fleets of virtual instances to consolidate into compact two-socket servers.

Questions for hardware sizing:

• What is peak observed memory bandwidth across consolidated workloads?
• How much overhead is needed for host virtualisation management?
• What sustained storage IOPS are required alongside peak bursts?
• How much flash capacity is lost to parity and filesystem reserves?
• Will storage use dedicated arrays or software-defined nodes?
• How many spare nodes are needed on-site for hardware failures?
• What are manufacturer lead times for processor and storage orders?

Data centre white space is contracted on power availability rather than floor area. Sizing colocation deployments requires calculating actual power draw rather than power supply nameplate ratings.

Nameplate ratings state theoretical maximum draw. Sizing to nameplate values leads to contracting unneeded power capacity. Sizing should rely on thermal design power modelling, vendor calculators, or bench tests under production loads.

Dense servers concentrate thermal loads into single cabinets. Standard halls support lower power allocations per rack, forcing dense equipment to spread across extra cabinets. High-density halls use aisle containment or in-row cooling to manage heat. Floor load limits must also be checked to confirm that slabs and raised floors support fully populated racks.

Questions for cooling and power diligence:

• What is the maximum continuous kilowatt rating permitted per cabinet?
• What cooling design prevents hot exhaust air from recirculating into intakes?
• What intake air temperature and humidity bands are specified in the contract?
• Does the floor slab or raised floor support fully loaded cabinets?
• Can the facility support rear-door heat exchangers or direct liquid cooling later?

Migrating workloads out of cloud platforms shifts network dependency from software APIs to physical carrier circuits. External connectivity depends on the facility carrier ecosystem and meet-me room design. Carrier-neutral data centres let tenants buy transit, dark fibre, and private circuits from competing providers. Sourcing teams must verify physical entry diversity, ensuring fibre enters through separate, isolated conduits. Hybrid deployments connect colocation cabinets to public clouds through dedicated on-ramps or metro fibre. Private on-ramps bypass public internet transit to deliver predictable latency and lower egress fees. We match capacity buyers with carrier-neutral operators whose on-ramp ecosystems, meet-me rooms and fibre routes fit their connectivity and power requirements.

Questions for carrier and connectivity diligence:

• How many carriers maintain active points of presence in the building?
• How many physically separate meet-me rooms operate in the facility?
• Are cross connects routed along diverse pathways to meet-me rooms?
• Which cloud providers maintain native on-ramps inside the building?
• What are measured latency figures and route distances to target cloud regions?
• What are monthly cross-connect fees, and do they escalate annually?

Colocation agreements involve multi-year commitments, physical fit-out capital, and structured power billing that differ from public cloud subscriptions.

Operators package power through three main models:

• Reserved power: A fixed monthly fee for contracted capacity in kilowatts or kVA, regardless of consumption.
• Metered power: A baseline space fee plus actual electricity consumed, recorded by sub-metered units with a power usage effectiveness multiplier.
• Bundled power: A fixed monthly charge per cabinet bundling a set power allocation, space, and cooling.

Agreements should include power ramp schedules so billing aligns with staged server installation. Tenants should negotiate reservation options over adjacent cabinets to prevent fragmented footprints. Contracts must define dual-feed power availability, environmental thresholds, and service credit remedies for downtime.

Questions for commercial contract diligence:

• How are power charges calculated, and is power usage effectiveness billed at a fixed ratio or actual efficiency?
• What power ramp schedule applies during installation and migration phases?
• Can adjacent cabinet positions or expansion power be reserved, and what holding fees apply?
• What service credits apply if an electrical feed, temperature band, or cross connect fails?
• What notice periods and decommissioning terms apply at contract expiry?

A thorough facility evaluation reviews critical engineering systems before shortlisting candidate data centres. Technical teams should examine:

• Electrical topology: Does the facility use an N+1, 2N, or distributed redundant topology from substation to rack distribution units?
• Generator fuel: How many hours of generator fuel are stored on-site, and what are contracted delivery response times during emergencies?
• UPS systems: Are units static battery or rotary flywheel systems, and when were cells last tested?
• Fire suppression: What suppression systems are installed, and does discharge trigger an automatic electrical cut-off?
• Physical security: What access controls protect site perimeters, building entrances, data halls, and private cages?
• On-site support: Are qualified technicians present 24 hours a day for remote-hands tasks and visual inspections?
• Incident history: What is the facility record of unscheduled electrical switching, outages, or environmental excursions over the past five years?

When comparing shortlisted facilities, buyers should obtain written engineering confirmation across each of these seven areas alongside commercial lease terms.

Migration schedules depend on hardware lead times, carrier circuit delivery, and site preparation. Server, storage, and network procurement depends on component supply chains. Private cage fit-outs, structured cabling, and high-density power rails must be completed before servers arrive. Once hardware is powered, testing protocols confirm stability:

• Load bank testing: Portable resistive heaters test cooling airflow and breaker stability under full thermal load before servers are installed.
• Hardware stress testing: Synthetic stress tests across processors, memory, and drives identify component faults while warranties are active.
• Failover verification: Disconnecting primary network uplinks and power feeds confirms that automated routing and dual-cord supplies function properly.

Parallel running is an unavoidable cost during cutover. Organisations must budget for simultaneous cloud expenditure and colocation lease costs while synchronising data and validating production workloads.

Repatriating steady workloads only delivers the modelled savings if your colocation contract matches your actual kW draw, rack density, cloud on-ramp latency and power ramp schedule. Data Centre Axis sources colocation, wholesale and AI-ready capacity across our core markets.

What you get when you source through Data Centre Axis:

• Find capacity (/colocation/, /colocation/wholesale/): A screened shortlist of facilities matched to your kW or MW load, rack density, carrier and cloud on-ramp needs, and target cutover date.
• Operator proposals compared (/colocation/pricing/, /colocation/leasing/): Side-by-side comparison of metered versus reserved power pricing, PUE multipliers, cross-connect charges, power ramp schedules and expansion options.
• Find tenants (/colocation/offer-capacity/, /colocation/anchor-tenants/): For operators with live halls or upcoming phases, we introduce enterprise, cloud and AI customers.

Our coverage spans 115 facility profiles and 44 regional operator profiles across Australian data centres and UK data centres, 2,004 project signals tracked across Australia (215) and the UK (1,789), and colocation markets across the United States and Canada.