At 18:40 on Friday, an operations lead studies a dashboard after a bad week. The order gateway stayed up, but p99 response time widened during the open, a cloud bill jumped after a replay job, and a cross-connect fault took hours to isolate. The dashboard shows symptoms. It does not show whether the infrastructure model caused them, or whether the team measured the wrong things.
Colocation and cloud are not opposing answers to one question. They package power, compute, network access and operational responsibility differently.
A broker or prop firm comparing them needs a workload definition, a cost boundary and measurements taken on the same route to the same endpoint. Without those, a low server quote or a latency figure from a provider presentation is not a decision.
Compare the same service, not two invoices
Start by drawing the production path: market data source, feed handler, strategy or risk service, order gateway, broker or exchange endpoint, then persistence and monitoring. Mark which components require deterministic performance, which can tolerate a restart, and where data crosses a paid network boundary. A matching, risk or gateway service with strict tail-latency objectives has different requirements from a reporting cluster that runs overnight.
Colocation usually means leased rack space and power in a data centre selected for network access. The firm supplies or leases servers, arranges cross-connects and maintains the operating system and application. Cloud usually means virtual machines, managed services or bare-metal instances in a provider region. The provider runs the facility, but the customer still owns configuration, capacity planning, security and the behaviour of its software. Neither model removes operations work.
Choose a service boundary before comparing prices. Include the same number of production and standby nodes, the same retention period, backup policy, traffic volume and support coverage. Count development and replay workloads separately if they can be scheduled away from the trading path. For cloud, include block storage, snapshots, load balancers, private connectivity, logging and data transfer. For colocation, include hardware amortisation, cross-connects, carrier services, spares and remote hands.
Trading Technology's infrastructure coverage tracks the platforms and providers behind broker operations. For a procurement comparison, the useful unit is not a server or rack. It is the cost and performance of a defined service at a defined level of availability.
Build the monthly cost from its parts
The figures below are typical commercial market ranges in US dollars for a modest production footprint of roughly four to eight compute hosts, with a primary site or cloud region and ordinary production support. They are planning ranges, not quotes. Location, power density, contract term, provider discounts, network routes and exchange access can move costs materially. The ranges exclude staff, software licences, market-data fees and exchange or broker port charges.
| Cost component | Colocation, typical monthly range | Cloud, typical monthly range | What changes the figure |
|---|---|---|---|
| Facility and power | $2,000-$8,000 for a 5-10 kW allocation | Included in service pricing | Metro, committed power, density and contract term |
| Compute | $1,000-$5,000 amortised hardware cost | $2,000-$12,000 for always-on instances or dedicated capacity | CPU generation, memory, licensing and utilisation |
| Storage and backup | $200-$1,000 | $300-$2,500 | Performance tier, snapshots, retention and restore tests |
| Network connections | $300-$2,000 for two to four cross-connects | $300-$2,500 for private links, ports and related services | Connection count, carrier, region and bandwidth |
| Traffic and egress | $100-$1,000 | $500-$5,000 or more | Feed distribution, replication, replay and outbound volume |
| Hands-on support | $200-$1,500 | $300-$2,000 for managed services or support allocation | Coverage hours, response time and task volume |
| Indicative monthly total | $4,000-$18,500 | $3,700-$24,000 or more | Scope and accounting treatment must match |
Hardware amortisation is often missed in a colo spreadsheet. A server bought for $12,000 and kept for 36 months represents about $333 a month before spares, installation and financing. Conversely, a cloud estimate based on an on-demand calculator can overstate steady-state cost if the workload qualifies for a commitment, or understate it if it omits snapshots, inter-zone traffic and egress.
The table is deliberately broad. A rack near a venue with specialised access can cost far more than a general-purpose cabinet, while a cloud estate with steady utilisation may secure lower compute rates.
Cross-connect pricing and cloud private-link charges are not equivalent products. Ask vendors to itemise recurring port fees, installation charges, traffic billing and minimum commitments. A useful comparison has a base case and a peak-month case, not just a single monthly total.
Latency depends on the route and the tail
Colocation can shorten the physical and network path to a market or execution venue. It does not guarantee a fast order acknowledgement.
A cross-connect may lead to a carrier hand-off, a managed gateway or another network provider before reaching the destination. Each hop, queue and protocol layer can add delay or variability. A server in the right building but on the wrong route can perform worse than a well-placed cloud deployment.
Cloud adds another set of variables: virtualisation, shared host resources, virtual networking and placement across zones. Dedicated hosts, bare-metal services and carefully pinned instances can reduce some sources of variation, but they do not remove network distance or provider control of the facility. Typical intra-region network round trips can range from fractions of a millisecond to several milliseconds, depending on placement and service.
Metro circuits and exchange access vary by site and route. These are broad engineering ranges, not service guarantees.
Measure application-level latency as well as network round-trip time. Capture timestamps at market-data receipt, decision, order send and acknowledgement. Use synchronised clocks and record the clock source and estimated error. Report median, p95, p99 and p99.9, along with packet loss, retransmits and queue depth.
A median can look healthy while a small share of orders experiences damaging pauses. Test during busy market conditions and controlled bursts, not only on an idle weekday.
Keep a stable test design across locations. Send the same message sizes and rates, use the same protocol and software build, and compare the same destination where access permits. Separate host scheduling delay from network delay with packet captures and process-level timestamps. For market-data systems, measure feed gap recovery and message processing capacity as well as latency. A faster average is of little value if the system falls behind during a burst.
Failure behaviour belongs in the benchmark too. A second cloud zone can improve resilience but introduces replication traffic, extra cost and a recovery path that must be tested.
A second colo site needs hardware, connectivity and operational procedures. Confirm what happens after a host failure, a link loss, a power event and a regional outage. Recovery time, data loss and manual steps are measurable design properties, not assumptions to leave in a runbook.
Choose by workload and test the operating model
Many firms use both models. A latency-sensitive gateway or feed handler may run in colocation, while analytics, reporting, test environments and burstable replay workloads run in cloud. Hybrid architecture can preserve a short execution path and make non-critical capacity easier to scale.
It can also create a costly data-transfer boundary and more failure modes. Keep data close to the service that consumes it, and make ownership of each link explicit.
Use this sequence before committing to a migration or renewal:
- Set service objectives. Write the latency distribution, throughput, availability, recovery time and data-loss limits for each workload. Distinguish hard limits from targets.
- Draw the dependency and traffic map. Record endpoints, protocols, message rates, peak bursts, storage flows and every paid network boundary. Include failover routes.
- Build a like-for-like cost model. Price production, standby, backup, support and expected peak traffic over a 36-month period. Show capital amortisation and committed cloud discounts separately.
- Run a representative test. Use production-shaped traffic and the same software build. Measure tails, resource contention, packet loss, recovery and operational effort at each candidate site.
- Test the failure path. Remove a host or link under load, then verify failover, recovery and reconciliation. Record what requires an operator and how long it takes.
- Set exit and review criteria. Agree thresholds for cost, p99.9 latency, capacity headroom and recovery. Recheck them after traffic growth, provider changes or a network redesign.
For procurement, ask for port and cross-connect schedules, power assumptions, network route details, cloud billing exports and the terms behind any latency commitment. Ask operations to estimate rack visits, incident escalation and patching effort, not just monthly charges. Provider comparisons and platform coverage from Trading Technology can help frame the supplier landscape, but the decision still depends on the firm's own route, traffic and service objectives.
Stop doing these things:
- Comparing a cloud compute quote with a colo quote that excludes hardware and connectivity.
- Using average latency as the sole performance measure.
- Assuming a provider's nearest region or data centre is on the shortest route to the venue.
- Leaving egress, standby capacity and recovery tests out of the business case.
Frequently Asked Questions
Is colocation cheaper than cloud for trading infrastructure?
It can be, particularly for steady workloads that use dedicated servers and predictable network connections. Colocation requires hardware, support and connectivity costs to be included, while cloud bills can rise with storage, traffic and always-on capacity. Compare the same service scope over a multi-year period.
Does cloud provide low-latency trading?
Cloud can support low-latency services, especially when compute and network endpoints are placed carefully or dedicated capacity is used. Virtual networking, host contention and physical distance still affect latency and jitter. Benchmark p99 and p99.9 on the actual application path.
What costs are commonly missing from a trading cloud estimate?
Common omissions include data egress, cross-zone traffic, snapshots, backup retention, private connectivity and standby capacity. Some estimates also leave out support plans and logging charges. Use billing exports from a representative workload and model a peak month as well as normal usage.
Platforms · Software · Providers · Markets