Colocation is not automatically faster than cloud.
Colocation is a location, not a latency guarantee
Colocation puts your equipment in a data centre, often in the same facility as an exchange matching engine or market-data provider. That can remove distance and network hops. It does not remove queueing, switching, application processing or the delay between your system and the venue.
Inside a facility, a short cross-connect may add only a few microseconds, while the path through your own switches, firewalls and software can add much more. Across fibre, propagation is roughly 5 microseconds per kilometre one way, before equipment and routing delays. A server in a nearby building can therefore lose to a well-connected host in the exchange facility.
Cloud can deliver low and consistent latency for many workloads, particularly where the broker's matching, risk or client-facing services already run there. But a cloud region is not an exchange rack. A route may cross provider networks, shared virtual switches and several availability zones. Typical intra-region round trips can range from hundreds of microseconds to several milliseconds, depending on service, placement and traffic. Measure the route you will use, not a provider's best-case figure.
The network path matters more than the hosting label
Map the full path for each critical flow: market-data source to feed handler, order service to venue gateway, and execution report back to risk and client systems. Record physical and virtual hops, protocol conversion, encryption, multicast handling and where traffic leaves the provider network.
Market data is often the deciding constraint. Some feeds use multicast and require specialised network design, feed arbitration and hardware timestamping. A cloud service may support multicast only in specified configurations, or require a partner connection. Exchange connectivity also brings port, cross-connect, certification and market-data licensing costs that do not disappear when compute moves to cloud.
Compare production measurements at busy periods, not just idle ping tests. Capture median, 99th and 99.9th percentile latency, packet loss and burst behaviour.
Keep separate numbers for network transit and application processing. An average can look healthy while a small tail of delayed packets creates stale quotes or risk checks that miss their target.
- Use the same order flow, feed volume and software build in each test.
- Measure from the application process, with timestamps from synchronised clocks.
- Repeat during market open, news bursts and planned failover tests.
- Check the route after every network or provider change.
Trading Technology's coverage of platform and infrastructure decisions is useful context, but the decisive evidence is still a trace from your own systems.
Compare the bill, not the server line
Infrastructure comparisons often put a cloud instance price beside a rack quote and call the cheaper one the winner. That omits hardware refresh, cross-connects, bandwidth commitments, egress, support and the engineering time needed to operate the stack.
The figures below are typical market ranges in USD per month, not quotes. Location, contract term, redundancy and utilisation change them substantially.
| Cost item | Colocation, typical range | Cloud, typical range | Scope and caveats |
|---|---|---|---|
| Compute capacity | $500-$2,000 for rack allocation, plus hardware | $2,000-$8,000 for a comparable small server fleet | Cloud range assumes steady general-purpose compute; colo hardware purchase and refresh are separate. |
| Rack space | $500-$2,500 for a quarter rack | No direct equivalent | Power density, facility and location drive colo pricing. |
| Exchange or private connectivity | $100-$500 per cross-connect | $100-$1,000 for private-connect port or service charges | Carrier fees, exchange fees and data transfer may be additional in either model. |
| Network traffic | $300-$2,000 for a committed bandwidth service | $100-$2,000 for 1-10 TB of internet egress | Commit level, destination, discounts and traffic shape affect the bill. |
Cloud bills can rise quickly when systems send large market-data volumes between zones or regions. In-colo bandwidth can be more predictable, but a second facility, diverse carriers and spare capacity add fixed cost.
Model normal load, peak load and recovery load. Include the cost of running both sites during a migration or failover, not just the steady-state diagram.
For a useful comparison, price the same capacity and service level over three to five years. Include server depreciation and replacement, software licences, power, remote hands, cloud commitments, support plans and network charges. Then add the cost of the people who patch, monitor and recover it.
Resilience changes the architecture
Neither model makes a system resilient by itself. A single colo rack can fail through power, switching, carrier or human error. A cloud deployment confined to one zone can fail through a regional service issue or a bad configuration pushed across the fleet.
Colocation usually gives direct control of hardware, firmware, NIC settings and cabling. It also makes capacity planning physical: procurement and installation can take weeks, and spare parts must be on site or reachable. Cloud gives faster provisioning and simpler elastic capacity, but the dependency chain includes provider APIs, quotas, service limits and the control plane.
Design recovery around the trading function. A warm standby may be enough for reporting or overnight services; order entry and risk controls may need tested capacity ready to take traffic. Keep independent network paths where the risk justifies them, and test whether market-data entitlements and venue sessions can move during an incident. Recovery time objectives are not met by a second environment that has never processed live-like load.
Choose by workload and operating model
Many firms do not need a single hosting answer. A latency-sensitive market-data handler or venue gateway may sit in colo, while client APIs, analytics, monitoring and development environments run in cloud.
A hybrid design works only if the boundary is explicit. Cross-environment calls add latency and failure modes, and can turn a simple deployment into a network dependency that is hard to diagnose.
Colocation tends to fit when:
- Venue access, multicast feeds or deterministic network paths dominate the requirement;
- Load is stable enough to justify owned hardware and a multi-year facility commitment;
- The team can manage physical infrastructure, firmware and on-site recovery.
Cloud tends to fit when:
- Services need rapid environment creation, variable capacity or broad regional reach;
- The critical path does not depend on physical proximity to an exchange;
- The team can control service limits, egress, placement and provider-specific failure modes.
Some teams treat cloud as a test environment and later discover that a cloud-native production stack has different network behaviour. Others buy colo capacity for a latency target that the application cannot meet. Match the placement to a measured bottleneck, not to a general belief about either model.
Run this decision test
Before renewing a rack contract or moving a trading service, make the choice testable. Use this checklist with network, application, operations and finance owners:
- Set a service target. Write down the required median and tail latency, recovery time, throughput and availability for each flow.
- Trace the current path. Identify every network hop, provider boundary, cross-connect and application stage between data source and trading decision.
- Benchmark a representative workload. Run the same build and traffic profile in both candidate locations, including peak bursts and recovery conditions.
- Build a three-year cost model. Include hardware refresh, connectivity, egress, facility power, support, engineering and duplicate capacity.
- Exercise failure modes. Test lost links, a failed host, a provider-zone outage and a market-data interruption, then record recovery time and data integrity.
- Assign an owner. Name who approves changes, watches latency tails and leads recovery in each environment.
The useful result is not a universal verdict on colocation or cloud. It is a documented placement decision for each service, with a measured path, an operating cost and a tested failure plan. This week, run a timestamped trace of one critical order path from application process to venue and record its median, 99th and 99.9th percentile latency.
Frequently Asked Questions
Is colocation always faster than cloud for trading?
No. Colocation can reduce distance to an exchange, but application processing, switching and network routes still determine end-to-end latency. Measure the full order path and its tail latency in the actual deployment.
What latency should a trading firm expect in cloud?
There is no single figure: placement, network service, instance type and traffic load all matter. Intra-region round trips can range from hundreds of microseconds to several milliseconds, so benchmark the exact route and workload.
How much does exchange colocation typically cost?
A quarter rack can typically cost around $500 to $2,500 per month, before hardware, cross-connects and exchange charges. Power density, facility, redundancy and contract terms can materially change the total.
Can a broker use both colocation and cloud?
Yes. A firm may place feed handlers or venue gateways in colocation and run client services, analytics or development systems in cloud. The design needs clear boundaries because cross-environment calls add latency, cost and failure dependencies.
What should I measure before moving trading infrastructure?
Measure end-to-end median, 99th and 99.9th percentile latency, packet loss and throughput under representative peak traffic. Also test failover, recovery time, data integrity and the cost of running normal and recovery capacity.
Platforms · Software · Providers · Markets