Colocation and cloud solve different trading problems

Server racks in a data centre, illustrating colocation and cloud choices for trading infrastructure
Photo by panumas nikhomkhai on Pexels

If the order path needs predictable access to an exchange, putting it in a general-purpose cloud region is a compromise, not a like-for-like substitute for colocation. That does not make every broker a colocation customer. It means the workload, network path and failure model need to be separated before anyone compares monthly bills.

Consider a fictional multi-asset broker with 45,000 active accounts, around 2,000 concurrent users at peak and 400 to 900 order requests a second during busy periods. Its platform also serves mobile clients, runs overnight reports and distributes market data. Those services do not share the same latency budget or operating profile.

The decision is about the network path

Colocation places the broker's equipment in a facility near an exchange, liquidity provider or trading venue. A cross-connect provides a direct physical link between cabinets or networks in the building. This can reduce the number of network hops and avoid parts of the public internet, but it does not remove latency inside the broker's own stack. Feed handlers, risk checks, order routing and venue acknowledgement all contribute.

Cloud infrastructure replaces owned hardware with provider-managed compute, storage and networking. An instance in the right region can be close to a venue, and private connectivity may be available. But the broker has less control over physical routing, host placement and noisy-neighbour effects. Cloud latency can be entirely adequate for many workflows; its distribution and tail behaviour still need measuring against the actual order path.

For the fictional broker, 700 orders a second is not, by itself, a reason to rent cabinets. The deciding questions are whether the execution strategy has a tight latency target, whether a venue requires a particular connection, and how much burst capacity the service needs.

Compare like with like

The figures below are typical market planning ranges, not audited results or vendor quotes. They vary by facility, region, contract, hardware, traffic profile and service level. Latency figures describe network transport, not a complete order round trip.

MeasureColocation, typical rangePublic cloud, typical rangeWhat changes the figure
Network RTT within a metro area0.1 to 1 ms on a direct, nearby path0.5 to 3 ms, depending on region and pathVenue location, routing and network design
Monthly infrastructure cost for a small production footprint£8,000 to £30,000 for two cabinets, power and connectivity£15,000 to £50,000 for persistent compute and networkingHardware, licensing, bandwidth and support are often extra
Initial production deployment4 to 12 weeks, including circuits and hardwareDays to 4 weeks, including security and connectivityProcurement, approvals and integration work
Capacity changeDays to months if hardware or power must be addedMinutes to hours for supported instance typesQuotas, capacity availability and application design
Recovery to a secondary siteSeconds to minutes with engineered replication and routingMinutes to hours, depending on architecture and data recoveryRecovery objectives, replication lag and operator procedures

The cost rows are not directly equivalent. Colo estimates exclude the purchase and refresh of servers in many contracts. Cloud estimates can omit market data fees, software licences, logging, support and egress. A service streaming large volumes of quotes to users may see network charges that are immaterial for a low-traffic API. Model a full year, including spare capacity and failover, rather than comparing a rack quote with a single compute bill.

Put latency-sensitive work where it earns its keep

A sensible starting design is not "everything in colo" or "everything in cloud". Keep the narrow, venue-facing path close to the venue if measurements and business requirements justify it. Put services that benefit from elasticity or managed storage elsewhere, provided the added network hop does not affect trading controls.

In this example, the broker could put market-data capture, normalisation, pre-trade risk and order gateways in colocation. The customer portal, CRM integrations, reporting and research environments could run in cloud regions. A cloud-based risk dashboard can consume a replicated event stream, while the authoritative risk check remains on the execution path.

  • Colocation candidates: exchange gateways, low-latency market-data handlers, deterministic risk checks and services tied to venue connectivity.
  • Cloud candidates: customer-facing APIs, batch analytics, development environments, backups and workloads with variable demand.
  • Either location, with evidence: pricing services, internal risk aggregation and market-data fan-out, where throughput, cost and latency requirements vary by product.

That split adds an operational boundary. Engineers must define which system owns each state change, how timestamps are synchronised, and what happens when the link between sites fails. Trading Technology's coverage of broker platforms follows the infrastructure choices that sit behind those systems.

Design the failure path before the migration

Hybrid designs often fail at the joins. A primary colo gateway and a cloud-based backup are not resilient merely because they occupy different locations. They need independent network routes, tested credentials, replicated configuration and a controlled failover process. If both sites depend on the same carrier, DNS service or management plane, the diagram overstates the separation.

For the broker in this example, a practical first phase might retain primary execution in colo and use a second facility for standby gateways. Cloud can hold replicated operational data and host recovery tooling, but it should not be assumed to take orders automatically. Failover needs explicit sequencing for venue sessions, duplicate-order prevention, open-order reconciliation and market-data freshness.

  • Measure p50, p99 and p99.9 latency at the application boundary, not only with a network ping.
  • Record packet loss, retransmits, queue depth and feed gaps during normal trading and venue reconnects.
  • Test recovery with stale data, unavailable links and partial venue outages, then verify order and position state.
  • Include cross-connects, remote hands, cloud egress, support and hardware replacement in the cost model.

Use production-like traffic to test both steady load and bursts. A platform can meet its average latency target while queueing badly during an open or a volatile market. Keep a baseline for each route and compare changes after network, firmware or instance-type updates.

Make the choice workload by workload

Colocation buys physical proximity and greater control over the local network path. It also brings hardware lifecycle work, fixed capacity, facility contracts and dependencies on people who can act on site. Cloud buys faster provisioning, managed services and easier capacity changes. It brings less control over the underlying host and network, plus costs that move with usage and data transfer.

For a small firm, cloud may be the operationally sound choice even when its measured latency is higher. For a strategy sensitive to microseconds or a broker with venue-mandated connectivity, colo may be necessary for a narrow slice of the stack. Most decisions are less ideological: price the whole service, measure the real route, and assign a failure owner to every boundary.

The industry still measures distance in milliseconds and invoices it by the rack.

Frequently Asked Questions

Is colocation faster than cloud for trading?

It can be, particularly when equipment is in the same facility as a venue and uses a direct cross-connect. The full order path also includes application processing, risk checks and routing, so measure end-to-end latency and tail behaviour rather than relying on network distance alone.

Can a broker run trading systems in the cloud?

Yes. Cloud can suit customer APIs, reporting, analytics and some execution services, depending on latency and connectivity requirements. Test the actual route and define how the platform behaves if a region, network path or provider service fails.

What does trading colocation typically cost?

A small production footprint can typically cost around £8,000 to £30,000 a month for two cabinets, power and connectivity. Hardware, market data, software licences, support and remote hands may be additional, so compare full annual costs with a cloud design.

Which trading workloads should stay in colocation?

Exchange gateways, low-latency market-data handlers and time-sensitive risk checks are common candidates. The case depends on venue access, measured latency targets, traffic volume and the cost of operating hardware close to the venue.

How should a broker test cloud and colocation latency?

Measure latency at application boundaries under realistic traffic, including p50, p99 and p99.9 results. Record packet loss, queueing and feed gaps, and repeat tests during bursts, reconnects and failover rather than relying on a single ping.