Network infrastructure for low-latency trading

A design guide to the decisions that set a corridor's latency: which medium carries which signal, where the microseconds are actually spent and how to hold a vendor to the number in the brief.

01Latency is a physics problem first

Light in silica travels at roughly two thirds of its speed in vacuum. That gives about 5 µs of one-way delay per kilometre of fibre, or near 10 µs per kilometre round trip once the path is traversed both ways.

On a London–New York corridor of about 5,500 route kilometres, propagation accounts for the overwhelming majority of the delay. Equipment tuning matters, but no switch selection recovers the cost of a path that is 200 km longer than the shortest one available. Route choice is the design decision; everything else is refinement.

02Choosing the medium

Terrestrial fibre

~1.47× vacuum

Highest capacity and the only medium that carries full market data. Path length, not equipment, dominates the latency budget: a route 40 km longer costs roughly 200 µs round trip.

Microwave / millimetre wave

~1.0003× vacuum

Line-of-sight tower chains follow a straighter path than fibre and travel through air rather than glass. Capacity is narrow — tens of Mbps — and rain fade must be engineered for, not assumed away.

HF / shortwave RF

Near vacuum, ionospheric

Intercontinental trigger signals of a few bits. Availability varies with the ionosphere by hour and season, so HF is a layer above other media rather than a replacement for them.

Hollow-core fibre

~1.0005× vacuum in the core

Useful over short metro spans, typically inside a campus or between adjacent data centres, where the cost per metre is justified by the propagation saving.

03Building a latency budget

A latency budget lists every component of delay on the path and states what can be done about each. It is the document a vendor should be held to.

Propagation
Route kilometres ÷ medium velocity. Usually 80–95% of the total.
Shorten the path or change medium
Optical amplification & regeneration
Tens of µs across a long-haul span
Fewer regen sites, better spans
Layer-1 vs layer-2/3 handoff
Layer-1 transport avoids per-hop switching
Buy layer 1 where offered
Cross-connects & metro tails
Often the largest avoidable component
Colocate closer to the matching engine
Switching & NIC stack
Sub-µs on a tuned stack
Cut-through switching, kernel bypass

04A design sequence that survives procurement

  1. State the target, then the failure model

    01

    A design brief needs a latency target with a tolerance, and an explicit statement of what happens when the fastest path drops. A design with no stated failure model is a single-path design regardless of how it is drawn.

  2. Measure the corridor before pricing it

    02

    Establish the great-circle distance, then the shortest realistic path for each medium. This sets the physical floor for the corridor and shows immediately whether any vendor quote is credible.

  3. Tier the media by signal type

    03

    Trigger signals, order flow and full market data have different bandwidth and availability requirements. Assigning each to the cheapest medium that can carry it is where most of the cost saving lives.

  4. Survey who is already on the corridor

    04

    Licence databases such as the FCC ULS and national equivalents reveal existing tower chains and their owners. This determines whether you are buying capacity on an existing path or funding a new build.

  5. Price direct, then benchmark resellers

    05

    Direct carrier pricing and reseller quotes for the same underlying path can differ substantially. Run both, and review term sheet, SLA and MTTR commitments alongside the headline price.

  6. Commission against acceptance criteria

    06

    Acceptance testing measures delivered latency, jitter and failover behaviour against the written brief before the circuit goes into production and before handover to an in-house team.

05Resilience is part of the latency design

A microwave path that carries order flow will lose seconds to rain fade and minutes to a tower fault. What matters is what the trading system does in that window.

Practical designs pair the fast path with a fibre backstop on a diverse route, automate failover so it happens without a human in the loop, and select vendors on realistic MTTR rather than headline SLA credits. Diversity must be verified at the fibre-path level: two vendors frequently resell the same conduit.

06Common questions

What is low-latency network infrastructure in trading?
It is the transport layer that carries market data and orders between venues and trading systems, engineered so propagation delay and processing delay are as small as the physics and budget allow. In practice it combines fibre, microwave or HF/RF paths with colocation, layer-1 transport and a stated failover design.
Is microwave always faster than fibre?
Over the same corridor, microwave usually is, because radio travels through air at close to the speed of light in vacuum while light in glass travels at about two thirds of it, and tower chains follow a straighter path. The trade-off is capacity of tens of Mbps and weather-dependent availability, so microwave carries triggers and orders while fibre carries market data.
How much latency does distance actually cost?
In terrestrial fibre, roughly 5 µs per kilometre one way, so about 10 µs per kilometre of route length round trip. That is why route selection, not equipment selection, decides most of the latency budget on an intercontinental corridor.
Should a firm buy from a carrier or a reseller?
Both should be priced. Resellers can aggregate paths a single carrier cannot offer, but they also mark up capacity you could buy directly. The decision should compare price, contract term, SLA credits and realistic MTTR on the specific path rather than the vendor category.

07Work through a corridor

Route design, vendor sourcing and feasibility studies are the practice's core work. Describe the corridor or the problem.