Where the infrastructure sits

A market has a street address 

Behind the word venue there is a room. Rows of steel cabinets, cold air pushed through them, and in the cabinets the servers that hold the order book and pair one instruction against another. Around that room sit the cabinets of the firms that want to reach it: banks, wholesale market makers, brokers, and the operators of the networks that connect them. The building is a data centre, and a handful of them carry a disproportionate share of the world's financial traffic. Interbank currency trading concentrates in a small number of facilities near London, New York, Tokyo and Singapore, because that is where the firms quoting currencies put their machines, and every subsequent arrival had a reason to join them rather than sit somewhere quieter.

The software doing the pairing is the matching engine. It holds the resting orders, applies the venue's priority rules, and produces a trade when two instructions are compatible. Everything a participant sees of a venue, the quotes, the depth, the confirmations, is that one program's output, distributed outward. The engine exists in one place at a time, which is what makes location a real variable rather than an abstraction.

What colocation means 

Colocation is the practice of renting space for a firm's own equipment inside the building that houses the venue, rather than connecting to it from elsewhere. A firm takes a cage or a cabinet in the same hall, installs its servers there, and buys a cross connect, a physical cable run by the data centre operator from that cabinet to the venue's own. The distance the signal covers stops being a national or intercontinental one and becomes a walk down a corridor.

Serious operators go further than proximity. Many exchanges equalise the cable length to every cabinet in the hall, coiling the surplus so that a cabinet close to the engine has no advantage over one at the far wall, because the alternative is an auction for the nearest rack. A firm that colocates is therefore buying entry into a group whose members all reach the engine in roughly the same time, not a position at the front of it.

The reason firms pursue it is not glamour and it is rarely raw speed for its own sake. A quote is only good for as long as the market it describes has not moved. A firm sitting far from the venue is always working from a slightly older picture, and it prices that staleness in: it quotes a little wider, or it declines a little more often, because the risk of being dealt against on a price the market has already left is carried by whoever is quoting. Proximity narrows that window. It does not close it.

Distance is a cost physics imposes 

Latency is the delay between sending something and it arriving. Round trip time is the delay between sending an instruction and receiving the answer, which is the number that actually matters to an order, because an order is not complete until its confirmation has come back. Both are usually quoted in milliseconds, and a meaningful part of both is not an engineering choice at all.

Light moves slower through glass than through vacuum, at roughly two thirds of the speed, and a signal in a fibre optic cable is bound by that. Distance therefore sets a floor beneath which no amount of money can push a connection, and the floor is easy to calculate.

Read the last two rows together and the design of the industry stops being mysterious. An order crossing an ocean spends tens of milliseconds in transit on the physics alone, before a single piece of equipment has looked at it, and a busy market can print many updates in that interval. The instruction arrives describing a moment that has passed, and the venue answers it against the moment it actually arrived in.

What a round trip is actually made of 

The measured delay a retail participant experiences is a sum of legs, not a single number, and the legs are owned by different parties. A click travels from the device to a local router, usually over a wireless link. From there it passes to an internet provider's network, then across whatever route that provider has to the broker's gateway. Inside the broker, the instruction is validated, checked against the account's margin, and passed to the routing layer, which is the point at which the lessons earlier in this module take over. Only then does it cross to the venue. Then all of it happens again in reverse, carrying the answer.

The shape is the lesson. The segment a broker builds and pays for is often the smallest one in the chain, while the segments between a household and that broker are longer, more variable, and owned by companies with no involvement in financial markets at all. This is why a published figure for a broker's internal path and a figure for what a client observes on a laptop are answers to different questions, and why comparing them directly produces nonsense.

Where YAL's own infrastructure sits 

A broker states two things about its own placement: which facility its servers occupy, and how far those servers are from the firms whose quotes it aggregates. YAL's trading infrastructure is located at Equinix LD4 in London, with under 1 ms to its liquidity sources. The first of those is a place, and the second is the length of the leg drawn as one row in the arithmetic above.

The client's own connection is part of the path 

A retail participant is not connected to a venue. A retail participant is connected to a broker, over the ordinary public internet, using ordinary consumer equipment, and every property of that link becomes a property of the order. A wireless link shares the air with the neighbours. A mobile network hands a device between towers and reprioritises traffic under load. A residential provider routes packets by whatever arrangements it happens to have, which can send one a long way round for reasons unrelated to the traveller.

There is a further point that is easy to miss. The screen is downstream of the same path. Quotes reach a chart by travelling the identical route in the other direction, so a slow or congested link means the picture on the screen is older than the market, in exactly the same proportion as an instruction sent from it arrives late. Distance is not a delay at the moment of sending. It is a persistent offset between what is displayed and what is currently true.

One practice sits on the boundary and is worth naming precisely. Some participants rent a virtual server in a data centre near their broker's, so that an automated strategy runs from there rather than from a home machine. The convention is bounded and its limits are known: it shortens the leg between the strategy and the broker and does nothing to the legs beyond it, it introduces a further supplier whose availability now matters, and it is irrelevant to anything a person decides by hand, because a human deciding takes far longer than any of the intervals under discussion. It is described here as a thing that exists in the market, not as a step this page puts forward for anyone.

Speed and consistency are different measurements 

An average delay describes the middle of a distribution and says nothing about its edges. The variation around that average has its own name, jitter, and for an order it is frequently the more consequential of the two. A path that is reliably slower behaves predictably. A path whose delay swings widely produces instructions that arrive at unpredictable points in a moving market, which is the mechanism behind the difference between the price on the screen when an instruction was sent and the price it was completed at.

The connection to fills is direct and mechanical. Where a counterparty holds the right to a final check before accepting, a quote is compared against the market at the instant the instruction lands, not the instant it was sent. Longer and more variable transit means a wider spread of arrival moments, which means a wider spread of outcomes: more instructions that land within tolerance, and more that land outside it and are declined or answered at a different level. The consistency of fills is therefore partly a statement about a network, which is an uncomfortable but accurate way to put it.

What is and is not in a client's hands 

Dividing the chain by who owns each segment produces a short and slightly deflating list. Two of the segments belong to the client. Everything else belongs to somebody the client has no relationship with.

• In the client's hands: the equipment at the near end, and the link from it to an internet provider. Wired against wireless, a congested machine against an idle one, a household connection against a rented server in a facility. These are properties of a home or an office, and they are the only part of the path a client is a party to.
• Chosen, but not controlled: the internet provider and, separately, the broker. Both are selections made once, and each carries a set of arrangements that the person selecting cannot inspect in detail or alter afterwards.
• Outside anybody retail: the route traffic takes between networks, the broker's placement and cross connects, the panel quoting behind it, the venue's engine and its rules, and the speed of light in glass. These are not preferences. They are the industry's physical layout, arranged long before any individual account was opened.

Stated plainly: the segment a client owns is the one nearest the client, and on a residential connection it is often the largest single contributor to the total. That is a useful thing to know, and it is also the end of what this page can honestly say about it, because what any particular household ought to do about its own network is not a question a lesson on market structure is in a position to answer.

Where practitioners disagree 

The loudest disagreement is about whether any of this matters to a retail participant at all. One position holds that milliseconds are the concern of firms whose entire method is arrival order, and that for someone deciding over minutes or days transit time is swamped by every other variable. The opposing position holds that this confuses two different things: nobody claims a slow path changes a view about a market, but the price an instruction is completed at is set at the moment it arrives, so transit is a cost paid on every order whatever horizon it was formed on. The disagreement persists because it is really an argument about which cost dominates, and that differs by participant.

A second and more technical dispute concerns whether published latency figures can be compared between firms at all. There is no shared convention for where the measurement starts and stops. A number can be taken from a gateway to a venue, from an order's acceptance to its acknowledgement, or across a whole internal stack including validation, and the same infrastructure produces very different figures depending on which is chosen. Practitioners who work on these systems tend to treat cross firm comparison of a single published number as close to meaningless without the methodology attached, and methodology is rarely attached. That scepticism applies to every firm's figures, including the ones on this site.

The third disagreement is about diminishing returns. Every improvement in a path is a subtraction from a total, so the same spending matters less as the remaining delay shrinks, and beyond some point the binding constraint stops being distance and becomes something else: whether a provider is quoting size at that moment, whether a counterparty accepts, how a venue prioritises. Where that point sits is contested, it moves as the technology moves, and it is not the same point for a wholesale firm and for a household. That it exists is not contested, which is why speed alone has never been a complete description of execution.