The Geography of Latency: Why Physical Distance Still Matters When Markets Are Electronic

Electronic markets did not abolish geography; they changed its units. A trade that once depended on which broker stood closest to the specialist's post now depends on which server sits closest to the matching engine, and the distances that matter are now measured in meters of fiber. What has not changed is that information takes time to travel. Light in glass covers about 200 kilometers every millisecond, which sounds fast until it is set beside the few microseconds a modern matching engine needs to process an order. On that scale the Atlantic is a wide place, the suburbs of Chicago are a long way from downtown, and a decision made in London about a price set in Aurora, Illinois, is older than it looks.

This piece maps the physical footprint of the markets that matter to a firm with offices in Chicago, London, and Bahrain, from the cities where people sit to the sites where orders are matched, and works through what the map implies. We compute the theoretical minimum time for information to travel between the hubs, compare that floor with the speed of light in free space, and ask what one microsecond is worth to strategies of different speeds at different distances. The conclusion is not that speed is everything. It is that distance sorts participants into those for whom a given piece of information is live and those for whom it is already history, and that a firm has to know which of the two it is before deciding where to put its machines.

Light Is Fast, but It Is Not Instant

The speed of light in a vacuum is 299,792 kilometers per second. In the silica glass of an optical fiber it is slower, because the glass has a refractive index of roughly 1.47, which puts the signal at about 204,000 kilometers per second; the industry rounds this to 200,000 and speaks of five microseconds per kilometer.^[1] Microwave links through the air travel at nearly the vacuum speed, which is why towers went up along the corridors where a few milliseconds mattered most, but they need line of sight, carry little bandwidth, and degrade in heavy rain. Hollow-core fibers that guide light through air rather than glass narrow the gap over short distances. For anything that crosses an ocean, glass is the medium and 200,000 kilometers per second is the number to plan around.

Two other facts set the scale. First, cables do not follow great circles. Submarine systems run between landing stations chosen for geology and politics; terrestrial fiber follows railways, pipelines, and highways because that is where the rights-of-way are. Engineering rules of thumb put a real route anywhere from a few percent to a third or more longer than the straight line, depending on the corridor, and amplifiers, regenerators, and switches add their own delay, so every figure in this piece is a floor rather than an estimate.^[2] Second, the electronics at each end are now fast enough that propagation dominates. A matching engine turns an order around in single-digit microseconds; a well-built trading system reacts in a few more. A participant in the same building as the engine spends most of its round trip inside silicon. A participant a thousand kilometers away spends almost all of it inside glass, and no engineering at either end changes that.

Where the Matching Actually Happens

Figure 1 draws the five hubs that matter most to a firm with offices in Chicago, London, and Bahrain, together with the sites around them where orders are actually matched. The first thing the map shows is that the exchanges are not where the cities are. Chicago's principal futures matching engine sits in Aurora, roughly 60 kilometers west of the Loop. The two largest New York equity venues match in Mahwah and Carteret, New Jersey, close to 57 kilometers apart and neither within sight of Manhattan. London's derivatives matching happens in Basildon to the east, and the colocation campus that hosts much of the currency market is in Slough to the west. Frankfurt's derivatives exchange matches outside the city. The reasons are prosaic: power, floor space, physical security, and room to grow were available in industrial suburbs when the engines were built. The consequence is that, for latency purposes, Chicago is Aurora and New York is a triangle in New Jersey.

The arc on the map joins the two office cities on either side of the Atlantic and carries the theoretical minimum: the great-circle distance divided by 200,000 kilometers per second, one way. The Atlantic is about 32 milliseconds wide between Chicago and London, and about 28 between New York and London; London to Manama, a pair the map leaves to Figure 2, is a little over 25. At the scale of this map the matching sites collapse into their cities; Aurora is the same dot as Chicago, and the New Jersey sites lie on top of New York. That is the right picture for a strategy that thinks in seconds and the wrong one for a strategy that thinks in microseconds, for which Carteret and Mahwah are separated by close to 290 microseconds of glass, an interval that is many multiples of the engine's own processing time. Geography in electronic markets is scale-dependent: the same map is a point or a continent depending on the clock the reader brings to it.

The Theoretical Minimum, Pair by Pair

Figure 2 computes the floor for every pair of hubs and sets the fiber figure beside the free-space figure, the time light would take through a vacuum along the same great circle. The ordering is what matters. Two hops are short: London to Frankfurt at about 3 milliseconds and Chicago to New York at under 6. The transatlantic pairs cluster between 28 and 35 milliseconds. The pairs that reach the Gulf run from 22 milliseconds for Frankfurt to more than 56 for Chicago. The free-space bars are about a third shorter than the fiber bars in every case, and that third is the prize microwave networks were built to collect on land corridors where towers can be placed. Over water no such option exists, and the fiber floor is the floor.

The numbers explain a good deal of market history. The Chicago to New York corridor is short enough that a few milliseconds of advantage over fiber could be bought with towers, and it was; the same arithmetic on London to Frankfurt produced a similar build.^[3] The transatlantic crossing offered no such shortcut, so the competition there moved to cable route length and to the placement of equipment at each end. And the Gulf, at more than 50 milliseconds from the American hubs, is far enough that no participant sitting there can react to Chicago in Chicago's own time, whatever it spends. Distance in these cases is not a cost to be minimized. It is a fact that determines which contests a participant is in at all.

What a Microsecond Is Worth Depends on Where You Stand

To turn the floor into economics we need a model of how an opportunity loses value while a participant is still on the way to it. The simplest one, which a companion piece in this Library develops at length, lets the value of an opportunity decay exponentially in the delay τ between the event that created it and the arrival of an order at the venue: V(τ) = V₀ · 2^(−τ/h), where h is the opportunity's half-life. A participant located d kilometers from the venue cannot see the event or deliver the order faster than light in fiber, so its delay is at least d / v, and the fraction of the opportunity still alive when its order can first arrive is 2^(−d/(v·h)). The marginal value of shaving one microsecond off its systems is proportional to that fraction: a microsecond is worth the most to the participant who still has something to win, and nothing to one who has already lost.^[4]

Figure 3 plots that fraction against distance for three stylized strategies whose half-lives differ by orders of magnitude: a time-sensitive strategy at 200 microseconds, an event-driven strategy at 20 milliseconds, and a slow strategy at one hour. Distance runs on a logarithmic axis from one kilometer to twenty thousand; three hub distances from the map are marked on the middle curve.

The three curves describe three different relationships with the map. For the time-sensitive strategy, distance is binary. The value of a microsecond halves every 40 kilometers and is effectively gone within a few hundred; such a strategy either runs in the building where the matching happens or does not run. For the event-driven strategy, distance is a graded tax. The value halves every 4,000 kilometers, so New York is close to Chicago, London is meaningfully farther, and the Gulf keeps only a small fraction; a firm can choose to pay the tax or to route around it by placing the decision closer to the event. For the slow strategy the curve is flat, and the map might as well not exist. This is the sense in which colocation is a necessity for some participants and an irrelevance for others. It is a matter not of ambition or budget but of which curve the strategy sits on.

Every Pair, and the Last Mile

Figure 4 gives the full table: each hub pair and each metro hop, with the great-circle distance, the one-way and round-trip fiber floors, and the free-space floor. The metro hops repay attention. They are the distances that separate a matching engine from a participant who has chosen the wrong building, or a participant who needs to watch two engines at once, as anyone trading the same instrument on two New Jersey venues does. At five microseconds per kilometer, the 24 kilometers between Manhattan and Carteret cost about 120 microseconds one way; Carteret to Mahwah is close to 290. In the fast regime of Figure 3 those are not rounding errors. They are the whole contest.

Two cautions apply to reading the table. The floors assume a straight line, and there is no straight line: the London to Manama route runs through the Mediterranean, is constrained at Suez, and follows the Red Sea, so the realized path is longer than the great circle by a margin that depends on the cable. And the floors are one way, while most decisions involve a round trip: seeing the event, deciding, and delivering the order. A participant who wants to know how late it is at a remote venue should start from the round-trip column and then add whatever the route and the equipment impose. The result is always larger than the floor, but the ordering of the pairs never changes, which is why the floor is the useful quantity to reason with.

How We Think About Geography

Oak St. does not treat the map as a race course. We treat it as a sorting device. Every decision our systems make has a horizon, the time over which the opportunity it acts on decays, and the first question we ask about a decision is which of the three curves in Figure 3 it lives on. Decisions on the flat curve are placed wherever the data and the people are, and geography plays no part. Decisions on the graded curve are placed as close to the event as the infrastructure allows, and the residual delay is modeled explicitly as a known cost, so that a strategy is never surprised by its own lateness. Decisions on the steep curve are made from inside the building where the matching happens or not at all; we do not attempt them from across an ocean, and we do not pretend that a faster link changes the physics.

The map has a second reading that has nothing to do with light. Chicago, London, and Manama span eight or nine hours of clock, depending on the season, which means the firm's people are awake across most of the trading day of every major market without anyone working through the night. That is a human geography, and it matters for reasons the latency numbers do not capture: someone is awake when a system misbehaves, and someone is rested when the next problem arrives. The machines are placed by the arithmetic in this piece. The people are placed by the sun.

Geography did not disappear when trading moved into data centers. It shrank to the size of a building at one end of the scale and stretched to the width of an ocean at the other, and it stopped being visible to anyone who was not measuring in microseconds. The discipline is to know which end of the scale each decision is on, to place it accordingly, and to resist the temptation to buy speed where the map says it cannot help.