19Technology Strategy

Virtu Financial’s communication and data processing expense was $209 million in 2019 and $249 million in 2025. In between, its revenue went from $1.5 billion to $3.2 billion, back down to $2.3 billion and up to $3.6 billion; the expense rose by a fifth and fell only once, by less than 1%. A trading firm does not choose each year whether to be fast. The latency tier it pays for is a decision about which businesses it will be in, taken once and paid every year, whatever the year brings.

19.1 Architecture as strategy

Book 13 built the low-latency stack and Book 14 the networks under it. This chapter is about the decision above both: how much speed to buy, for which strategies, and what else the same money could do. The decision is strategic because the costs are fixed (chapter 1’s fixed cost base; Figure 19.1) and the benefits depend on what competitors buy.

The chapter’s three tiers are priced from venue fees Book 14 read from published fee schedules, plus the firm’s own hardware, routes and people, which are inputs (Box 19.1): $0.29 million a year for tier 1, $1.35 million for tier 2 and $4.45 million for tier 3, per strategy. The venue fees are the small part: $36 000, $450 000 and $946 000 a year.

19.2 What speed is worth, strategy by strategy

Speed earns money in races: two or more firms react to the same public event, and only the first gets the trade. Aquilina, Budish and O’Neill (2022) measured races directly in stock exchange message data for FTSE 100 stocks: about one race a minute per symbol, lasting 5 to 10 millionths of a second in the modal case, about a fifth of trading volume, with six firms accounting for over 80% of race wins and losses. Races were roughly a third of the cost of liquidity. For a strategy, the question is what share of its revenue is decided by races.

The chapter’s race model splits a strategy’s revenue pool $V$, shared by $n$ competitors, into a race share $\alpha$ and the rest. Races go to the firms in the fastest tier present, shared equally; the rest is shared equally among all $n$. A firm earns

where $m$ is the number of firms in the top tier (Listing 19.1). Five strategies differ in $V$ and in $\alpha$: futures market making ($60 million, 0.8), ETF arbitrage ($30 million, 0.9), equity statistical arbitrage over minutes ($20 million, 0.1), options market making ($40 million, 0.5) and crypto basis trading ($10 million, 0.3). The values are illustrative; the race shares are set against the finding that races are a fifth of volume and a third of the cost of liquidity in a market where market makers dominate both.

def revenue(s, mine, others):
    """mine, others: tier indices (higher is faster)."""
    n = 1 + len(others)
    top = max([mine, *others])
    p_win = 1 / (1 + sum(o == top for o in others)) if mine == top else 0.0
    return s.pool * (s.alpha * p_win + (1 - s.alpha) / n)


def best_tier(s, tiers, others):
    profits = [revenue(s, i, others) - t.annual for i, t in enumerate(tiers)]
    return max(range(len(tiers)), key=lambda i: (profits[i], -i)), profits


def arms_race(s, tiers, n, start=0, max_rounds=100):
    """Best-response dynamics from everyone at tier `start`; raises if they cycle (no pure equilibrium found)."""
    prof = [start] * n
    for rounds in range(1, max_rounds + 1):
        changed = False
        for i in range(n):
            b, _ = best_tier(s, tiers, prof[:i] + prof[i + 1:])
            if b != prof[i]:
                prof[i], changed = b, True
        if not changed:
            return prof, rounds
    raise RuntimeError("best responses cycle: no pure equilibrium reached")

19.3 Tutorial: the fastest tier

Against its rivals, the firm should buy tier 3 for futures market making (profit $13.95 million a year, against $2.11 million at tier 1), ETF arbitrage ($9.65 million against $0.31 million) and options market making ($19.55 million against $3.71 million); tier 2 for the crypto basis trade ($3.05 million), where no rival has more; and tier 1 for statistical arbitrage over minutes ($3.31 million), which loses $2.16 million a year by buying tier 3. The same firm is a speed business in three strategies and not in two, and its architecture should say so.

With three tiers the argument applies step by step: for futures market making $\alpha V=48$ and the step from tier 1 to tier 3 is $4.16 million, so up to eleven firms would race; with five, all five do. Options market making gives $20/4.16$, four firms; the crypto trade, $3/1.06$ for the step to tier 2, two firms; statistical arbitrage, $2/1.06$, one.

19.4 Latency tiers and the arms race

The equilibrium spends $22.2 million a year on speed in futures market making and in ETF arbitrage, against $1.4 million if no one upgraded: 93.6% of the spending buys nothing for the industry, since the pool is the same. The industry’s profit in futures market making falls from $58.6 million to $37.8 million; in ETF arbitrage, from $28.6 million to $7.8 million, about a quarter of what it would be. Options market making wastes 92.1%, the crypto trade 59.8% and statistical arbitrage 42.7%. Across the five, the industry spends $68.6 million where $7.2 million would do: 89.6% of its speed spending is waste in the model’s sense. This is the argument Budish, Cramton and Shim (2015) make for markets in general: competition does not shrink the arbitrage opportunities, it raises the speed needed to capture them.

For a single firm the conclusion is the opposite of the industry’s: in a race that others are running, not buying speed means losing the race’s revenue. The firm’s decision is where to race and where not to, and where the pool is small or the races few, to leave.

19.5 Platform teams and product teams

The latency tier cuts across the two. Tier 3 infrastructure is usually a platform, since venues, cabinets and routes are shared by every strategy that races; the tutorial charges each strategy its own tier only for clarity. What belongs in a platform is what several product teams need in the same form; what belongs in a product team is what makes a strategy different. Skelton and Pais’s Team Topologies (2019) gives the vocabulary; chapter 21 applies it to a trading firm’s engineers (Figure 19.4).

19.6 The technology budget

The chosen tiers cost $14.97 million a year and production support $1.2 million: $16.2 million of run. The year’s projects, new venue connections, an FPGA feed-handler upgrade and a research platform, cost $4.0 million of change: run is 80.2% of the budget. The hook’s filings show why the split matters: a cost that did not fall when revenue fell by 29% from 2020 to 2023 is a run cost, and every tier decision adds to it. Flow Traders’ annual report puts its technology expenses at € 70.6 million in 2025, against € 66.6 million in 2024, beside a net trading income of € 485.8 million: another fixed line in a variable business.

19.7 Build: latency tiers

19.8 Exercises

19.9 Problem: The Fastest Tier

19.10 Interview questions