Latency has been on the capital markets agenda for more than two decades, and for most of that history the proposition was simple: go faster, win more. A panel at A-Team Group’s recent ExchangeTech Summit London, titled “Building the Ultimate Low Latency Highway: Execution Gateways, Order Routing and Ensuring Near Zero Jitter and Determinism”, spent much of its time challenging that proposition. Speed, the room broadly agreed, has become table stakes, a floor below which a firm cannot fall, rather than an edge it can press. The harder question is what differentiates a trading operation once everyone is fast enough.
The session was moderated by Steve Hamilton, an Independent Non-Executive Director and Market Structure Consultant, and brought together Anvar Karimson, Group Chief Technology Officer at Kepler Cheuvreux; Irina Sonich-Bright, Head of Market Access Product Development for UBS Global Markets; Deepak Dhayatker, Chief Technology Officer at Rapid Addition; and Tom Swarbrick, Head of EMEA Sales at Exegy.
An opening audience poll asked where latency still drives alpha most, and placed order execution and routing ahead of market data processing and signal generation. One panellist pushed back on the framing immediately, arguing that the answer depends entirely on the participant. Principal trading firms building on FPGA boards for both feeds and order access tick every box; large agency brokers, serving varied client bases across high-touch, programme trading and electronic desks, generate value in places that have little to do with raw access speed. The point that held the panel together was that tick-to-trade latency is a chain, and optimising one link leaves the others untouched – there is little use executing quickly on poor data.
Determinism, Not Speed, Is the Real Target
If the panel had a consensus, it was that determinism and predictability now matter more than absolute speed. Several speakers returned to the distinction. Software is flexible but less deterministic; hardware is more deterministic but carries a cost of build and use that keeps it short of mainstream accessibility. One panellist captured the trade-off with an image of a fast car stuck in heavy traffic: for a benchmark-driven execution model such as VWAP, knowing the route and avoiding roadblocks matters more than outright pace, and paying per-client FPGA costs for a strategy that is indifferent to microseconds makes little sense.
A second consensus was subtler and arguably more consequential: the ability to change a system has overtaken the ability to run it quickly. With market microstructure shifting – the rise of bilateral trading and systematic internalisers, extended trading hours, new order types and venues – one panellist argued that the fastest possible system that takes a long time to alter is now less of an advantage than it once was. Agency clients assume their brokers can turn around new venues and order types rapidly, and that agility competes directly against the slower evolution that specialised hardware imposes.
The most detailed contribution came on architecture. Low latency, one panellist pointed out, is not a component choice but a discipline: a holistic approach spanning the physical layer, hardware, operating system and application, governed by the theory of constraints – the principle that a complex system has a small number of bottlenecks, and that optimising anything other than those bottlenecks yields no improvement. The practical inventory ranged across colocation and proximity cloud offerings from the major vendors, precision time protocol clocks, thread pinning and interrupt isolation, NUMA node configuration, kernel bypass through frameworks such as DPDK and Solarflare Onload, and CXL 3 remote memory pooling. The framing that landed was that observability must be designed in from the start, because it cannot be retrofitted across an estate after the fact.
The FPGA Question Becomes a Question of Where, Not Whether
The acronym came up repeatedly during the session, but the panel’s treatment of FPGA was notably more measured than evangelical. The case for it – that the quickest and most deterministic route for market data and execution runs through programmable hardware – sat alongside a recurring caveat about cost, a narrower talent pool and slower evolution. The position the panel settled on was selective deployment: place FPGA at the highest-throughput bottleneck, run pre-trade risk checks there where they are genuinely fast, and offload enrichment and standardisation to a software application via direct memory access.
That logic is now shaping product strategy. Hardware acceleration once reserved for latency-obsessed proprietary trading is increasingly being repackaged for agency use, with market-data and execution functions split apart and configuration exposed in higher-level languages, so that teams without specialist hardware skills can work with it. These capability claims are necessarily self-reported, but the direction of travel matters more than any single product: a market moving past speed-at-any-cost toward a return-on-investment calculation, where the real prize is server consolidation and freeing up estate for higher-value work.
Where the Regulation Turns the Argument
If the speed economics framed the session, market structure and regulation gave it edge. One panellist argued, pointedly, that MiFID II had killed the venue competition that MiFID I created: where the earlier regime allowed firms to seek a better European best bid and offer across venues, the consolidated reference price removed that incentive, leaving firms frequently obliged to trade at the speed of the slowest relevant market. The observation that exchanges introducing speed bumps appear to have lost no market share was offered as evidence that engineered fairness does not cost venues their flow – at least in EMEA, where, on the same argument, meaningful venue competition is largely absent.
Compliance obligations, the panel agreed, are not going to ease. The weight of the argument fell on a single asymmetry under DORA: a firm remains accountable to its regulator for the resilience of critical third-party providers, yet several of those providers, particularly large US firms, decline to answer resilience questionnaires at all. The moderator relayed accounts of providers refusing such requests outright, raising the question of how effective a regime can be when the entities it designates as critical simply will not engage. The panel set this against a wider regulatory picture – incoming AI rules, MiCA for crypto, and a UK digital-assets consultation that looked, to at least one speaker, more protective than lighter-touch rhetoric would suggest.
Which returns the discussion to the question that ran underneath the whole session. If market access is eventually commoditised, with shared deterministic tooling and identical cable lengths, speed can no longer be the differentiator. Where the edge moves next – into agility, into cost discipline, into the intelligence layered on top of execution – is the question the industry is now working out, and the one this panel was sharp enough to put on the table.
Subscribe to our newsletter
