Is Speed No Longer the Hard Part in Trading Infrastructure?

A panel of high-performance trading infrastructure architects, convened at A-Team Group’s recent TradingTech Summit New York to discuss the blueprint for speed, spent remarkably little of its time talking about speed. The tick-to-trade race still runs, and firms still shave nanoseconds where a strategy justifies it, but raw latency is the part this end of the market has long had a discipline for. The harder problems sat elsewhere – in the supply chains that no longer deliver, the refresh cycle that no longer holds, the security overhead that eats into processing capacity, and, dropped into the room mid-session, a regulatory proposal that unsettles two decades of assumptions about why the architecture looks the way it does.

Speed has become the floor. The competitive edge has moved to the decisions around it: what to build, what to buy, what to consolidate, and how to source the hardware to do any of it at all.

Determinism

Latency variance, one line of argument held, is the canary in the coal mine for the stability and resilience of a platform. A system that slows down precisely when the market gets busy is not just ceding position to competitors. It is signalling that it may start dropping data, acting on stale prices, or falling over entirely. The whole infrastructure therefore has to be designed holistically, so that processing data, making decisions and acting on orders all keep pace with the rate at which inputs arrive. The raw network bandwidth coming into the platform has to sustain the rate at which inputs actually arrive, at all times – so that the market spikes on busy days are handled with the same assurance as any other.

Build, buy and talent

The build-versus-buy question ran through the whole discussion. One panellist framed it as a time-to-market decision rather than a technology one, and underneath that, a talent decision.

FPGA programming remains arcane – the skill set is scarce, expensive to acquire, expensive to retain, and slow to grow internally. The same panellist’s framing was that the build-or-buy line should be drawn around what is genuinely strategic. Where a capability is genuinely strategic – the part that captures alpha – build it in-house and keep the engineers to run it. Everything else is a candidate for a proven, battle-tested platform that the rest of the industry has already stress-tested. The example offered was a firm running two FPGA engineers across trading on 22 global markets, having concentrated its own people on algo and alpha capture and taken a platform for everything else.

Writing your own feed handler against your own FPGA, one panellist noted, lets you discard the portion of a feed you do not need – the 30% of the data that never gets used – whereas a commodity feed, built to serve everyone, cannot. Drop what you will not read, and you are faster than the commodity alternative. The strategic slice is where owning the code pays; abstraction layers – APIs that let FPGA logic be written in something close to C – are lowering the barrier for firms not doing anything exotic enough to need to write it the hard way.

On whether AI widens access to the technology, the panel was blunt: it does, for anyone with the base capability to use it well. The value of an LLM in learning FPGA work, or anything else, depends on the operator already knowing enough to recognise when the output is wrong – the experience that fires a warning when a plausible answer does not smell right. New technology does not remove the need to understand fundamentals; it raises the cost of not understanding them.

Supply chain

Where the panel divided most sharply was how to respond to a hardware market that no longer delivers on demand.

One position held that the answer is to slow down, with commodity servers reportedly costing 23% more, memory scarce and GPUs effectively spoken for. The three-year refresh that has been industry norm becomes five, then seven, with firms paying maintenance to keep older hardware in service rather than replacing it. A 2023-vintage FPGA, on this view, is not losing you anything yet. The advice is to stockpile: get your own GPUs into the field before you need them, buy into the supplier’s process rather than the part, commit to years of purchasing in advance to get in line behind the hyperscalers. Quantised models running on smaller memory footprints make the wait more bearable.

The opposing position rejected the premise. The business, on one panelist’s account, does not want to wait even three years – it wants to embrace technology the instant it hits the market, and stretching the cycle is at odds with staying relevant. The answer is not to slow down but to source differently: look outside the traditional big providers to niche firms stockpiling components for delivery at a premium, and build the relationships to inject yourself into that flow. Underlying it was a pointed claim – that any suggestion an exceptionally important market participant cannot get hold of parts is simply not correct. There are angles, and the firms that matter find them.

A third line threaded between the two: consolidation as a supply strategy in its own right. Much hardware in production has never been properly performance-benchmarked, and where it has, spare capacity is often sitting unused. Break the estate down by asset class, find what can be consolidated onto bare metal or shared classes, prioritise the applications that genuinely need the headroom, and a good deal of the lower-end infrastructure problem can be self-served through hygiene work that went unaddressed for years – shuffling workloads across VM farms, cloud presence and physical grids that a large firm already owns.

The scarcity is also reshaping the physical footprint. Colocation space that used to be sold in cabinets is now sold by the room, with AI demand setting the minimum buy, and vendors using the constraint as leverage – monthly calls, one panellist reported, that open with the price rise for the month.

Security

Security drew some of the most concrete numbers of the session, around a consistent principle: the right tool for the right place, because the wrong tool is a performance cost you cannot afford.

In a colocation environment, the approach described was subtraction. Strip the box to its minimum – IP addresses rather than DNS, no routing daemon, only the routes to exactly where it needs to reach, one protocol, logging off wherever possible. The effect is a drastically reduced attack surface without the overhead of an endpoint agent that, by one panellist’s account, consumes 11% to 21% of CPU – enough that in a latency-sensitive context it takes you out of the game. That box, stripped and locked down in a colo, was described as the most protected part of the industry precisely because it is neither public cloud nor open corporate network.

The principle that closed the topic was defence in depth: no single answer, no layer perfect, inside risk weighed against outside risk. Like the operational problems elsewhere in the session, security is not solved once but managed continuously.

Rule 611

On the day of the panel, the SEC proposed rescinding Rule 611 of Regulation NMS – the trade-through rule, in force since 2005 – along with Rule 610(e), the prohibition on locked and crossed markets. Comments are due by 17 August 2026.

Much of the architecture built over the past twenty years exists partly to serve a regulatory structure, not only the laws of physics. Rule 611 requires trading centres to avoid executing at prices worse than protected quotations displayed elsewhere, which is a large part of why minimising latency between the New Jersey data centres that make up the US national market system has mattered so much. Remove it, and the questions reopen: what does best execution now require, how is fiduciary responsibility discharged in its absence, and how does that change what infrastructure is built to optimise for?

The SEC cites the emergence of tokenised NMS stocks and on-chain trading venues as a reason to revisit the rule – the same blurring of the line between cloud, exchange and settlement that one panellist flagged as a security paradox in the making, as pre-trade and post-trade analytics migrate into public cloud environments where exchanges themselves increasingly sit. In a connected action the same day, the SEC extended the compliance date for its 2024 Reg NMS amendments, citing the cumulative burden of multiple market structure initiatives landing across 2026 – expanded trading hours among them, the 24-hour trading question that ran through the summit’s other sessions.

The blueprint

The lightning round that closed the session put determinism first and security last – build for the platform to hold its shape, and do not end up on the front page of the newspaper – with a final instruction to know precisely what makes the business special and sustainable, then find the right partners for everything else.

That last point is where the answer to the panel’s own question settled. Speed is no longer where the hard differentiation happens for most firms, because the problems that decide competitiveness have moved to sourcing, talent, consolidation, security economics and a regulatory floor that has just been shown to be less solid than it looked. Building a platform that stays stable, stocked, defended and relevant while the ground underneath keeps moving is the harder discipline, and it is the one that now decides who competes.