by Stewart Orrell, Vertical GTM Lead at ZutaCore
Speed has long been the language of high-frequency trading, shaping every decision from where firms place their servers to the optimisation of every step between market data and order execution. Even a small gain can change the outcome, which is why trading infrastructure has always been built around one central question: how much performance can firms place as close to the market as possible?
For years, that work focused on faster chips, shorter network paths and sharper trading models. Now another limit is appearing inside the rack. As firms put more powerful CPUs and GPUs into colocation environments, performance depends not only on proximity to the market, but also on whether they can remove the heat those processors create.
In financial services, that constraint is harder to work around. Latency-sensitive trading systems need to remain close to the markets they serve, making these colocation sites part of the execution model itself. Rather than shifting workloads to locations where space and power are easier to secure, firms must extract more performance from fixed, highly contested environments.
That pressure is changing how firms think about cooling. Many strategic colocation facilities were built for air-cooled racks and lower-density workloads, yet today’s processors generate far more heat. As trading firms push more compute into the same environments, they are asking existing facilities to support levels of performance they were never designed to handle. Compute demand keeps climbing, while traditional cooling capacity cannot stretch at the same pace.
The same rack-density challenge is now reshaping the wider data centre market. In its 2025 Cooling Systems Survey, Uptime Institute found that 69% of respondents cited higher rack densities as a primary driver for direct liquid cooling adoption. For high-frequency trading firms, that finding reflects a familiar problem: how to raise performance inside sites that are already built, occupied and hard to change.
In some cases, colocation providers are already asking firms to make a proportion of new infrastructure liquid cooled. In shared facilities, adding more air cooling can mean major works and disruption for other customers, so trading firms need cooling choices that help them use the sites they already occupy more effectively.
When heat changes execution
The effect becomes clearest during volatile trading sessions. A platform can look stable in quieter markets, then behave differently when market data jumps, order flow rises and processors run hard for sustained periods.
In those moments, predictability carries the same commercial weight as raw speed. Trading systems have to absorb heavier data volumes and execute without unwanted variation entering the tick-to-trade path. If processors run too hot, they can throttle when firms need consistent behaviour most.
That loss of predictability can also make performance harder to diagnose. When latency becomes less consistent, firms have less visibility into how their systems behave under stress and less confidence in the link between a trading strategy and the machines running it. In an environment where firms compete on fine margins, that uncertainty can quickly become a performance issue.
This is why trading infrastructure is starting to be judged as a full system, rather than as a collection of faster components. Raw compute power still counts, but it only delivers value when the surrounding architecture lets it run reliably during peak demand. Deterministic tick-to-trade execution depends on the full system, including the cooling design beside the servers.
Waterless, two-phase direct-to-chip cooling gives firms a practical way to remove heat at the source. The approach takes heat away from the chip, supports sustained CPU and GPU performance and keeps water away from the server level. That last point is critical in financial environments, where teams often raise concerns about leakage, corrosion and risk around sensitive IT equipment.
The right cooling design also has to fit the realities of colocation. Some can connect to facility water loops, while others need a cooling design that can operate without that infrastructure. A system that can connect to available facility services or use an air-based cooling distribution unit gives trading firms more control over where they deploy high-performance hardware.
The pressure will grow as AI-driven analytics and real-time modelling move closer to live trading workflows. Some modelling can run outside latency-sensitive environments but workloads that feed real-time execution have to come back into the same colocation sites already under strain.
High-frequency trading will still reward faster compute and lower latency. Increasingly, however, those gains will depend on whether firms can keep infrastructure stable when markets move fastest. Cooling is becoming part of the execution stack, and firms that act on that shift will be better placed to scale performance inside the sites that count.
