Direct-to-chip cooling — once a specialist technique reserved for high-performance computing labs — has become the default design assumption for any facility intended to host current-generation AI accelerators. Roughly two-thirds of the liquid cooling market today runs on this approach, and for good reason: it offers a practical, increasingly standardised path to managing the heat output of modern GPU platforms without the operational complexity of full immersion.
How the System Actually Works
In a direct-to-chip, or cold plate, system, coolant never directly contacts electronic components. Instead, a metal cold plate — typically a copper micro-channel block — is mounted directly onto the CPU or GPU package, conducting heat into a secondary coolant loop that circulates within the rack and data hall. A coolant distribution unit manages the interface between this secondary loop and a primary facility loop, which ultimately rejects heat to the outside environment through dry coolers or cooling towers. This isolation between the fluid touching the chip and the fluid exposed to the wider facility is central to making liquid cooling commercially practical at scale — it limits the consequences of a leak and allows facility-side equipment to be serviced independently of the IT environment.
Why 45°C "Warm Water" Changes the Energy Equation
One of the more consequential recent developments is the move toward higher coolant supply temperatures. Current-generation GPU platforms are increasingly specified to support liquid cooling at supply temperatures around 45°C — warm enough that heat can be rejected through dry coolers using ambient air, rather than energy-intensive mechanical chillers, across a much wider range of climates and seasons.
- Eliminating or reducing mechanical chiller reliance is one of the most significant available levers for improving facility PUE
- Warm-water operation extends the number of hours per year that free cooling is viable, even in temperate and warmer climates
- This shift has been enabled by hardware manufacturers explicitly designing accelerators to tolerate higher coolant temperatures, rather than facilities working around colder hardware requirements
The move to warm-water direct-to-chip cooling did not happen because facilities demanded it — it happened because hardware manufacturers redesigned chips to make it possible.
What This Means for Facility Design
Designing for direct-to-chip cooling from the outset, rather than retrofitting it, materially simplifies a project. It allows piping headers, coolant distribution unit placement, and floor layout to be planned around the technology rather than worked around an existing air-cooled design. It also requires earlier coordination with server and rack original equipment manufacturers, since coolant connection standards, quick-disconnect fittings, and flow rate requirements vary by platform and are not yet fully standardised across the industry.
Facility operators also need to budget for water quality management — treated, filtered coolant loops require ongoing monitoring to prevent fouling or corrosion that could degrade heat transfer performance or, in the worst case, damage IT equipment over the facility's operating life.
Standardisation Pressure Is Building Across the Supply Chain
One of the more important developments to watch is the push toward greater standardisation of coolant connections, fitting types, and flow rate specifications across server and rack manufacturers. Today's relatively fragmented landscape — where coolant connection standards can vary meaningfully between hardware vendors — adds engineering overhead and limits interchangeability for facility operators managing mixed hardware fleets. Industry bodies and major hardware manufacturers are increasingly aligning around common reference designs, which should reduce integration complexity for facilities over the next several hardware generations, much as standardisation in rack form factors and power connectors simplified earlier generations of data center design.
The Maturity Curve Is Moving Quickly
What was a bespoke engineering exercise just a few years ago is rapidly becoming a more standardised design pattern, with established vendors offering increasingly interoperable cold plate, CDU, and piping solutions. This maturation is good news for developers: it reduces the technical risk associated with adopting direct-to-chip cooling and shortens the learning curve for operations teams who have not previously managed liquid-cooled environments.
DATAPERT's engineering teams support clients through direct-to-chip cooling design and integration as part of our broader data center development and technology integration services. Start a project to discuss cooling strategy for your next AI-ready facility.
