AI is rapidly becoming embedded in business operations, consumer applications, and critical infrastructure worldwide. As adoption accelerates, the computational demands behind it are growing at an equally dramatic pace, placing enormous strain on the data centers that power it all.
Google has raised the bar for AI data center cooling with Project Deschutes, an advanced coolant distribution unit (CDU) framework built on open-source standards that supports both direct-to-chip and rack-level solutions. But for manufacturers looking to meet its design requirements, ultra-low harmonic (ULH) drives – which control the pumps that circulate coolant through the system – are essential.
The heat is on
As the AI race intensifies, providers such as OpenAI and Anthropic are pushing the limits of compute, placing pressure on infrastructure providers to scale capacity quickly and reliably. This has fueled a surge in new data center construction, with firms such as Google, Microsoft, and Meta packing facilities with increasingly powerful processors. More power inevitably means more heat, and that heat must be managed effectively. Any performance gains will quickly turn into reliability risks without robust cooling.
Traditional data centers already generate significant heat, but AI processors push existing cooling systems even further. Filling racks with high-performance processors creates an environment comparable to rows of industrial ovens running simultaneously at maximum heat. If heat accumulates unchecked, sensitive components fail and force costly shutdowns. Without major improvements in cooling technology, progress in AI development could slow or even stall.
Recognizing these rising challenges, Google unveiled Project Deschutes in late 2025, reshaping industry expectations while establishing stricter requirements for the components that support these systems.
Form follows function
Project Deschutes introduces a new approach to CDU structure, built around openness in both form and philosophy. Physically, the system uses an open-frame design, allowing technicians to easily access components for maintenance or replacement.
The design mirrors the accessibility of a Formula One car, exposing every part for quick servicing and supporting uptime and efficiency. Technicians can access parts directly without removing panels or working inside tight server cabinets, which becomes increasingly important at the scale and performance levels modern AI infrastructure demands.
Equally important is the openness of the specification itself. Google released the Deschutes design through the Open Compute Project (OCP), making it available for any manufacturer to produce and sell. This marks a departure from its earlier, closely guarded, CDU designs. In doing so, it has provided the industry with a shared reference point for next-generation CDU development.
Liquid cooling remains central to this evolution and Google has relied on this method since 2018 – using water or dielectric fluids that can remove heat up to 3,500 times more efficiently than air. By 2024, Microsoft and Meta had also shifted toward liquid-cooled server designs, and Project Deschutes builds on that momentum.
High-spec cooling systems for the AI era
At its core, a Deschutes CDU functions like a huge, powerful heart, continuously pumping liquid coolant through the server to manage extreme thermal loads. Each unit can cool a two-megawatt AI server, well beyond today’s typical requirements. In fact, Google, Meta, and Microsoft jointly announced plans for the first one-megawatt server.
Meeting these demands requires strict engineering standards. Components must operate reliably at elevated temperatures, such as sustained ambient conditions of 50°C, without losing capacity. Filtration and fluid quality are equally critical factors. Cooling fluids must be filtered to remove particles larger than 0.2 microns to protect sensitive chips.
The Project Deschutes standard’s attention to detail also extends to the variable speed drives controlling the coolant pumps. These drives must provide advanced monitoring capabilities while maintaining total harmonic current distortion (THDi) below 5 percent, a threshold met by ULH variable speed drives (VSDs).
Google is aiming for near-continuous uptime (99.99%) that’s supported by predictive maintenance and detailed system monitoring, making reliability a core objective of the standard. To achieve this, VSDs must provide granular data on factors such as torque, DC power levels, and insulated gate bipolar transistor (IGBT) temperature to the programmable logic controller. This data enables early detection of potential issues before they impact operations.
Managing harmonic distortion is particularly critical in this environment. Conventional drives draw current in a non-linear manner rather than following a smooth sinusoidal pattern, creating distorted waveforms known as harmonics. Because they cause components to run hotter and less efficiently, such disturbances can affect power quality and disrupt sensitive systems in data centers. ULH drives address this by drawing current smoothly, generating minimal electrical noise.
Shaping tomorrow’s AI data centers today
Looking ahead, it’s likely that cooling technologies will continue to evolve alongside AI workloads. As compute demands grow, so too will the need for more efficient and scalable heat management. Open initiatives like Deschutes may serve as a foundation for future development, providing a shared starting point for addressing increasingly complex challenges.
And with the specification now public, manufacturers have a clear path to developing compliant CDUs. Because the requirements make it evident that achieving the necessary stability and efficiency is far easier with ULH drive technology integrated into pump systems, these standards are likely to shape the broader direction of AI infrastructure as adoption grows.
Meanwhile, innovation continues at a rapid pace. As open cooling standards continue to develop, the next generation of thermal management solutions is already taking shape. Technologies that meet today’s demanding specifications will need to evolve again. The manufacturers and component suppliers best placed to respond will be those already building to the most demanding specifications available.
