Since the GPT-4o family of models broke the one-trillion parameter barrier in April 2024, the need for data centers to meet evolving AI needs to both scale out (larger clusters and campuses) and scale up (higher power density per rack) is greater than ever and is rewriting the playbook for data center design.

Against this backdrop, direct-to-chip liquid cooling (DLC) has crossed the chasm from niche, high-performance computing labs to mainstream production. What looked ambitious in 2023 is the desired specification for supporting AI leading-edge workloads in 2025, and it will become the minimum specification for even denser GPU servers in 2026.

Data center infrastructure must be able to handle GPUs' increasing density requirements while maintaining the ability to pragmatically support the revenue-generating and operationally live hardware of the last few years.

For instance, Nvidia’s five-year-old A100 chip technology remains in use today. Current leading-edge GPUs need over 125 kW of power in a single rack alone – well beyond what traditional air-cooling technology can support. Next year’s GPUs are expected to break the 200 kW per rack mark, trending towards 300 kW per rack.

But, we’re not done – at the 2025 OCP EMEA Summit, Google revealed its Project Deschutes design for a 1MW rack, notable for its lack of blinking lights and more so for its resemblance to a car engine under the hood. This is only possible with DLC, which paves the path for continued densification and thereby innovation at the AI hardware level.

From data center tail wagging to server tail wagging

DLC is one of the hottest considerations at this stage of AI advancement, enabling closely coupled and denser computing systems to thrive. While it’s not a new technology, most deployments of DLC before this year have been in niche environments, such as national laboratory supercomputers.

For years, hyperscalers were resistant to upgrading the gigawatt-scale fleets of data centers optimized for air-cooled efficiency, which kept pressure on CPU and GPU server manufacturers to remain within air-cooled thermal limits of 30-40kW per rack. At the same time, Nvidia had 100+ kW rack designs, but they weren’t selling at scale until generative AI drove broader use cases for the exponentially higher horsepower of larger GPU clusters. The data center “tail” was wagging the server “dog.”

Google was the exception to the hyperscaler orientation around air cooling. They began implementing DLC in their data centers starting in 2018 in support of their TPU chip architecture. This is unsurprising, given Google’s innovations are at the leading edge of ML and generative AI in support of their dominant search business. In developing systems for their own use, Google didn’t have the obstacle of making those chips sellable to customers and, therefore, likely has the largest fleet of DLC at multi-gigawatt scale today.

As Nvidia’s market share for AI chips became dominant in their own right (estimated at 70-95 percent), by 2024, they finally had the leverage to press for advancements at the data center level and assert in their reference architecture for the GB200 NVL72 rack system where DLC would be required (with lower density systems still able to live in air-cooled environments).

The table thereby turned – instead of working backwards from the thermal limits of most data centers for scalability, Nvidia was drawing a new line. In order to support their leading-edge hardware, DLC was a must-have, not just a nice-to-have.

Looking at Nvidia’s density roadmap alone, the projection of liquid cooling demand will rapidly accelerate, seeking to double per annual product innovation cycle. Blackwell GB300s will reach a peak rack density of 163kW by the end of this year, Vera Rubin NVL144 racks may require 300+ kW in 2026, and by 2027, the Rubin Ultra NVL576 rack may push above 600 kW per rack – with further densification expected for 2028 and beyond. Google announcing 1MW rack designs points to the likelihood that Nvidia will get there as well.

What does DLC mean for Nvidia? DLC-enabled data centers will become the landing place for ever more powerful accelerated computing systems and escalating power densities. If the facility liquid is present at scale, getting to 500+ kW per rack is about distribution, not capability. Furthermore, a DLC-enabled facility can still typically support air-cooled hardware as well, but the opposite won’t be true for much longer, with air-to-liquid-cooled solutions reaching their limits this year (closer to the 100kW per rack range).

DLC adoption and investment outlook

Last year, the global DLC market size was valued at $1.85 billion. Over the next decade, it’s projected to reach $11.89 billion, primarily driven by DLC’s ability to handle increases in chip density.

DLC's value to data center environments will amplify this year, and by 2026, it will likely be considered the industry standard for state-of-the-art facilities. During this timeframe, the data center sector will move from simulated performance for Nvidia’s latest GPUs to real-world deployment and testing. This shift will be essential in determining the most effective cooling methods for supporting higher-density racks.

Densities are only expected to increase alongside chip transistor density growth, as seen with Nvidia’s shift from a two-year innovation cycle to a one-year cycle. Knowing this, we can expect investments in DLC at the data center and hyperscaler level will likely skyrocket over the next few years.

Promising growth calls for industry preparation

The full benefit of DLC cannot be achieved without considering the necessary design changes across future data center infrastructure. Campus designs must be specifically planned to support DLC as a core component.

Due to inefficiency, existing data centers with current air-to-liquid cooling deployments will likely need to be retrofitted, and in many instances of older deployments, it will be physically impossible to meet the high demands of DLC. Regardless, the lift to support DLC will be high, and ultimately, these older campuses will have hard caps on scalability.

With density requirements continuing to climb from regular campus densities of 300MW to figures well north, DLC deployments will be best suited for these purpose-built data centers specifically designed to serve DLC’s unique energy and cooling requirements from the outset. Greenfield, purpose-built data centers will be able to support the newest technologies as well as predicted future scale and densities. Consequently, DLC capability will be one of the key requirements for the “AI factories” frequently mentioned by Nvidia.

Data center developers can start preparing now by getting the “bigger” parts right. For instance, as GPU deployment advances, they should focus on ensuring adequate gross tonnage of chilled water to support air- and liquid-cooled solutions.

As the largest buyers of GPUs, hyperscalers will also need to continue preparing for this next cooling era. Since hyperscalers develop their own designs for DLC, they will need to take a leading position to finalize designs. Preparations at the hyperscaler level will allow the data center industry to distribute resources better and meet rapidly increasing densities at the rack level.

Embracing the next generation of data centers

With Nvidia leading the GPU race and Google remaining at the cutting edge for AI and ML, their mutual embrace of DLC solutions serves as baseline points for where the market should be.

Unlocking the future of AI doesn’t come without looking closely at data center design and finding ways to get the most out of critical systems, like cooling. AI transforms everything at the data center level, and with direct-to-chip liquid cooling here to stay – and expected to grow as a critical market segment, embracing its potential is essential for long-term success.