By Tom Traugott, SVP of Strategy at EdgeCore Digital Infrastructure

As we saw in 2024, AI is rapidly transforming what we consider "the modern data center." Facilities require significant upsizing to meet the high-power and density requirements of AI and GPUs, and the industry must be prepared to support continued innovation.

Guided by AI, the data center sector is entering another year of unprecedented growth and evolution. However, for the industry to best meet the heightened demands of AI and hyperscalers (the world's largest cloud and internet companies), we can expect developers, operators, and owners to be laser-focused on several fundamental areas.

Data centers must be designed with resiliency at the core

Considering the high-power requirements of AI and GPU clusters, everything is now larger and denser, which introduces new challenges to designing the modern data center. Facilities and campuses over the next 5-10 years will need significant upsizing, from 100-300 megawatts for most hyperscale campuses today to the gigawatt level corresponding with 1M+ projected GPU clusters.  2025 will be the year developers more broadly plan and advance data center designs to meet those requirements.

Nonetheless, AI data centers must still be designed with uptime in mind to ensure GPUs don't suffer outages at the wrong time, mid-training or during an inference call. While some degree of failure is expected in large GPU clusters, that is at the hardware level. A data center outage has the potential to be catastrophic and damage very expensive GPUs in a thermal runaway event.

The largest AI and cloud companies still provide their services with underlying always-on service level agreements which flow through to the data center providers supporting them. New data center entrants, predominantly from the crypto mining sector, appear to be developing data centers faster, yet with lower resiliency - making them a risky match for AI and cloud deployments.

While data center investment is an increasingly smaller proportion of overall infrastructure spending over time (as related to hardware/networking for GPUs), building to lower resiliency is like putting low octane gasoline in a high-performance sports car - it may work in a pinch, but runs the risk of killing the engine eventually.

AI has already added pressure to data center developers and operators regarding design - and that trend will continue into 2025 and beyond. The road toward the modern data center is much more sophisticated, housing new tech components like liquid cooling methods.

Data centers need to be ready for liquid cooling

The next-generation GPUs will need over 100 KW of power in a single rack - beyond what air cooling technology can typically manage. Closed-loop, direct-to-chip liquid cooling is necessary to support this density level, enabling closely coupled and denser computing systems.

This cooling method will become more predominant in 2025 and will likely be considered the industry standard for state-of-the-art facilities by 2026. Over the next year, we will move from simulated performance for the latest GPUs to real world deployment and testing to determine the best cooling methods for supporting ever-denser racks. Investments in liquid cooling will continue to grow given that densities are only expected to increase as chip transistor density increases further and faster (as seen with NVIDIA's shift to a one-year versus a two-year innovation cycle). 

Data center providers can prepare by getting the big parts right - ensuring adequate gross tonnage of chilled water that can support both air-cooled and liquid-cooled solutions for large quantities of megawatts. That said, the largest buyers of GPUs, the hyperscalers, will need to take a leading position to both finalize their designs and back particular liquid cooling methods so the industry can distribute resources effectively and meet rapidly increasing densities at the rack level.

Constructing modern data centers isn't a one-and-done fix - they require a brand-new approach, demanding technology upgrades, campus design pivots, strategic planning and most importantly, capital investment. 

Access to capital will remain critical for AI data center development

In 2025, whether a veteran or new entrant in the market, raising the necessary capital to design, develop, and operate AI-ready data centers will remain critical to meeting the next phase of growth, especially as requirements continue to get larger and denser. A challenge for new entrants will be to ensure that investors have sufficient confidence in a platform's track record to provide the necessary capital to guarantee certainty of delivery and execution, not to mention ongoing operations. 

This next growth phase will also depend on the development of new power generation, transmission, and distribution in support of data centers, which will only increase with AI. The industry needs to see private capital more meaningfully enter the power generation, utility, and grid reinforcement arena. However, many utilities aren't structured to make the entrepreneurial investments required. Since 2000, their fiduciary focus to investors has been to provide stable dividend growth and protection more so than opportunistic returns.

The easy megawatts have been found and billions of dollars of new investment is needed. Our industry will share those costs, but new sources of risk-taking capital are also required. Next year we will likely see more creative joint ventures, spinoffs emerging from utilities, or more Independent Power Producers (IPPs) emerge to help support continued development. The Liquefied Natural Gas (LNG) and data center industries have found a way to sleeve tens of billions of dollars in recent years to fund long-term projects. Moving into 2025, the data center industry requires new ventures and structures to get the capital needed, especially on a national level.

AI and power infrastructure investment and expansion remains a national priority

If AI compute scaling maintains its current growth trajectory, GPU clusters will surge in size from 100K+ to 1M+ clusters, reaching gigawatt scale before 2030. The U.S. is currently the leader in AI globally, maintaining a computing, chip, and technology advantage over its nearest rival, China, with the U.S. having approximately 2X the number of installed computing servers as China.

However, since 2000 China has outpaced the U.S. in terms of adding power infrastructure (adding 925 GW of generation via the U.S.'s increase of 51 GW) primarily in support of its manufacturing base but readily able to pivot to support data center infrastructure.  For the U.S. to maintain its advantage, power infrastructure investment needs to materially expand to the 100+ GW range.

Thankfully, this appears to have become a bipartisan area of political concern, and I believe the prioritization around national economic and security interests will help accelerate infrastructure development. However, a question remains: will investment be fast enough to maintain the U.S. technological advantage or will innovation be bottlenecked due to capital or regulatory constraints?

There's no question that 2025 will be a pivotal year for the data center industry, exacerbated by AI innovation. Next year will be all about strategically planning for AI's expansion via designing and perfecting the new "modern data center" - one that is larger, denser, and can accommodate emerging supporting technologies like closed-loop, direct-to-chip liquid cooling. Continued and new sources of capital are required for data centers to keep pace with AI and for the U.S. to maintain its advantage. In 2025, expect developers, operators, and owners to come together to propel AI innovation throughout the year and into the next decade.

##

ABOUT THE AUTHOR

Tom leads EdgeCore's strategy and activities around client leasing and marketing initiatives focused on client-centric growth solutions for top hyperscale and technology companies.

Tom brings to EdgeCore 19 years of experience in wholesale data centers, enabling him with valuable insight into client needs for tailored data center developments and tailored commercial models designed to enable EdgeCore clients to ensure scalability, reduce cost, and maximize flexibility. Supporting these objectives, Tom leverages his experience in site selection, acquisition, development, and leasing supporting hyperscale CSPs and information-intensive businesses, including completion of transactions totaling over 600+ MW and valued at over $1.8B. Prior to joining EdgeCore, Tom worked for Amazon Web Services, where he was accountable for strategy and execution for new and existing regions across EMEA, APAC, and the Americas, along with diligence and strategy for additional regions under evaluation. At AWS, Tom worked in a multi-stakeholder environment executing on accountabilities to scale data center infrastructure planned to exceed $1B in construction investment. Prior to Amazon, Tom was co-practice leader of Cassidy Turley's (now Cushman & Wakefield's) Data Center Advisory Practice, focused on end-user representation and capital markets transactions. Tom previously worked with many members of the EdgeCore team at CoreSite Realty Corporation as regional VP of Sales.