The data center industry today is markedly different from what it was a decade ago, largely due to a confluence of real-world factors over the past few years: the widespread adoption of AI technologies, the slowing pace of Moore’s Law, and the pressing challenge of sustainability.

Uptime Institute predicts that, as operators focus on and plan for challenges related to power supply, cooling, management, high-density deployments, and regulatory compliance, the data center industry as a whole will experience another wave of significant transformation by 2025.

Although it did not top Uptime’s ranking, given the widespread attention on AI, we’ll start there. Over the past 12 months, major cloud service providers and hyperscale infrastructure operators have rapidly deployed sizable GPU clusters. Uptime estimates that NVIDIA alone sold 600,000 H100 GPUs in 2023—though we believe the actual figure may be closer to 710,000. By the end of 2023, the chipmaker’s GPU shipments are expected to climb even further, to between 1.5 million and 2 million units.

There’s no need to worry—building AI infrastructure is not as complicated as many people imagine.

However, it must also be acknowledged that, in the face of such large-scale deployments and the seemingly insatiable market demand for generative AI–enabling technologies, the data center industry is—and must—be prepared to address the surge in demand, particularly the thermal and power-consumption challenges posed by the deployment of large-scale GPUs and other accelerators.

While HPC professionals are no strangers to ever-increasing accelerator performance and power density, the new facility pushes both of these metrics to unprecedented levels compared with typical dual-socket systems.

NVIDIA’s H100 and the upcoming H200 both exceed 700 watts in rated power—yet that figure represents the power consumption of a single chip. An AI cluster typically deploys four to eight GPUs, pushing the total thermal design power straight into the kilowatt range.

However, Uptime estimates that the AI infrastructure boom will have only a limited impact on most operators. This is primarily due to ongoing chip supply constraints and the relatively small number of companies capable of mobilizing large-scale deployment resources.

In any case, data centers that deploy such systems on a large scale will inevitably face dual challenges in power supply and thermal management. Fortunately, several approaches can address these specific issues, with one of the simplest being to distribute the system across a larger footprint—this approach also imposes the least disruption to the facility’s operating environment.

For example, suppose the existing infrastructure can support 25 kilowatts of power per rack along with the corresponding heat load; in that case, the operator might choose to distribute DGX nodes across twice as many racks. While this would naturally result in substantial underutilized space within the cabinets, for certain workloads—provided that the cost of additional space is not prohibitive—it may actually be the simplest and easiest-to-implement option.

However, as we previously learned from Chris Sharp, Chief Technology Officer at Digital Realty, while disaggregated systems do address heat dissipation and power delivery challenges, they are less well suited to training workloads that rely on dedicated interconnect fabrics. For instance, NVLink has a limited reach, so it works best when paired with higher-density deployment configurations.

Direct Liquid Cooling Shows Its Advantages

The second option is to upgrade facilities to liquid cooling, particularly direct liquid cooling (DLC) designs. Uptime analysts predict that, under mounting pressure from chip heat dissipation, system density, and sustainability concerns, operators will deploy DLC solutions more widely by 2025, trading short-term installation convenience for enhanced hardware performance.

Direct liquid cooling generally outperforms traditional air cooling in terms of efficiency, as liquids are far better thermal conductors and the technology largely eliminates the need for cabinet-mounted fans. According to our understanding, adopting liquid cooling can reduce system power consumption by up to 20%. However, Uptime also notes that because liquid-cooling power consumption is often aggregated with overall IT-system power consumption, precise quantification is extremely challenging.

Moreover, realizing the energy-saving potential of direct liquid cooling is far from as simple as mere lip service. Uptime explains that many facility operators tend to maintain the coolant at a lower temperature in order to enhance cooling performance for the infrastructure. As we understand it, this approach can reduce the design load on the infrastructure, thereby extending the service life of IT hardware. However, from an energy-efficiency perspective, this practice is less effective than using room-temperature coolant, since it requires additional electricity to actively cool the fluid down.

It must be acknowledged that direct liquid cooling at low temperatures does offer performance advantages. Lower coolant temperatures ensure that the CPU and accelerators consistently operate at lower junction temperatures, enabling them to sustain higher overclocking levels—and correspondingly higher power consumption—for extended periods.

The real concern is that the cost savings achieved through direct liquid cooling may be offset by higher system power loads, rendering the retrofit unprofitable.

Achieving sustainability requires taking a new approach.

Uptime also points out that direct liquid cooling is unlikely to contribute to the achievement of sustainability goals, while imminent regulatory requirements leave operators with no alternative.

In principle, every major cloud service provider and hyperscale data center operator has made a net-zero emissions sustainability commitment over the past few years. As for tech giants such as Microsoft and Google, they are now just a few years away from achieving their first major milestone.

Uptime predicts that data center operators will face a challenging period if they are truly committed to delivering on their promises. Because data center deployment sites do not always have abundant renewable energy resources, the integration of new energy sources such as wind, solar, and tidal power often fails to make a meaningful difference.

Moreover, governments around the world have been urging data centers to enhance transparency regarding their electricity consumption and carbon footprint.

The EU Corporate Sustainability Reporting Directive, adopted last September, along with California’s Climate Corporate Data Accountability Act and other similar regulations, will soon require more companies to report their carbon emissions and climate-related risks arising from business operations.

Uptime reports that the U.S. Securities and Exchange Commission (SEC) has also begun to prioritize this issue, requiring large publicly listed companies to disclose certain emissions data in their quarterly reports.

Undoubtedly, the most stringent of the regulatory requirements is the EU Energy Efficiency Directive issued last autumn. This document sets out reporting obligations for data centers and other IT and network operators. However, it is important to emphasize that the directive aims to collect data on usage patterns and does not yet impose prescriptive operational requirements on data center facilities.

Although these regulatory requirements are well-intentioned, Uptime’s survey reveals that fewer than half of the data center operators surveyed have actually begun tracking metrics such as carbon emissions.

Intelligent upgrading of data centers has become imperative.

For many years, Uptime has been advocating the widespread adoption of data-driven automation in the data center industry. Analysts believe that 2025 may mark the milestone when this goal is fully realized.

The root of the problem lies in the fact that, despite fundamental changes in data center hardware, the evolution of management tools has stalled. Most building management systems (BMS) and data center infrastructure management (DCIM) software still offer relatively limited automated analytics capabilities.

It is easy to foresee that, with only moderate improvements at the management level, operators can significantly enhance energy efficiency while also lowering the compliance threshold for subsequent regulatory oversight and emissions reporting requirements. A typical use case for automating system operations in this context is adjusting environmental systems during periods of low demand, thereby avoiding the wasteful consumption of electricity to cool idle systems that are not operating at high intensity.

Uptime also believes that more advanced automation technologies will leverage AI models trained on facility data sets to flexibly and predictively optimize data center operating modes.

Applying AIOps-like capabilities to the data center as a whole undoubtedly offers numerous benefits; however, Uptime analysts remain pessimistic about the ability of incumbent DCIM software vendors to adapt. Instead, they anticipate that such capabilities will first emerge from a new generation of startups. Uptime is currently monitoring six such companies at various stages of development and believes their R&D efforts hold promise for addressing the challenges of intelligent infrastructure operations.

Although the report does not specify the names, we suspect that one of them is likely Oxide Computer. The company was co-founded by Bryan Cantrill, a former software engineer at Sun Microsystems, and Steve Tuck, president of Joyent, and focuses on rack-level computing scenarios. It has even developed its own board management controller (BMC) in-house to avoid relying on industry-standard controllers such as those from Aspeed.

Ultra-Large-Scale Facility Parks May Become the Norm

A range of emerging trends, particularly the increasing demand for higher compute density to support the growing uptake of AI, are driving operators to invest in building hyperscale server campuses composed of multiple data centers.

According to Uptime, these data center campuses can span several million square meters, are designed to meet the power and connectivity needs of multiple tenants, and often prioritize the use of greater amounts of clean energy from the outset.

The defining characteristic of such new data center campuses is the achievement of gigawatt-scale capacity. While these campuses, as a key concept in facility planning, will not initially operate at such power levels, they are likely to incorporate design margins to accommodate future expansions and avoid capacity bottlenecks over the lifecycle.

Last year, we also witnessed numerous more ambitious exploration initiatives, including the use of emerging energy sources such as hydrogen fuel cells and small modular reactors to deliver power on the order of several gigawatts.

Moreover, beyond the advantage of shared power infrastructure, competing data center operators may also choose to collaborate for another reason: low-latency communication between facilities.

Uptime’s trend forecasts for these data centers—which may in the future be more appropriately termed “data cities”—will help reduce colocation and connectivity costs, enhance facility resilience, and improve operational sustainability.

Whether these predictions will ultimately come to fruition remains to be seen, but one thing is certain: data centers are poised to continue their relentless march toward larger scale, greater capacity, and higher energy consumption.