AI data center reconstruction enters the era of power constraints: from GPU stacking to the 800V DC infrastructure race

AI Data Center Reconfiguration Enters the Era of Power Constraints: From GPU Stacking to the Infrastructure Race for 800V DC

AI infrastructure is entering a new stage: the bottleneck is no longer just chip supply, but power, cooling, and distribution capacity. According to the Los Angeles Times’ reporting on the industry, as vendors such as Nvidia continue to release higher-performance GPUs and rack systems, the power density of AI data centers is rising rapidly. Traditional data center architectures designed around CPUs and general-purpose workloads are already struggling to meet the demands of the new generation of AI training and inference workloads. The report also notes that companies including Nvidia, CoreWeave, Google, Vertiv, and GE Vernova are promoting solutions such as liquid cooling, simplified power distribution, and 800V DC systems to reduce losses and increase compute output per unit area.

What makes this shift important is not merely that “larger machine rooms” are emerging, but that the design logic of enterprise IT infrastructure is being rewritten: future cloud platforms, private data centers, and hybrid cloud environments may all have to answer “how much power is available, how will it be cooled, and how will it connect stably to the grid” before asking “how many GPUs can be installed.” For CTOs, CIOs, and enterprise architects, this means the feasibility assessment of AI projects is being upgraded from a software and procurement issue to a coordinated energy and infrastructure issue.

Technical Analysis: Why AI Data Centers Are Different from Traditional Data Centers

Traditional data centers mainly handle enterprise applications, databases, storage, web services, and office collaboration workloads. These tasks are typically CPU-centric, and rack power density is relatively manageable. Industry data cited in the report indicates that a standard server rack may require only 25 to 40 kilowatts, whereas AI racks, due to the dense deployment of GPUs, are seeing power demand rise rapidly. As model scale, distributed training, and inference throughput continue to grow, the power requirements of a single rack have already jumped from earlier low-density stages to the hundreds-of-kilowatts level, and may even approach the megawatt level in the future.

The key change behind this is that AI is not a problem that can be solved simply by “running a few more virtual machines.” Training and inference depend on highly parallel GPU clusters and require higher-bandwidth networks, faster memory access, and more stable power delivery. In other words, AI infrastructure is a systems engineering effort in which compute, networking, and power are co-designed. If any one link falls behind, overall performance will be constrained.

The report also highlights a commonly overlooked fact: a significant portion of the electricity entering a data center is not directly used for computation, but is consumed by cooling systems, long-distance transmission, and multiple stages of voltage conversion. As power density rises, these losses will be amplified further. As a result, the industry is beginning to extend its optimization target from “server efficiency” to “end-to-end efficiency from the grid to the chip.”

Enterprise Impact Analysis: AI Infrastructure Budgets Are No Longer Just About Compute Procurement

For enterprise users, this trend will directly affect the structure of CAPEX and OPEX.

First, capital expenditures are rising and becoming more front-loaded.First, capital expenditures are rising and being brought forward. Traditional IT projects typically revolve around servers, storage, and software licenses, but AI data centers require simultaneous investment in power distribution equipment, liquid cooling systems, facility retrofits, backup power supplies, and higher-grade grid connection capabilities. In other words, before an enterprise deploys an AI platform, it may first need to invest in the “infrastructure for infrastructure.” This raises the initial barrier to entry, especially for companies that want to build their own inference platforms or private AI clouds.

Second, operating expenses are expanding from electricity costs to energy management. As rack power density increases, electricity prices, peak load, cooling efficiency, and grid connection stability all become long-term cost variables. The report mentions that Nvidia is working with energy-saving software companies such as Emerald AI to try to avoid placing excessive stress on the power grid during peak usage periods. This shows that AI infrastructure operations are increasingly becoming a kind of “energy scheduling” task, rather than just IT operations and maintenance.

Third, deployment pace is constrained by power and approvals. Even if a company has the budget, it may not be able to scale up immediately. Grid capacity, substation facilities, land conditions, local regulations, and community backlash over data center electricity consumption can all delay delivery. For companies that depend on the speed of AI capability rollout, this means the “compute procurement cycle” may become longer again, and planning windows must start earlier.

Fourth, security and compliance requirements are rising at the same time. When companies deploy data centers in a distributed manner to get closer to energy sources or renewable power, cross-regional data governance, sovereign cloud compliance, backup strategies, and disaster recovery design all need to be reassessed. AI platforms are no longer just compute pools, but complex systems that span power, network, and regulatory boundaries.

Market Competition Analysis: Cloud Providers and the Infrastructure Supply Chain Are Being Reshuffled

This change is not happening only inside server rooms; it is reshaping the entire supply chain.

On the cloud provider side, the core of competition is shifting from “regional coverage” to “power availability.” AWS, Azure, and Google Cloud used to compete on the number of global regions, network quality, and service ecosystems, but in the AI era, whoever can secure stable power faster, deploy high-density racks, and provide liquid cooling and GPU clusters is more likely to win high-value AI workloads. For enterprise customers, the criteria for choosing a cloud platform will also change: it is no longer enough to look at price and features; they also need to consider whether there is sufficient AI capacity, whether high-density deployment is supported, and whether the energy and cooling infrastructure is mature.The collaboration between chip and server vendors is also becoming closer. Reports note that Nvidia is pushing for higher-integrated chip-and-rack systems and testing device designs that simplify the power delivery path. This means chip makers are no longer just selling accelerator cards; they are defining data center architecture standards. For server and systems integrators such as Dell, HPE, and Supermicro, this is both an opportunity and a source of pressure: if future AI racks become increasingly standardized and modular, then whoever can first provide complete solutions around the new power supply and cooling systems will be more likely to secure project entry points.

The importance of power and energy equipment companies is rising. Traditionally, the “invisible layer” of data center infrastructure consists of UPS, power distribution, cooling, and substation equipment. As solutions such as 800V DC, solid-state transformers, and fewer conversion stages are being explored, the role of infrastructure suppliers like GE Vernova and Vertiv becomes even more critical. For data center operators, this is not just about purchasing equipment, but about choosing the architectural path for the next decade.

Industry trend observation: AI-native data centers are taking shape

In the long run, this round of change may lead to four trends.

1. AI-native cloud will replace the idea of “general-purpose cloud plus AI.” In the past, cloud platforms added GPUs on top of existing server architectures; in the future, AI cloud may be designed from the ground up, starting with power, cooling, networking, and rack layout. AI workloads will become the starting point of architecture, rather than an add-on feature.

2. Liquid cooling will shift from an optional item to a mainstream configuration. As GPU power density continues to rise, the marginal effectiveness of air cooling will decline. Liquid cooling is not simply an energy-saving tool, but a prerequisite for sustaining high-density computing. For enterprises, data room design and equipment selection must include liquid cooling in standard planning.

3. DC power supply and fewer conversion layers will become an important direction. Reports note that multi-stage conversion from the grid to the chip causes energy losses, and the industry is exploring power delivery paths with fewer steps as well as 800V DC solutions. If these mature, the energy efficiency, space utilization, and renewable energy integration capacity of future data centers may all improve.

4. Energy constraints will affect the geographic distribution of AI deployments. When grid capacity becomes a critical resource, AI infrastructure may not necessarily remain concentrated in traditional cloud regions, but may instead move toward areas with more abundant power, faster approvals, and a better energy mix. For enterprises, this will affect data residency, disaster recovery planning, and multicloud strategies.

Reference points for enterprise decision-making: what should be watched now

• For enterprises planning to expand AI capabilities, the most important thing right now is not to blindly pursue more powerful GPUs, but to establish a “feasibility assessment framework for AI infrastructure.” At minimum, it should include the following dimensions:- Compute demand forecasting: how many GPUs and how much rack density are needed for training, fine-tuning, and inference respectively;
• Power capacity assessment: whether existing data centers and target cloud regions can support growth over the next 12 to 24 months;
• Cooling solution fit: whether liquid cooling is needed, and what the retrofit timeline and cost would be;
• Deployment model selection: whether public cloud, colocation, private cloud, or hybrid cloud is more suitable;
• Compliance and sovereignty requirements: whether data flows across regions and whether industry regulations are involved;
• Peak cost control: how to balance electricity prices, expansion, and utilization.

From this perspective, AI infrastructure budgets are becoming a “hard constraint expense” in enterprise digital transformation, and their priority may gradually approach that of networking, security, and core ERP systems.

CloudTechDaily Insight

The most important significance of this industry shift is that it reveals the true underlying logic of cloud computing in the AI era: competition no longer happens only in software features or model parameters, but in power, cooling, and infrastructure design. For enterprise IT strategy, AI capabilities are no longer a service that can be “bought and used right away,” but a long-term project that requires coordinated support from the power grid, data centers, cloud regions, and operations systems.

Over the next five years, the truly leading cloud platforms and enterprise IT organizations will not necessarily be the ones with the most GPUs, but the ones that can integrate compute power, energy, and operations into a stable delivery capability. In other words, the next winners in AI infrastructure will likely not be simply chipmakers or cloud providers, but enterprises that can redesign data centers into an infrastructure system that is “power-first, AI-native, and sustainably operated.” For CTOs and CIOs, power and cooling should now be incorporated into the AI roadmap; otherwise, the bottleneck in the future may not be the model, but the racks that cannot be powered on.