Data center power design has always been built around predictability. Training workloads consume power at a known rate for a known duration. Inference workloads, while variable, follow demand curves that capacity planners can model with reasonable confidence. The entire discipline of data center power architecture is built on one assumption: you can characterise your load and design for it. UPS sizing, redundancy design, and utility interconnection agreements all depend on that premise. Agentic AI is, however, breaking that assumption in ways that the industry has not yet fully reckoned with.
Agentic AI workloads are fundamentally different from the training and inference tasks that data center power infrastructure was designed to support. An AI agent executing a multi-step task orchestrates other models, accesses external data sources, and adapts its approach based on intermediate outputs. The result is a power demand profile that is neither the sustained high-density draw of training nor the predictable throughput pattern of inference. It is, instead, a highly variable, spike-heavy demand curve. It can shift from near-idle to full load and back within seconds, repeatedly, over durations determined by task complexity rather than by any parameter the operator controls.
The Power Architecture Nobody Designed For This
The core problem is not that agentic workloads consume more power than training or inference. In many cases they consume less, because individual agentic tasks are less computationally intensive than training at scale. The problem is, rather, the shape of the demand. Data center UPS systems are sized for the peak load they are expected to serve. Redundancy systems are designed to maintain that load through a defined interruption window. PDUs, switchgear, and cooling infrastructure are, consequently, all specified against expected maximum draw.
When the load profile shifts from predictable to chaotic, each of these systems is exposed to stress it was not designed for. A UPS adequate for a training cluster at sustained 90% capacity may be inadequate for an agentic workload that spikes to 100% unpredictably and repeatedly. The dynamics of power switching, battery management, and load transfer are fundamentally different under variable demand than under sustained load. Conventional data center power design, in other words, assumes a level of demand predictability that agentic AI workloads simply do not provide.
The Grid Stability Dimension
The stress does not stop at the facility boundary. As we have covered in our analysis of how AI infrastructure is permanently restructuring the US power sector, the load ramps associated with AI training workloads already stress regional grid stability in ways that conventional industrial loads do not. Agentic AI compounds this problem by making the ramps less predictable.
A utility managing a 300-megawatt campus running sustained training workloads can model the facility’s demand and plan generation dispatch accordingly. A utility managing the same campus running predominantly agentic workloads, however, faces a demand profile that can shift significantly within minutes. The factors driving that shift are invisible to the utility. The nature and volume of tasks being executed by AI agents are not predictable from any external signal. That unpredictability makes demand response, grid balancing, and generation dispatch more difficult. It also increases the risk of voltage and frequency events that affect not just the data center but the broader grid segment it connects to.
Battery storage has emerged as the primary engineering response to this challenge. A data center placing adequate battery storage between its grid connection and compute infrastructure can absorb demand spikes and smooth the load profile presented to the grid. The unpredictable agentic demand curve is, consequently, converted into a more manageable sustained draw. Battery storage at the scale required to buffer gigawatt-level agentic workloads, however, adds significant capital cost and operational complexity. Facilities not originally designed with this architecture in mind face a challenging retrofit.
Cooling Infrastructure Faces the Same Challenge
The power demand variability of agentic AI workloads creates a parallel challenge for cooling infrastructure that is, in some respects, even harder to solve. Cooling systems are designed to remove heat at rates that match the thermal output of the compute infrastructure they serve. When that output is predictable, cooling design is straightforward. When it is highly variable, however, cooling systems face a difficult choice. They can be oversized to handle peak thermal output at all times. That is expensive. Alternatively, they can be designed to respond dynamically to variable load. That, however, requires response characteristics that conventional systems do not provide.
Direct-to-chip liquid cooling, which we have covered extensively in our analysis of how it scales from the rack to the AI factory, has significantly better thermal response characteristics than air cooling but still operates within engineering constraints that assume a degree of load predictability. When agentic workloads create thermal spikes that exceed those envelopes, the cooling system has limited options. It must throttle compute performance, accept temporary thermal limit violations, or fail to manage the heat.
The Software-Hardware Interface Problem
The deeper challenge is that agentic AI power demand is determined by software behaviour rather than hardware configuration. An agentic task is non-deterministic. The power consumed depends on how many sub-tasks the agent spawns and how complex each proves to be. It also depends on how many external API calls are made and how the agent adapts its strategy based on intermediate results. None of these variables are visible to the infrastructure layer at the time the workload begins.
A fundamental disconnect emerges as a result. The software layer determines agentic AI behaviour. The infrastructure layer must respond to it. Neither has reliable visibility into the other’s constraints. Infrastructure operators can instrument their facilities to measure agentic demand patterns after the fact. The data collected, however, reflects past workload behaviour. It does not provide reliable predictions of future demand. The engineering discipline required to manage infrastructure for non-deterministic workloads is genuinely different from anything the data center industry has previously developed. Training and inference at scale, by contrast, are deterministic enough to design around reliably.
What Operators Are Beginning to Do About It
The industry’s response to the agentic power challenge is still in its early stages, but several approaches are emerging. The most immediate is the adoption of more conservative power density planning. Operators who previously planned facilities to 85% or 90% of rated capacity are, consequently, targeting 70% or 75%. The extra headroom absorbs the demand spikes that agentic workloads generate. That approach costs money, both in the larger facility footprint and in the underutilised infrastructure it implies. It is, however, more manageable than running into capacity limits during peak agentic demand periods.
Dynamic power capping at the hardware level is a second response that is gaining traction. Modern GPU architectures support power management features that allow operators to set instantaneous power limits on individual devices. That capability is, in effect, a software-controlled power ceiling. Agentic workloads are forced to operate within an envelope the infrastructure can support. The trade-off is, however, computational performance. A power-capped GPU executes agentic tasks more slowly than an uncapped one. For many agentic applications, task completion latency is measured in seconds or minutes rather than milliseconds. In those cases, the performance reduction is acceptable. For applications where response time matters, it is not.
The Infrastructure Design Response
The most significant response to agentic power demand is happening at the infrastructure design level. Operators building new facilities specifically for agentic workloads are incorporating power architecture decisions that were not required for training or inference campuses. These include larger UPS systems relative to IT load and more sophisticated battery management systems capable of handling repeated charge-discharge cycles at high frequency. Cooling infrastructure with faster thermal response characteristics is also required. Furthermore, grid interconnection agreements increasingly include demand response provisions allowing the facility to shed load rapidly in response to grid conditions.
These design choices add cost and complexity to AI infrastructure development. They are, however, the cost of building infrastructure that actually matches the workload it serves. Agentic AI is not simply a new type of workload that can be accommodated within existing infrastructure design. It is a fundamentally different demand profile that requires infrastructure designed around its specific characteristics rather than adapted from designs optimised for training or inference. As we have covered in our analysis of the infrastructure gap agentic AI is about to expose, the operators who understand this and are designing for it now are building facilities that will serve agentic workloads effectively. Those deploying agentic applications into infrastructure designed for conventional AI workloads are accumulating technical debt. That debt will eventually require significant capital investment to remediate. The alternative is operational compromises that limit the scale and responsiveness of the agentic applications they can support.
