Transformers Are Now Harder to Get Than GPUs: Why Power Is the Real AI Data Center Bottleneck
For a while, the center of the generative AI race was GPU supply. But buying GPUs doesn't help if you don't have the power to run them — and the electrical infrastructure to deliver that power into the data center.
Lately, delays in power equipment — transformers, breakers, switchgear — are increasingly what push back an entire data center build. In the US, lead times for some transformers have passed 160 weeks, forcing utilities and data center operators to lock in equipment years in advance. The AI infrastructure bottleneck has expanded beyond chip supply into power infrastructure as a whole.
This article looks at why power and transformers matter so much for AI data centers, and what companies need to do to raise GPU utilization and operational efficiency within a limited power budget.
Why Power Matters So Much in AI Data Centers
Traditional data centers were built to run servers and networking equipment reliably. AI data centers are different — they pack high-performance GPUs densely, which demands far higher rack power density and cooling capacity.
As GPU server density and power draw increase, you have to satisfy all of the following at once:
How much power capacity a rack can actually be supplied
How much load the transformers and switchgear can handle
Power distribution losses inside the data center
Cooling capacity to remove the heat GPU servers generate
The maximum power allowed for the facility as a whole
Next-generation AI data centers are pushing power demand and heat load so high that they're exposing the limits of existing power supply structures. That means how you supply and manage power now matters as much as GPU performance itself.
If GPUs are the production equipment that runs AI compute, power infrastructure is the foundational system that keeps the production line running. Without power infrastructure in place, even the most expensive GPUs can't be converted into actual AI compute or service throughput.
Check Power Before You Buy GPUs
Building an AI data center starts with confirming that the surrounding power grid actually has enough capacity around the site. From there, you determine the required receiving capacity and voltage, and design and manufacture transformers and switchgear to match.
The problem is that bottlenecks can appear at any point along the way. If the nearby substation or transmission network has no spare capacity, you may need new lines or substation expansion. Delays in negotiating with the utility, permitting, design and construction schedules, or simply failing to secure transformers, breakers, and switchgear in time — any of these can push back when the entire data center actually goes live.
In fact, an analysis of major US data center construction using satellite data found that roughly 40% of projects scheduled to go live in 2026 may miss their planned timeline. Power and equipment shortages aren't the only cause — permitting, skilled labor shortages, and construction delays are all compounding factors.
Grid interconnection times are also stretching out. In some saturated data center hubs like Amsterdam and Tokyo, it can reportedly take up to 10 years for a new data center to connect to the grid (JLL, 2025). In the US, PJM-territory projects that reached commercial operation in 2025 spent an average of 7–8 years just in the interconnection queue. In grid-constrained regions, you can build the data center first and still have no fixed date for when power actually arrives.
That's why building an AI data center shouldn't start with picking a GPU model and quantity — it should start with confirming how much power you can actually secure at that site, and when. If power supply turns out to be impossible or slower than expected, GPUs bought early can sit unused for a long time, generating nothing but depreciation and maintenance cost.
Why Transformers Became the Core Bottleneck
Transformer lead times have changed dramatically: standard transformers went from roughly 50 weeks in 2021 to about 120 weeks in 2024, and large transformers now regularly exceed 160 weeks — about three years — with some specialized units taking up to four.
Among all the power equipment involved, transformers have emerged as the single biggest bottleneck recently.
A transformer converts the power delivered from the external grid into a voltage that the data center's internal equipment and GPU servers can actually use.
But a data center transformer isn't a commodity part you can order and receive right away. It has to be engineered to the specific voltage, capacity, footprint, cooling method, and redundancy requirements of that data center. Once manufactured, it still needs testing for electrical performance, insulation, temperature rise, and safety.
AI data center growth has rapidly increased demand for large-capacity transformers, but manufacturing capacity, skilled labor, raw materials like electrical steel and copper, and testing facilities can't scale up quickly. That's why standard power transformer lead times have already more than doubled — from about 50 weeks in 2021 to about 120 weeks in 2024 — and large transformers for substations and generators now regularly exceed 160 weeks (roughly three years), with some specialized units reportedly taking up to four years.
According to an analysis citing Wood Mackenzie data, demand for generator step-up transformers in the US grew 274% between 2019 and 2025, while demand for substation power transformers grew 116%. With data centers, expanding power generation, and aging grid replacement all competing for the same supply at once, the shortage has become structural.
A delayed transformer doesn't just delay one piece of equipment — it can cascade into delays for the switchgear, UPS, wiring, cooling systems, and commissioning schedule that all connect to it.
Four Things Companies Need to Prepare for in the Power-Constrained Era
Since building new power supply infrastructure can take years, companies can't afford to just wait for new power to arrive. They need an operational strategy that gets more out of the power and GPUs they already have.
1. Raise the utilization of the GPUs you already have
Before buying new GPUs, check the utilization and idle time of your current resources. Automatically reclaiming and reallocating unused GPUs, and letting multiple workloads share a single GPU through resource partitioning, can meaningfully increase throughput without adding any power draw.
2. Let multiple teams share GPUs safely
Because GPUs and power are both limited resources, you need to prevent both long-term monopolization by a single team and uncontrolled sharing. Role-based access control, per-project resource allocation, environment isolation, and priority policies let multiple departments use one GPU infrastructure without conflict.
3. Tune GPU power to the workload
The optimal power setting for a GPU depends on the hardware and the workload, so you shouldn't apply one power limit across every job. Set different power caps for training, development, testing, and inference, and use benchmarking to find the right balance between performance and power consumption for each.
4. Measure GPU usage and cost
You need to track GPU occupancy time, actual utilization, and power consumption by team and by project to make reasonable resource allocation decisions. That data lets you analyze cost per project, identify idle or over-allocated resources, and run showback/chargeback.
AIPub: The GPU Operations Layer for the Power-Constrained Era
These four challenges aren't separate problems. Raising GPU utilization requires identifying and reallocating idle resources. Sharing resources across teams requires permissions and isolation policies. Tuning power to the workload requires real-time visibility into GPU state and performance. And reasonable resource allocation requires tracking usage and cost by team.
Managing GPU resources, user permissions, workloads, and power and cost through separate tools makes it hard to see the full operational picture. It becomes difficult to tell whether a given GPU is sitting idle because of resource allocation, permission policy, workload characteristics, or a power setting.
AIPub brings these GPU operations challenges together on a single platform.
GPU resource partitioning and reallocation
Split GPUs so multiple workloads can share them, identify idle resources, and reallocate them to the users and projects that need them.
Multi-tenant resource management
Isolate GPUs, storage, and development environments by user, team, and project, and set access permissions and resource limits by role.
GPU power and clock control
Set power limits per GPU card and adjust SM and memory clocks based on workload characteristics to find the right balance between performance and power efficiency.
Unified monitoring and cost management
Track GPU utilization, power, temperature, network, and storage status in real time, and build a showback/chargeback system based on usage by team and project.
AIPub is an enterprise GPU operations platform built not just to help you acquire more GPUs, but to help you run the GPUs and power you already have more efficiently.
Conclusion: Power Is Becoming AI Infrastructure's Real Competitive Edge
With transformer lead times exceeding three years and grid interconnection taking 7–10 years, power is no longer just a line item in data center operating costs. What determines your AI infrastructure's competitive edge isn't how many GPUs you've secured — it's when, where, and with how much power you can actually run them.
It will take time for new power and transformer supply chains to normalize. In the meantime, the fastest thing a company can do is operate the GPUs and power it already has more precisely. Before buying more GPUs, run the ones you have better.
Does any of this sound familiar?
Your GPU rollout timeline is uncertain because of power availability or transformer lead times at a new site
You've added GPUs, but you're hitting power limits and can't run them at full capacity
You can't decide whether to build a large data center or expand in stages
You can't accurately track GPU and power usage — or cost — by team
If any of this applies to you, let's design your AI infrastructure and GPU operations strategy for the power-constrained era together, with TEN's experts.
👉 Explore RA:X infrastructure consulting
Reference
Wood Mackenzie, Transformer & Electrical Equipment Lead Time Survey (2025~2026)
Data Center Knowledge, "AI Data Center Boom Rewires US Power Supply Chain" (2026)
JLL, Average Power Grid Connection Lead Time for Data Centers by Market (2025)
PJM Interconnection, Interconnection Queue & Time-to-Energize Data (2025)
U.S. Department of Energy, Large Power Transformer Resilience Report (2024)
Related reading