AI’s Power Problem Isn’t Just Supply — Why I’m Watching Eaton (ETN)

1 MW racks, 800 VDC, and why power delivery is becoming part of the AI computing system

Most discussions about AI data centers begin with the same concern: electricity demand.

That concern is real. More accelerators require more power, and a delayed grid connection can delay an entire data-center project.

But focusing only on the amount of electricity misses another problem.

An AI data center is not merely a facility that consumes a great deal of power. It can also behave like an enormous computing load whose power demand changes very quickly.

The challenge is no longer just producing enough electricity.

It is delivering that electricity efficiently, absorbing sudden changes in demand, and controlling the interaction between the data center and the grid.

More power and harder-to-manage power are different problems

Traditional data centers also experience changes in workload.

But their servers, storage systems, and applications often operate on different schedules. That diversity can make the total power profile relatively smooth.

A large AI cluster behaves differently.

When thousands of GPUs or other accelerators work on the same training or inference job, they can increase or reduce their activity together. Their power demand can move together as well.

These rapid changes are often described as AI power bursts.

The grid, transformers, switchgear, and distribution equipment must therefore handle more than the total amount of energy consumed over a day.

They must also tolerate how quickly the load changes and how those changes affect voltage, frequency, heat, and equipment stability.

That creates two separate questions:

Can the data center obtain enough electricity?

And:

Can it deliver and control that electricity without destabilizing the system around it?

Building more generation may help answer the first question.

It does not solve the second one by itself.

At 1 MW per rack, voltage becomes part of the architecture

Many current server systems distribute power inside the rack at relatively low voltages, including 54 V architectures.

That approach works at conventional server densities.

It becomes much more difficult when a single AI rack approaches hundreds of kilowatts or even 1 MW.

The reason is basic electrical physics.

Power equals voltage multiplied by current.

As a simplified illustration, delivering 1 MW would require approximately:

  • 18,500 amperes at 54 V

  • 1,250 amperes at 800 V

A real data center contains multiple conversion and distribution stages, so these figures are not literal design specifications.

They do, however, show why the industry is moving toward higher-voltage distribution.

Excessively high current requires more copper, larger busways, heavier connectors, and more physical space. It also increases resistive losses and heat.

At some point, the infrastructure used to deliver power begins competing for the same space that operators would rather use for computing equipment.

This is why 800 VDC is more than a change in voltage.

It represents a redesign of the power path so that electrical infrastructure can scale alongside the density of AI compute.

Power conversion is moving farther outside the rack

In a conventional data center, incoming AC power may pass through several layers of transformers, UPS systems, distribution equipment, and local power supplies.

Each conversion stage introduces some combination of energy loss, heat, equipment cost, and maintenance complexity.

Higher-voltage DC architectures move more of that conversion upstream.

Instead of repeatedly converting power close to every server, a facility can convert AC to high-voltage DC at a larger scale and distribute it closer to rows or racks.

That does not mean existing AC data centers will suddenly disappear.

Older facilities may first adopt hybrid designs, using dedicated power racks or row-level equipment to supply 800 VDC to high-density AI systems.

New AI factories may eventually be designed around high-voltage DC from the beginning.

The transition is therefore likely to happen at different speeds.

Existing facilities will retrofit what they can.

New facilities will have the opportunity to redesign the entire power architecture around much denser compute.

Batteries are becoming control systems, not just backup systems

For years, batteries and UPS systems were mainly understood as emergency equipment.

Their job was to keep servers running during a power interruption or provide enough time for generators to start.

AI data centers are expanding that role.

Battery energy storage systems and supercapacitors can help supply short bursts of additional power when a GPU cluster suddenly increases its demand.

They can also absorb rapid reductions in load instead of forcing the grid and distribution system to handle the full change immediately.

From the grid’s perspective, that can make an AI data center less like an unpredictable industrial load and more like a controllable energy asset.

This matters because obtaining enough electricity is only one part of securing a grid connection.

Utilities also need confidence that the facility can behave predictably without creating unacceptable stress on local infrastructure.

In regions where grid connections can take years, the ability to buffer and control load changes may have significant economic value.

Batteries, UPS systems, supercapacitors, power electronics, sensors, and control software are therefore moving beyond their old role as supporting equipment.

They are becoming part of the production system that allows AI compute to operate.

Why I’m watching Eaton

The company I am watching along this transition is Eaton, ticker ETN.

I am not watching Eaton simply because it sells transformers, switchgear, or electrical equipment.

The more important point is that Eaton is attempting to address the AI data-center power problem across a broad section of the system — from the grid connection to equipment located much closer to the chips.

Its AI infrastructure portfolio includes areas such as:

  • medium- and low-voltage power distribution

  • electrical protection and switchgear

  • conversion from AC to high-voltage DC

  • 800 VDC and 1,500 VDC busway systems

  • UPS and battery energy storage

  • supercapacitors for rapid load changes

  • power-quality monitoring and control

  • liquid cooling for high-density accelerators

  • prefabricated modular power and cooling systems

The significance is not merely that Eaton sells products in each category.

The company is attempting to combine them into a repeatable infrastructure design.

Its Beam Rubin DSX platform, developed around NVIDIA’s Vera Rubin DSX architecture, brings power distribution, high-voltage DC, energy storage, cooling, and control into a more integrated system.

The goal is to make AI factories easier to design, assemble, and deploy at scale.

If the value of an AI data center is no longer determined only by how many GPUs it contains, Eaton may occupy a wide portion of the infrastructure required to make those GPUs usable.

This demand is already beginning to appear in the numbers

This is not entirely a future product story.

In its second-quarter 2026 materials, Eaton reported that data-center orders in its electrical business increased by approximately 85% from the previous year, while related sales increased by roughly 65%.

Those figures should not be interpreted as revenue from 800 VDC alone.

They include demand for a much broader set of products, such as conventional power distribution, UPS systems, switchgear, data-center electrical equipment, and cooling infrastructure.

But that existing demand is still important.

It shows that Eaton is not merely preparing to enter the AI data-center market.

The company is already supplying the infrastructure used to build it.

If 800 VDC and integrated power-and-cooling designs gain adoption, Eaton would not have to create its entire market position from the beginning.

It could expand its role on top of an established data-center business.

Technology roadmaps and stock prices run on different clocks

The case for 800 VDC does not automatically become a case for an immediate rise in Eaton’s share price.

A new electrical architecture must pass through several stages before it becomes meaningful revenue.

Standards must mature.

Products must complete safety and reliability testing.

Customers must validate the design.

Data centers must incorporate it into their construction plans.

Equipment must then be manufactured, installed, commissioned, and operated.

Existing AC facilities may adopt only selected parts of the architecture.

Some new AI factories may use high-voltage DC from the beginning, while others may continue with hybrid systems for years.

Competition also matters.

Large cloud operators may develop more of their own infrastructure. Rival suppliers may win important projects. Grid delays or slower data-center investment could postpone deployment.

Eaton’s integrated platform may be technically compelling without producing the expected level of design wins or profitability.

The current valuation may also reflect a substantial amount of optimism about data-center growth.

That is why this is not an attempt to identify an immediate entry point.

It is a longer-term observation:

How far will AI data-center power architecture actually change, and how much of that transition can Eaton capture?

What I will watch from here

Several developments will determine whether this thesis strengthens or weakens.

  • Whether the first 800 VDC power-rack deployments appear in real customer facilities

  • Whether the architecture expands from individual racks toward row- and data-hall-level deployment

  • Whether Eaton’s integrated platform produces identifiable customer projects and design wins

  • Whether rapid growth in data-center orders converts into durable revenue and profitability

  • Whether integrated power and cooling systems materially reduce construction and deployment time

  • Whether energy storage and load control help customers obtain grid connections or operate more predictably

The thesis would weaken if 800 VDC adoption remains delayed, customers rely heavily on internally developed systems, competitors secure stronger positions, or rising orders fail to translate into attractive margins.

The next power question for AI may not be limited to how much more electricity the world can generate.

It may be about how that electricity reaches a massive GPU cluster.

How sudden changes in demand are absorbed.

How the facility avoids destabilizing the grid.

And how power, cooling, and energy-storage systems can be assembled quickly enough for AI compute to begin operating.

As AI data centers become larger and denser, electrical infrastructure stops being a utility in the background.

It becomes part of the computing system itself.

The company I am watching along that transition is Eaton (ETN).

I do not know when — or to what extent — the market will recognize this shift.

That is why I want to record the thesis before the technology transition becomes fully visible in reported results.



This reference price is not an entry price or a target price. It records the market price at the point when this technology and industry thesis was first published.

This article is an independent long-term observation based on publicly available technical roadmaps, industry materials, company announcements, and financial disclosures. Detailed source-discovery paths and research methodology are not disclosed.

This article does not constitute a recommendation to buy or sell any security.

#AIInfrastructure #DataCenterPower #800VDC #Eaton #ETN #EnergyStorage

Popular posts from this blog

AMD Is Trying to Turn Kernel Diversity Into One Software Asset

The first broad integration was rolled back, but the direction of the software stack is difficult to ignore The GPU software race is often described as a contest over who can write the fastest kernel. That is only part of the problem. A company may have several teams capable of producing high-performance matrix-multiplication kernels. Some may work through template libraries. Others may use compiler-generated code, handwritten assembly, or model-specific optimization stacks. The existence of those kernels does not automatically create one coherent platform. They still need to be: discovered validated matched to the correct problem given the correct arguments launched through a stable runtime and delivered through a reliable software package Recent public code changes inside AMD’s software ecosystem showed an attempt to connect these layers more directly. The central idea was to allow hipBLASLt’s GEMM selection system to choose kernels created by multiple internal kernel technologi...
Read more

PyTorch #188031 — Why Did Forward Use 64-Bit Indexing While Backward Still Used uint32_t?

PyTorch Spatial Softmax, Part 1 of 3 — An asymmetric indexing contract and the patch that aligned it A recent PyTorch change fixed a 64-bit indexing problem in the spatial softmax backward path. The bug was not in the softmax formula itself. It was in the width of the integer used to calculate memory locations. For the same large tensor, the forward path already selected either 32-bit or 64-bit indexing based on the input size. One backward path, however, continued to calculate sizes, strides, and offsets using uint32_t . That created a possible split in behavior: Forward → 64-bit address calculation → accesses the intended elements Backward → 32-bit address calculation → offset wraparound → gradients read from or written to different locations A tensor could therefore complete the forward pass correctly while the corresponding backward kernel could no longer represent every address in that tensor. The changes described in PyTorch #188031 were later landed in PyTorch’s main branch. S...
Read more

NVIDIA CUTLASS #3017 — Why load() and store() Mapped the Same Tile Coordinate to Different Addresses

CUTLASS #3017, Part 1 of 3 — An asymmetry in address calculation and the patch that fixed it In February 2026, while reading through NVIDIA CUTLASS’s RegularTileIterator<layout::PitchLinear> implementation, I noticed an odd asymmetry. load() and store() accepted the same tile coordinate, yet they calculated the corresponding memory address differently. The difference was a single line. load() divided the contiguous coordinate by kElementsPerAccess . store() did not. When each memory access handled only one element, the difference remained invisible. But in vectorized paths, where a single access handles multiple elements, the same logical tile coordinate could resolve to two different physical addresses. I reported the issue as CUTLASS Issue #3017. The missing division was later added to store() in PR #3049. CUTLASS and tile iterators CUTLASS is NVIDIA’s open-source library for high-performance CUDA linear algebra. It provides C++ template abstractions for composing GPU ope...
Read more