Grid Law: Why AI Data Centers Became a Battery Problem

This one is for the people who build cells, packs, BMS and thermal systems. AI data centers have quietly created a battery requirement that didn’t exist five years ago, and most of the specification is being written right now by people who aren’t battery people. If that’s your world too, the Electrification Academy community is where these questions get worked through.

Right now, inside data centers across America, machines are pulling megawatts of power in a single millisecond – on, then off, millions of times a day.

We are living through the largest infrastructure buildout in human history, and almost nobody can actually picture what it looks like.

Here is the number that stopped me cold: within just a few years, close to 10% of all the electricity generated in the United States could go into one single purpose — powering AI. Not homes. Not factories. Artificial intelligence.

And here is what most people miss. The thing holding AI back isn’t chips anymore. It’s power. It’s the grid. And increasingly, it’s batteries.

To understand what that actually means at the hardware level, I sat down with Paul Charles — co-founder and former CEO of American Battery Factory, and one of the few people actively involved in the design and construction of AI data centers today.

AI Has a Power Problem – Can Batteries Solve It?

“Grid Law”: The New Bottleneck

Most people still see AI as software. Paul sees it as energy and infrastructure — and that reframing changes everything.

He has a phrase for it: grid law.

The pace of AI is no longer determined by CPUs or GPUs. It’s determined by how quickly we can deliver electricity, transmission, and cooling.

The numbers behind it: roughly 10,000 to 11,000 data centers globally, about half of them in the U.S.

It’s worth being precise about that headline number, because it gets quoted loosely. Data centers took about 3% of total US power demand in 2024. The 9% by 2030 figure — roughly 415 TWh a year, per the IEA — is the higher-growth scenario, built on 15% annual growth. The more moderate scenarios land between 4.6% and 6.8%. So the honest framing isn’t “AI will consume 10% of American electricity.” It’s that the plausible range now has a ceiling nobody was modelling five years ago, and the most recent reports are pulling even that timeline in by a year or two.

Meanwhile, the queue to connect to the grid is running 3 to 7 years. Call it five.

That single mismatch — exponential demand against a five-year interconnection wait — is the reason batteries stopped being an afterthought.

Where Do the Electrons Come From?

Paul’s answer was refreshingly unromantic: anywhere you can get an electron is being considered.

• Natural gas — the number one source right now
• Solar — costs have dropped so far that much more is coming online
• Nuclear, wind, geothermal
• Anaerobic digestion — converting biomass and carbon stock into renewable natural gas, then burning that for electricity

And once you’ve created those electrons, the stewardship piece is battery storage. Paul sees it splitting into two distinct segments:

1. Traditional grid stability and enhancement
2. AI data centers specifically — including microgrid capability

The question developers are asking has quietly changed. It is no longer “where is the cheapest electricity?” It is “where can I get reliable electricity?”

From One Big UPS to Batteries Inside the Rack

This is the part I found most interesting, because it’s a genuine architecture shift, not just a scaling-up.

Look at what happens to the battery’s position between those two diagrams. In the top one it sits off to the side, a bridge waiting for the generator to start. In the bottom one it sits in the main power path — the site BESS is between the rectifier and the distribution bus, not parked next to it. That’s not a bigger battery. That’s a different job.

The traditional design: you connect to the grid, power is brought to site, and one centralized uninterruptible power supply (UPS) — historically a lead-acid battery — sits there and holds things up for a few minutes. Its entire job was to bridge the gap until the diesel generators outside kicked in.

With AI, the rules have totally changed.

The newer designs physically remove the centralized UPS and distribute that responsibility outward — and inward. Literally into the rack.

People call these battery backup units (BBUs), a term Paul doesn’t like, because it badly undersells what they do. And he’s right. Consider what’s sitting in one of those racks: hundreds of millions — in many cases billions — of dollars worth of GPUs. When an AI training run starts, that rack draws megawatts within milliseconds. Then it’s off again.

That isn’t backup. That’s power conditioning for one of the most violent load profiles in industrial electronics.

There’s also a second element Paul described: Power Sidecars — additional batteries placed at the end of a rack system.

What used to be one box is now four distinct jobs:

This is also where the “eight minutes” number gets clearer. Rack-level backup is measured in minutes — that’s ride-through, and it’s supposed to be short. Site-level storage is measured in hours. The problem Paul ran into wasn’t someone sizing the rack batteries wrong. It was a site-level system being specified as though it were still a bridge to a diesel generator.

415 V AC → 800 V DC

Alongside the battery shift, the electrical architecture itself is moving.

Facilities have traditionally run on roughly 415 V alternating current. The new AI factories are moving to 800 V DC architectures, and the reasoning is stacked:

• Reduced conversion loss from repeatedly going AC → DC
• Less copper — and copper prices have gone through the roof
• Better efficiency overall
• Power delivered much closer to where it’s actually consumed

For anyone who has worked on high-voltage EV architectures, this will feel extremely familiar. The industry is relearning the same lesson in a different building. ⚡

The Chemistry Question

So what’s actually in these systems?

LFP (lithium iron phosphate) dominates — less expensive than chemistries containing nickel and cobalt.

Sodium-ion is the growth story, and for a very practical reason. For the containers that sit outside the building, sodium designs can be built with no moving parts — which means no cooling system to install, power, and maintain. In a facility where every kilowatt of parasitic load matters, that’s a real argument.

Inside the rack, it’s more varied: some LFP, some more power-centric chemistries, and increasingly supercapacitors combined with lithium in hybrid approaches.

Paul’s own comparison tables are worth studying, because they show how differently the two positions score. At rack level, you are buying milliseconds of response and 1–5 minutes of duration, and the exotic options start to make sense — niobium-based lithium-ion at roughly ten times the power density of standard Li-ion, or nickel-zinc where non-flammability outranks cost:

At site level, the calculation changes completely — you’re buying hours, and long-duration technologies like iron-air and redox flow enter the frame alongside the diesel generators they’re increasingly meant to replace:

Note where LFP lands in both tables. It isn’t the highest-performing option in either one — it’s the option that is safe enough, cheap enough, and available enough in both. That is usually what winning looks like in this industry.

Paul expects significant advances here in utilization and power efficiency — and I’d agree. This is exactly the kind of application where the system requirement, not the material, defines the winner.

The $50 Billion Project With Eight Minutes of Runtime

This was the story that stayed with me.

Paul was working through the battery architecture on a project worth roughly $50 billion. The response he got: we’re perfectly fine here — we have eight minutes of runtime.

Eight minutes.

How do you spend that amount of money, with essentially all the resources in the world available to you, and end up with an architecture designed around assumptions from ten or twenty years ago?

Paul’s counter-position: the realistic requirement is a minimum of two hours, and more likely four to six hours of battery runtime.

That gap — between eight minutes and six hours — is not a rounding error. It’s an entire market that most of the battery industry hasn’t priced in yet.

And it’s worth doing the arithmetic on what that actually means in hardware. At two hours of duration and 0.5C discharge:

Paul’s conclusion from that slide is the line I keep coming back to:

Every AI data center factory requires a battery factory behind it.

For anyone in cell or pack manufacturing wondering whether this is a real demand signal or a headline: 400 containers is not a pilot project.

Offshore AI: Putting Data Centers on Retired Oil Rigs

The most unexpected part of our conversation was what Paul is working on now: offshore AI infrastructure.

The team recently announced 40 identified sites across the East Coast, West Coast, and the Gulf. The concept: place AI infrastructure in 20- and 40-foot shipping containers, GPUs and all, onto decommissioned offshore oil rigs.

Why it works:

• Those rigs already have power — from 25 MW up to 200 MW — and have been running for decades
• Many are being decommissioned right now
• Most sites already have fiber running directly to land, so latency stays low

It’s a genuinely elegant use of assets that are otherwise just deteriorating. 🌊

The Part We Don’t Talk About Enough: Communities

One of the groups Paul works with tracks over 100 different moratoria on data center projects — some permanent holds, some “let’s pause until we understand what’s going on.”

His view, which I share: those questions are legitimate, and citizens should be asking them.

If a data center is coming to your community, Paul’s three starting questions are:

1. Water. What cooling architecture is being used?
2. Power. Is it coming from the grid, generated on site, or a hybrid?
3. Impact. What does this actually do to your community — good and bad?

On that last point, he offered a number worth knowing. Data centers don’t employ many people directly — but Paul puts the multiplier at roughly 1 to 6: for every direct data center job, around six more are created in the area. Build a real technical hub and it extends further still, from electrical technicians all the way down to the sandwich shop.

His closing advice was simply: start a dialog. Keep asking questions. And encourage your elected officials to get educated on this too.

Why This Matters for the Battery Industry

For years, “battery storage” in this industry meant EVs first, grid second. This conversation convinced me there’s now a third category with its own requirements — and it is being defined right now, at speed, by people who mostly aren’t battery people.

Millisecond-scale load steps. In-rack thermal constraints. 800 V DC. Runtime expectations off by a factor of forty. Chemistry selection driven by whether you can avoid a cooling loop.

If you build cells, packs, BMS, or thermal systems, this is a specification sheet you should be reading — not a headline you should be scrolling past.

These are exactly the questions we’re working through inside Electrification Academy — a community of battery and electrification professionals answering each other’s questions across cell manufacturing, pack, safety, recycling and transport.

We are far from done learning here.

–Veronika

#EV Electrified Veronika

Expertise, Empathy, Enthusiasm 

Watch the full conversation with Paul Charles on the “On the Floor” series: https://youtu.be/rtSoC290ZuA