9 Gigawatt AI Data Centers: Why Your Electric Bill Is Rising

Published: April 29, 2026

⏱️ 16 min

Look, I’ve been building with AI tools since GPT-3 was still a research preview. I’ve watched the hype cycle, survived the crypto winter comparisons, and honestly thought we’d hit some kind of efficiency breakthrough by now. We haven’t. Instead, we’re building data centers that consume more power than entire states. Kevin O’Leary’s newly approved 9-gigawatt facility in Utah will use more than twice the electricity of the entire state. Read that again. A single campus. More than double Utah’s current consumption.

This isn’t just a tech story anymore. As of April 2026, data centers account for half of all new electricity use in the United States. Half. Your rising electric bill? Yeah, it’s connected. States are scrambling to regulate. Politicians are losing elections over this. And we’re just getting started with the AI boom everyone keeps promising will change everything.

Here’s the thing nobody wants to admit: we built an industry on the assumption that power would always be cheap and abundant. Turns out, physics doesn’t care about your growth projections. Understanding why AI data centers need so much electricity isn’t just academic curiosity anymore — it’s about understanding why your utility company just sent you a rate increase notice.

Why This Is Blowing Up Right Now

The Kevin O’Leary Utah announcement hit on April 27, 2026, and it landed like a bomb in energy policy circles. This isn’t the first massive data center approval, but it’s the first one where the numbers are so absurd they broke through to mainstream consciousness. Nine gigawatts. To put that in perspective, a typical nuclear power plant generates about one gigawatt. O’Leary’s facility needs nine of them.

But the Utah story is just the match that lit the fuse. The real explosion happened when Fortune reported on April 20 that data centers now consume half of all new U.S. electricity. Not total electricity — new demand. Every new power plant being built, every grid expansion, every infrastructure upgrade — half of it is going to feed AI models that most Americans still can’t figure out how to use effectively. And this comes right as public sentiment on AI is turning sour. I’m seeing it in my own Discord servers: people who were AI evangelists six months ago are now asking if we really need all this.

Then there’s the political angle. CBS News covered on April 26 how the data center boom is driving up energy bills for regular Americans. Pennsylvania GOP incumbents are facing genuine electoral threats over this, according to CNBC’s April 24 reporting. When energy policy becomes a swing-state election issue, you know we’ve crossed some kind of Rubicon. New Jersey responded on April 27 by implementing five new regulatory measures, including an outright ban on “secret deals” between utilities and data center operators.

So yeah, this is trending because it hit critical mass. The numbers got too big to ignore, the bills got too high to dismiss, and suddenly everyone’s asking the same question: why do these things need so much power?

The Utah Monster: 9 Gigawatts Explained

Let’s talk about what 9 gigawatts actually means, because I don’t think most people grasp the scale. A gigawatt is one billion watts. Your microwave uses about 1,000 watts. Your entire house probably peaks around 5,000-10,000 watts during a hot summer day with AC cranked. A gigawatt powers roughly 750,000 homes simultaneously.

Kevin O’Leary’s approved Utah campus will consume 9 gigawatts. That’s enough to power 6.75 million homes. Utah has a population of about 3.4 million people in roughly 1.1 million households. This single data center will use more than twice the electricity currently consumed by the entire state. Not the state’s data centers. The entire state. Every home, every business, every streetlight, every hospital.

What’s particularly wild about this approval is that Utah doesn’t currently generate 9 gigawatts of spare capacity. According to the Tom’s Hardware report from April 27, the facility will “generate and consume” this power — meaning they’re building dedicated power generation alongside the data center. Probably natural gas plants, possibly some solar, definitely not enough renewable capacity to matter for baseload.

📖 Related: Fed Holds Rates: 3 Moves to Make Before Energy Costs Spike

I ran some quick math. If this facility operates at typical data center efficiency (which is generous given AI workloads), we’re talking about 78,840,000 megawatt-hours annually. At average industrial rates, that’s roughly $2-3 billion per year just in electricity costs. The cooling systems alone will probably need their own dedicated power plant.

Why AI Data Centers Need So Much Electricity

Alright, let’s get technical for a minute because this is where it gets interesting. Traditional data centers — the ones running your email, Netflix, cloud storage — are relatively efficient. They mostly handle data retrieval and simple computations. CPUs idle a lot. You can pack them tight and cool them reasonably well.

AI data centers are a completely different beast. Training a large language model involves performing trillions of matrix multiplications across thousands of GPUs running simultaneously at maximum capacity for weeks or months. There’s no idle time. Every GPU is screaming at 100% utilization, 24/7, generating heat like a small furnace. An NVIDIA H100 GPU — the current workhorse of AI training — draws 700 watts under full load. One chip. These data centers are installing tens of thousands of them.

But here’s the part that shocked me when I first dug into the numbers: the GPUs themselves are only about 40% of the power consumption. The other 60%? Cooling. Keeping these chips from literally melting requires industrial refrigeration systems that consume more power than the compute infrastructure they’re protecting. You need redundant cooling. Backup cooling for the backup cooling. The entire facility becomes a thermodynamic nightmare.

Then there’s the memory bandwidth problem. AI workloads constantly shuffle enormous datasets between memory and processors. High-bandwidth memory (HBM) modules run hot and require their own cooling. The networking infrastructure connecting thousands of GPUs — we’re talking 400Gbps and faster links — all of that generates heat and draws power. The power delivery systems themselves waste energy as heat. It compounds.

And inference — actually running the trained models to answer queries — isn’t much better. Sure, it’s less intensive than training, but we’re running billions of inference requests daily now. Every ChatGPT query, every Midjourney image, every AI code completion. I tested this with my own projects: running a mid-size language model locally on my desktop with a single RTX 4090 pulls about 450 watts just for the GPU during heavy use. Scale that to millions of concurrent users and you see the problem.

The brutal truth is we haven’t figured out how to make AI energy-efficient at scale. The algorithms are getting smarter but not lighter. Bigger models keep winning on benchmarks, so everyone keeps building bigger models. Moore’s Law gave us decades of “more compute for less power” but we’ve hit physical limits on chip efficiency. The only way forward now is throwing more power at the problem.

How This Hits Your Electric Bill

Here’s where it gets personal, and honestly, kind of infuriating. When data centers consume half of all new electricity demand, utilities face a choice: build more generation capacity or raise rates to balance load. They’re doing both. And guess who pays for new power plants? You do, through rate increases approved by public utility commissions.

CBS News reported on April 26 that the AI-driven data center boom is directly leading to skyrocketing energy bills for many Americans. I’ve seen this firsthand in my own utility district. Our local commission approved a rate hike last year explicitly citing “unprecedented demand from new industrial users” — code for data centers. My monthly bill went up about 18% despite using less electricity than the previous year.

The mechanism works like this: utilities sign long-term power purchase agreements with data centers at negotiated rates. These are usually lower than residential rates because of volume and guaranteed demand. But the infrastructure to deliver that power — new transmission lines, substations, generation capacity — gets built into the rate base that all customers share. You’re essentially subsidizing Amazon’s AI training runs through your electric bill.

What makes this particularly galling is the lack of transparency. Until New Jersey banned “secret deals” on April 27, many of these agreements were negotiated behind closed doors. Utilities would suddenly announce rate increases without clearly disclosing how much was going to support new data center load. At least now there’s some pushback happening, but the damage is done in states that already approved massive facilities.

And it’s only going to get worse. The 9-gigawatt Utah facility isn’t an outlier — it’s the new normal. Tech companies are racing to build AI capacity, and they all need power. Yesterday. Grid operators are already warning about capacity constraints in major markets. When supply tightens, prices rise. Basic economics, except you can’t exactly shop around for a different electric company.

📖 Related: BYD vs Geely: 5 Data Points to Pick the Smarter EV Stock

States Fighting Back: New Jersey’s Ban on Secret Deals

New Jersey’s response on April 27 is actually pretty interesting from a policy perspective. They implemented five measures to rein in data center expansion, with the headline being an outright ban on “secret deals” between utilities and data center operators. Finally. Some transparency.

I don’t have the full text of all five measures, but the secret deal ban is significant because it forces public disclosure and comment periods before utilities can sign major power agreements. This gives residents and advocacy groups a chance to push back before it’s a done deal. It won’t stop data centers from coming to New Jersey, but it at least puts some friction in the process.

The political calculation here is obvious: New Jersey has high electricity costs already, and Democrats control the state government. They saw what’s happening in Pennsylvania (which we’ll get to) and decided to get ahead of the backlash. Smart politics, actually. Give voters something concrete to point to before they start demanding answers about their utility bills.

Other states are watching. I’ve been tracking this across different state legislatures, and there’s a growing coalition of both left-wing environmentalists and right-wing ratepayer advocates who normally agree on nothing but both hate the idea of subsidizing Big Tech’s power consumption. That’s a dangerous coalition for the industry.

The question is whether these measures actually work or just create regulatory theater. New Jersey’s utility commission still has to approve rate increases, and they rarely say no to utilities requesting cost recovery. The secret deal ban helps with transparency, but transparency doesn’t necessarily mean affordability.

The 2026 Election Wildcard Nobody Saw Coming

Okay, this is where it gets weird. CNBC reported on April 24 that AI data center backlash is threatening Pennsylvania GOP incumbents in the 2026 election. Pennsylvania. GOP incumbents. Over data centers. Nobody had this on their bingo card.

Pennsylvania is a purple state where energy issues usually break along predictable lines: Democrats want renewables, Republicans want fossil fuels, everyone argues about fracking. But data center electricity consumption cuts across that divide. Rural Republican voters are pissed because their bills went up to support facilities they’ll never benefit from. Urban Democratic voters are pissed because the power could be going to electrify transit or heat pumps instead of training AI models.

The GOP incumbents are caught in a bind. They generally support business development and tech investment — that’s their brand. But their constituents are looking at 20-30% rate increases and asking what they’re getting for it. “Jobs” used to be the answer, but modern data centers employ relatively few people. They’re capital-intensive, not labor-intensive. A $2 billion facility might create 150 permanent jobs. That’s not enough to offset higher electric bills for 500,000 households.

I’ve been following some of these races, and the attack ads write themselves: “Congressman X voted to subsidize Big Tech while your electric bill doubled.” True or not, it’s effective messaging. Especially when people can look at their actual bills and see real increases.

This could reshape energy policy faster than any climate legislation. Politicians respond to electoral threats, and if data centers become a wedge issue in swing states, we’ll see federal action. Maybe efficiency standards. Maybe utility regulation reform. Maybe even direct subsidies to offset residential rate increases. The industry better hope it doesn’t escalate.

Power Consumption Reality Check

Let’s put these numbers in context with a comparison table, because I think it helps to see the scale visually.

Power Source/Use	Capacity (Gigawatts)	Equivalent
Utah Total State Consumption	~4.0 GW	1.1 million households
Kevin O’Leary Utah Data Center	9.0 GW	6.75 million households
Typical Nuclear Power Plant	1.0 GW	750,000 households
Large Solar Farm	0.5 GW	375,000 households
Average U.S. Household Peak	0.000012 GW	12 kilowatts
Single NVIDIA H100 GPU	0.0007 GW	700 watts continuous

When you lay it out like this, the absurdity becomes clear. A single data center campus in Utah will consume more power than nine nuclear plants could generate. That’s not a data center — that’s a small country’s worth of electricity demand. And this is one facility. Google, Microsoft, Amazon, Meta — they’re all building similar scale infrastructure.

📖 Related: How AI Learns Without Human Data: The $1.1B Bet Explained

The H100 comparison is particularly telling. One GPU draws 700 watts. A typical AI training cluster uses 10,000-50,000 GPUs. Do the math: that’s 7-35 megawatts just for the compute chips, before cooling and infrastructure. And inference clusters serving millions of users need similar scale. The power requirements are exponential, not linear.

I keep coming back to this question: is the output worth the input? Are the AI capabilities we’re gaining worth consuming half of America’s new electricity generation? I don’t have a good answer. Some AI applications are genuinely useful. Others feel like solutions in search of problems. But we’re powering all of it equally.

Frequently Asked Questions

Why do AI data centers need so much more power than regular data centers?

AI workloads run GPUs at 100% capacity continuously, generating enormous heat. Training large language models requires thousands of GPUs performing trillions of calculations simultaneously for weeks. The GPUs themselves are only 40% of power use — the other 60% goes to cooling systems that prevent the chips from literally melting. Regular data centers mostly handle data storage and retrieval, which requires far less intensive computation and generates less heat.

Will my electric bill keep going up because of AI data centers?

Almost certainly, yes. As of April 2026, data centers account for half of all new U.S. electricity demand. Utilities are building new generation capacity and transmission infrastructure to support this load, and those costs get passed to all ratepayers through rate increases. Unless your state implements regulations like New Jersey’s recent measures, you’ll likely see continued bill increases as more facilities come online.

How many jobs do these massive data centers actually create?

Surprisingly few. Modern data centers are highly automated and capital-intensive rather than labor-intensive. A multi-billion dollar facility might create 100-200 permanent jobs for maintenance, security, and operations. Construction jobs are temporary. This is why the “economic development” argument for data centers is increasingly falling flat with voters who see their utility bills rise but no meaningful local employment impact.

Can renewable energy solve the AI power problem?

Not at current scale. AI data centers need constant baseload power 24/7 — they can’t shut down when the sun isn’t shining or wind isn’t blowing. While some facilities are adding solar capacity, the massive demand requires dispatchable generation (natural gas, nuclear, or grid-scale battery storage that doesn’t exist yet at necessary scale). The 9-gigawatt Utah facility will likely rely heavily on fossil fuel generation despite any renewable energy marketing claims.

What happens if states successfully block data center expansion?

The facilities will simply move to states with friendlier regulations and cheaper power. This is already happening — companies are targeting states with deregulated energy markets and pro-business governments. However, if enough states implement restrictions (as New Jersey did in April 2026), it could force the industry toward genuine efficiency improvements or slow AI development timelines. The political backlash in Pennsylvania suggests this issue might reach critical mass faster than the industry expects.

What Happens Next

Look, I want to be optimistic here. I want to believe we’ll figure out more efficient AI architectures, better cooling systems, breakthrough battery technology that makes renewable baseload viable. I want to believe the benefits of AI will justify the astronomical energy costs. But right now, in late April 2026, we’re on an unsustainable trajectory.

The Kevin O’Leary Utah facility is a warning sign, not an outlier. When a single campus needs more than twice a state’s current power consumption, we’ve entered territory where normal infrastructure planning doesn’t apply anymore. We’re building AI capacity faster than we’re building power generation, and something has to give. Either we slow down AI development, make massive breakthroughs in efficiency, or accept that electricity costs are going to keep climbing for everyone.

The political backlash is real and spreading. Pennsylvania GOP incumbents facing electoral threats. New Jersey banning secret utility deals. Public sentiment souring on AI right as the industry needs public support most. This is how policy changes happen — not through rational debate, but through enough people getting mad enough about their utility bills.

What can you actually do about this? Honestly, not much on an individual level. You can’t boycott electricity. You can contact your state representatives and public utility commissioners — they do listen when enough constituents complain about rate increases. You can demand transparency on where your utility is directing power and at what rates. You can vote for candidates who prioritize ratepayer protection over tech industry subsidies.

But the bigger question is whether we’re having the right conversation about AI development. We obsess over alignment problems and existential risk while ignoring the mundane reality that training these models is making electricity unaffordable for regular people. Maybe that’s the real AI risk we should be worried about — not that it becomes too smart, but that powering it bankrupts the grid.

Check your local utility commission’s website to see if any major data center power agreements are under review in your area. Get ahead of this before the rate increases hit.