The “once-in-a-hundred-years” storm now shows up about every three months.
The grid (almost all of us quietly depend on) was designed for a world where the big shock arrived once a generation. Now it arrives quarterly. And our electricity bill has gone from “not meaningful“ to something late-night hosts make jokes about - not in a good way.
Here’s the uncomfortable part: the thing straining the grid and the thing that has to rescue it are the same thing. AI.
For three to four decades, U.S. electricity demand was effectively flat. Every new appliance, screen, and server got absorbed by efficiency gains. That era is over, and it’s being replaced by three forces hitting the system at once:
Exponential, unpredictable load — data centers stacked on top of electrification, EVs, and heat pumps, growing faster and far less predictably than anything the grid was planned around;
A system built for gradual change — ISOs and TSOs designed for slow, centrally dispatched plants, now absorbing volatility on both the supply and demand side; and
Weather that no longer behaves — Winter Storm Uri blacking out Texas, 34 states near record highs last winter, blackouts on the Iberian Peninsula, wars rewiring European supply.
Sean Kelly has built his company around this change. He is co-founder and CEO of Amperon, the Houston-based forecasting company operating at the intersection of AI and energy data. Before Amperon, he spent over a decade as a power-markets trader - running plants and trading electricity across Texas and the eastern U.S. at firms like Tenaska, EDF, and E.On. In 2018, he and co-founder Abe Stanway started Amperon on a single thesis: forecasting wasn’t keeping up with how fast the grid was changing. Eight years later, that thesis is the entire industry’s headline.
And this isn’t a company riding the moment, it’s one the market has been validating for years.
Amperon was just named CleanTech Analytics Company of the Year at the 2026 CleanTech Breakthrough Awards 🏆, the latest in a track record that reads less like hype and more like consistent, independent confirmation that the thesis is working:
TIME - America’s Top GreenTech Companies (2024)
Forbes - America’s Best Startup Employers (2025)
a16z - American Dynamism 50, AI Edition
Energize Capital - Top 30 Software Innovators (2024)
DISTRIBUTECH (DTECH) - Initiate Winner, Emerging Energy Technology (2022)
GRIT Awards - Best Energy Workplaces (presented by Ally)
BigDeal Energy Load Forecasting Challenge - Winner, 2018 and 2022
In this conversation, part of our “AI Meets the Physical World” series, Sean walks through why the grid is structurally different now, why AI has to become the stabilizer rather than just the disruptor, and why energy ( unlike software) will never be won with a single product you build once and sell everywhere.
His view is grounded in 20 years of actually sitting in the trader’s seat, and in running forecasts today across 48 U.S. states, 16 European countries, Canada, and Australia.
Why the “once-in-100-years” storm now happens every quarter
The 50–100 hours a year that actually decide grid reliability
How forecasting went from a 15-day trading tool to grid-critical infrastructure
The ensemble engine: 6–8 models under every forecast, reweighted and retrained every hour
Why flexibility is not a feature, it’s the only alternative to blackouts
Data centers as the new demand response, and what the early-2000s playbook teaches us
Why there will never be one global grid and why that’s a moat, not a problem
Who captures the value: hyperscalers, utilities, incumbents, or startups
Why the company causing AI’s demand problem is also the one solving it
The old grid had a comfortable rhythm. Demand crept up slowly, efficiency clawed most of it back, and a few large, dispatchable plants were enough to keep the lights on. The institutions that run the grid ( the ISOs in the U.S., the TSOs in Europe) were built for exactly that pace of change.
Then the pace changed.
“The biggest challenge isn’t just how much load is coming. It’s how unpredictable and fast-moving it is. The grid was built for gradual change, and the AI load growth has been exponential.”
Sean’s framing of the climate side is the one that lands hardest:
“We used to have a one-in-100-year weather event. Now we’re having a one-in-100-year weather event about every three months.”
A hurricane reaching California. Uri freezing the Texas grid. A winter that pushed 34 states to or near record highs. Each of those events is a stress test the grid was never engineered to pass and they’re arriving faster than the system can adapt. Sean’s conclusion is blunt: over the next five years, AI has to win as the stabilizer, because the alternative is worse reliability and worse prices for everyone.
Forecasting used to be simple, and it used to be small. It started with the 15-day weather model (the European ECMWF model that went out 360 hours ), and almost everyone really only cared about the day ahead. It was a trader’s tool, not an infrastructure layer.
That’s no longer the answer. The European model now reaches 46 days, and even a 13-month horizon because the volatility that matters is now happening weeks out, not hours out.
“If you find out it’s not going to be windy next week, that means next week just got a heck of a lot more expensive. You need to understand that.”
This is where Amperon’s design philosophy gets interesting. The models retrain every hour, on hourly granularity in the U.S. and Canada and 15-minute granularity across all of Europe. And critically, no single model is trusted to be right.
“Any forecast you see from Amperon has about six to eight different forecasts underneath it that are automatically weighted and retrained each hour.”
That ensemble draws on roughly 40,000 weather points refreshed hourly, pulled from seven or eight different weather vendors -including a German vendor that’s simply better at German weather than anyone else, plus emerging AI weather models from the Europeans and Google now being onboarded.
“We look at about 40,000 weather points on an hourly basis. You can’t do that in Excel that macro would break really quick.”
It’s not the cheapest way to run a company. It is, Sean argues, the right way when your product is the future.
Ask where the new capacity comes from, and Sean doesn’t pick a favorite technology. He picks all of them because the math doesn’t work otherwise.
The demand projections from grid operators dwarf what we can realistically build in gas plants, wind farms, utility-scale solar, or batteries constrained by critical-mineral supply chains. So the flexibility stack (EVs and EV fleets, behind-the-meter storage, virtual power plants, and data centers that can actually power down ) stops being a nice-to-have.
“This isn’t an ‘if flexibility works.’ If flexibility doesn’t work, we are blacked out. That’s the answer. Prices keep rising and the power goes out.”
The unlock is precision. Because the crisis lives in just 50–100 hours a year, the choice is stark: get flexibility right in exactly those hours, or massively overbuild capacity you’ll pay for, and regret, for the next decade. As for what’s actually holding the grid back, Sean’s answer is unglamorous and correct: bureaucracy. Interconnection queues clogged with duplicate “phantom” projects, the years it takes to build a power line, a 100-year-old system that shouldn’t be run like a startup but should at least move faster than it does.
The most useful historical pattern in the conversation: Sean thinks data centers in 2026 look a lot like demand response in the early 2000s.
Back then, the idea of paying large commercial and industrial loads to “be on good behavior,” to power down when the grid asked, sounded impossible. How could a bottling plant possibly not bottle its cans on schedule? It was pioneered anyway, heavily by EnerNOC, whose co-founder Tim Healy now sits on Amperon’s board.
“I think you’re going to see this happen with data centers. People need compute but compute keeps getting more efficient, and the ones who get through the interconnection queue are going to be the ones who promise to be on good behavior.”
He points to companies like Emerald AI (where he’s an advisor) moving compute around in partnership with Nvidia. The bet is that “dispatchable” data centers, load that flexes with the grid rather than fighting it, become a core part of how this demand gets absorbed without blacking everyone out.
Here’s where energy diverges hard from software, and where Sean’s thesis has the most strategic bite.
“We’re never going to see one massive grid.”
Power markets are structured nationally and within Europe, practically regionally. Germany is not Spain is not the Netherlands is not the UK. New York is not Texas. When Amperon built out Europe, it couldn’t simply plug into one continental data source; it had to onboard a specialized German weather vendor, account for the Netherlands’ enormous behind-the-meter solar, and learn the nuances of 16 separate markets.
“AI is borderless. But on the other side of that, you still have to have the energy expertise.”
This is the part that founders chasing a “build once, sell everywhere” SaaS dream tend to underestimate. The subscription model is the easy part. The moat is the years of accumulated, market-by-market understanding underneath it- the thing an off-the-shelf LLM can’t simply reverse-engineer.
So in a market contested by hyperscalers, utilities, century-old incumbents, and startups — who wins?
Sean’s answer is a specific combination:
“The winner will combine AI-native infrastructure, proprietary operational data, and deep power-market expertise.”
His read on each player is clarifying. Hyperscalers have the compute but not the energy expertise. Utilities have the data but won’t hand it to big tech. Incumbents (many built in the 1990s) have the market relationships but struggle to hire the technical talent. Amperon’s wedge, he argues, is being cloud-native and AI-powered since its first model went live in November 2018, now forecasting across 51 million meters with under 5% customer churn.
The closing strategic note for allocators: in energy, the durable advantage isn’t the model architecture everyone will eventually copy. It’s the proprietary operational data and the market expertise that compound, market by market, year after year.
The frame worth holding onto is the one Sean ends on-the loop at the center of the whole story.
“The thing that got us into this situation of worrying about demand is also going to be the thing that saves us.”
AI is what made everyone suddenly care about electricity. AI is also the only realistic way to manage the load that AI itself is creating. That’s not a contradiction; it’s the operating reality of the next decade. The same intelligence layer that’s straining the grid is the one that has to make it smart enough and dynamic enough to absorb the strain.
And Sean is clear that this isn’t optional:
“It’s not ‘if’ we fix this. If we don’t fix this, we literally black out and rates go through the roof. So getting people educated on it and coming up with a solution together is how we keep the lights on.”
For investors, the takeaway is less about any single company and more about the shape of the opportunity: forecasting and grid intelligence have quietly graduated from a trading utility into foundational infrastructure for an electrified, AI-heavy economy. The market is national, messy, and resistant to one-size-fits-all disruption - which is precisely why the operators who understand it deeply, and got there early, are positioned to compound.
The grid that powered the last 30 years can’t power the next 10. Figuring out what does is one of the largest, least glamorous, and most investable problems of the decade.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.