The fastest way to get more power from the US grid is to promise to use less of it. This is fast becoming the defining trade of the AI buildout. Plugging a large data center into the grid takes years today and is the single biggest obstacle to the buildout. Yet, Duke's Nicholas Institute found the grid can absorb about 98GW of new load without building a single new power plant. Last month, the federal energy regulator, FERC, did something to try to shake that capacity loose, ordering the six organizations that run America's grid to publicly defend their wait times or change the rules that create them. Regulators do not issue ultimatums like that unless something is badly stuck.
Something is definitely stuck. The 98GW of headroom is accessible only if new loads can prove that they are flexible and agree to power down for 44 hours per year (or 0.5% of the time). So, we have nearly 100GW of energy sitting in plain sight, but we can’t unlock it today because becoming a flexible load on the grid that can power down, or “curtail,” on demand is very hard to pull off. We think the machinery that will enable data centers to do that is a massive venture opportunity we call the “energy flexibility stack.” This month's issue is about what goes into that stack.
A data center that can prove to the utility and the regulator that it will power down during peak events (e.g hot summer days) gets onto the grid years faster than one that can’t. That is the whole trade. Keeping the promise without dropping customer compute work means having backup power on site, shifting jobs between facilities, and above all, knowing when an energy crunch is coming. Predicting that well is a very hard math problem, and hard math problems that gate a trillion-dollar capex buildout tend to become great venture markets. The incentive to solve this one is only growing. The grid is getting more attractive as a power source, with ERCOT now charging a deposit for every megawatt requested to flush speculative projects out of a grid interconnection queue. As we argued in our GET piece, time-to-power is everything in this buildout, and we expect serious investments to be made in solving the flexibility problem.
We look at the energy flexibility stack through a supply chain optimization lens. That’s our background, so it’s what we know. During the pandemic, I was at a logistics technology company, watching global supply chains collapse under structural bottlenecks and fickle supply networks. Companies responded by building redundancy, durability, and ultimately more complexity into their supply chains. “Redundancy” was the term we all used in 2020, and it’s “flexibility” now. The lasting winners of that era were the companies that figured out the optimization of these complex supply chains (companies like Lyric in our own portfolio). We see the same pattern forming around energy for compute. Compute sits on top of the energy supply chain, the fastest-growing and most structurally constrained and complex supply chain in the world. The past year, from sold-out turbines to jammed interconnection queues to Hormuz closures, rhymes with what we saw in 2020 and 2021. The similarities are uncanny. The shortages will pass, but in the aftermath, complexity will grow. The value will go to the optimization engines that power flexibility.
Flexibility is how you buy speed
Flexibility is a win for every party at the table. For the data center, curtailing 44 hours a year costs at most 0.5% of its compute hours, and in practice far less, since backup power covers most of the gap. This is a better trade than waiting years. For the utility, a flexible load fits inside the grid it already owns, with no new capex outlay. For the regulator and the public, Duke's follow-up study found a 1 to 2% trim in data center peak demand cuts everyone's electricity rates 0.5 to 2.8% while protecting reliability. That is why flexibility is faster. Nothing new has to be built, and nobody at the approval table has a reason to say no. Data centers have noticed, and are increasingly trading curtailability for faster interconnection.
Regulators are now working to make that trade the new norm. FERC is asking directly whether flexible loads should get interconnection studies in 60 days instead of years. ERCOT went further last week, launching the Provisional Controllable Load Resource (PCLR). Before this, a data center could only connect as much power as the grid could guarantee to deliver at every hour of the year, and because that guarantee had to cover peak demand hours, it is what causes years of studies and construction. The PCLR splits the connection in two. The data center gets a guaranteed slice sized to what today's wires can handle, plus a much larger flexible slice it can use whenever the grid has room, in exchange for letting the grid operator dial that slice down during peak hours. If you are a large load buyer and can prove you can work within this system, you get connected sooner.
ERCOT also approved new market rules under which almost any commercial or industrial load, from a factory to a data center, can become both a buyer and a seller on the grid. This means every node on the grid can be a sophisticated trader now. That door used to be closed. A building bought electricity at whatever rate its provider set, and selling into the market was reserved for power plants and a handful of specialized large loads that could meet strict technical requirements. Now the meter works both ways for most parties. These new rules and PCLR together make it clear that flexibility gets you power faster, but as more nodes on the grid participate in open market trading, competition will get tighter. This means that operating anything on the grid, whether a data center, a battery, a power plant, will get more complex from here. We think that complexity is a great opportunity for intelligence.
Companies are already capitalizing on this opportunity. Emerald AI is an Nvidia-backed startup whose software slows and reshuffles computing jobs when the grid needs relief. In a live test in Phoenix, it cut a working GPU cluster's power use 25% for three hours during real grid peaks, without breaking any customer SLAs. Emerald, Nvidia, EPRI (the electric utility industry's research institute), Digital Realty (one of the world's largest data center developers), and PJM (the regional grid operator) are bringing Aurora online this year, a 96MW Virginia data center designed to be flexible the day it goes live. PJM's participation is major validation. It runs the grid for 65 million people and decides who gets connected to it. It does not attach its name to experiments. When the organization that controls the queue co-sponsors a flexible data center, it is signaling that flexibility is how it wants the next wave of load to show up.
Flexibility is hard to pull off
Most large data centers plan to get power from both the grid and the assets they build behind the meter. They will have a portfolio of grid power, on-site gas generation, batteries, potentially solar and nuclear, and sometimes sister facilities in other regions. Promising the grid flexibility means orchestrating all of it in real time. When the grid calls, something has to decide which jobs slow down, which site picks up the work, whether to discharge the batteries or spin the turbines, or whether the smarter move is to sell power back at that hour's price. Prices move every five minutes, and every asset in the portfolio can now trade. The energy supply chain feeding a single facility has the complexity of a global logistics network now.
Optimizing assets in this environment is definitely hard, and we believe the stack needed to pull it off boils down to three things:
- Forecasting. You need to see peak demand spikes and price swings coming before they arrive, both in power markets and in your own asset.
- Decision intelligence and execution. Something has to take those forecasts and autonomously decide when to curtail, where to draw power from, and when to sell it back to the grid. Something also has to execute on those decisions.
- Intelligent Hardware. The batteries, turbines, cooling systems, etc all have to be connected and fast enough to respond the moment the decision is made.
It should go without saying that this whole stack has to be AI-native. No human team can watch this many variables at five-minute resolution and act effectively.
What we are looking for
We think the energy flexibility stack is a theme that can support multiple venture scale categories. Large C&I loads are turning into intelligent assets that manage their own energy supply chain end to end. They need power market forecasts that feed their decisions, intelligence that make decisions, systems that execute on those decisions and hardware that responds.
Inside the data center, flexibility starts with compute and cooling, and that ground is already well covered. Emerald decides which AI jobs slow, pause, or move when the grid calls, without breaking customer performance guarantees. Phaidra applies machine learning to the cooling and power systems that set how fast a facility can respond.
The optimization layer is the brain, deciding what to run, store, buy, or sell across behind-the-meter and grid assets. It splits into two jobs that reward different skills. Companies like Lumora equip the traders, rebuilding grid simulation to run fast enough for real-time price forecasting. Others like Shatterdome do the trading, aggregating batteries and flexible demand into a virtual power plant it dispatches autonomously. Few teams will be great at both.
The orchestration layer carries those decisions into the physical world. We have met multiple seed teams pairing software with connected hardware so that the turbines, batteries, and switchgear on a behind-the-meter island actually move when the model says move. Important work, but downstream of the more valuable decisions made above at the optimization layer.
Every one of these companies stand to be a control layer application on someone else's physical assets, and our test for all of them is the same. Do they own the dispatch decision or merely advise it? Can they earn their way deeper into the asset's economics, and eventually earn the right to operate it? Those are the big questions we are focused on. We believe that energy assets will attract a staggering amount of capital, but most will earn infrastructure returns. The venture outcomes belong to the software that makes a complex, flexible portfolio more efficient.