The fastest-growing electricity load in America right now is AI data centers. The numbers have gotten genuinely large. A single large training campus draws as much power as a mid-sized city. The collective installed base of US data centercapacity is on track to roughly double in the next several years, with most of the growth concentrated in northern Virginia, Texas, and a handful of other hubs.

The grid was not built for this. I think it's worth being explicit about how mismatched the existing infrastructure is to the demand. The Federal Energy Regulatory Commission has roughly 2,000 gigawatts of generation projects in its interconnection queue, more than the entire current installed base of the country, and the average wait time is roughly five years. Transmission line projects take 10 to 15 years to permit. New baseload generation is even harder to bring online quickly. The grid is, in practical terms, full for the next several years in the regions where data centers want to build, and the datacenter operators know this.

What's actually happening, as a result, is that hyperscalers are increasingly building their own power infrastructure on site. Microsoft, Google, Amazon, and Meta are all pursuing on-site generation strategies. The mix varies — some combination of solar, batteries, natural gas turbines, small modular nuclear(eventually), and fuel cells — but the common thread is that the grid is too slow to be relied on as the primary power source, and the data center operators are willing to spend a lot of money to control their own supply.

The willingness-to-pay piece of this is more important than it looks. For mostof the last several decades, the economics of on-site power generation and storage have been gated by a comparison to grid electricity, which is cheap. That comparison has killed a lot of technologies that would otherwise work fine. A long-duration storage system that pencils at fifteen cents per kilowatt-hour gets ignored if grid power is four cents. A fuel cell that run sat sixty percent efficiency gets ignored if a CCGT runs at sixty-three. The technical case has been there for a long time. The economic case never closed because the value of the energy was too low.

Data centers change this. The value of a watt delivered to a GPU running inference at peak hours is not really about the wholesale price of electricity. It's about the productivity of the compute it powers, which is something on the order of dollars per kilowatt-hour of AI workload, not cents. When the value of the energy is high enough, technologies that were previously priced out of the conversation become economically rational. The constraint shifts from cost-per-kilowatt-hour to availability-at-scale.

The parallel I find clarifying is to the early automobile industry. The first cars were battery electric. Electrification was already the obvious technology in 1900 — quieter, simpler, more efficient. Battery EVs have always been around 80–90% efficient at converting stored energy to motion. Internal combustion engines, even today, are around 30% efficient at best. By any efficiency measure, the early electric car was the right answer.

But internal combustion won, decisively, for almost a century. Not because it was more efficient — it wasn't and isn't — but because it scaled. You could ship gasoline anywhere. You could refuel a tank in a minute. The energy density was high enough to take a car across a continent. Battery EVs had none of these properties at the time, and the value of mobility — the ability to drive far, fast, on demand — was high enough that customers were willing to absorb the efficiency penalty to get the scale and the convenience. Ford understood this. The Model T won because the supporting infrastructure for ICE could scale faster than the infrastructure for batteries.

The same logic applies to data center power, but in the opposite direction. The value of the energy is now high enough that efficiency stops being the gating constraint. What matters is whether you can deliver gigawatts of capacity, onsite, in eighteen months, at a scale the grid can't match. The technology that wins doesn't have to be the most efficient. It has to be the most scalable for the value the energy generates.

That reframe is what opens up a category of storage and generation technologies that would have looked too expensive a decade ago. Long-duration storage at hundreds of megawatts. On-site hydrogen production paired with fuel cells. Flow batteries sized to a campus. All of these were technically possible for years. They never penciled because the value of the energy didn't justify the capex. Now it does.

The interesting question, from an electrochemistry perspective, isn't really about backup generators. It's about long-duration storage. A data center campus draws hundreds of megawatts continuously. The standard battery answer to long-duration storage is lithium-ion, and lithium-ion is great for two to four hours of peak shaving, but the economics fall apart somewhere between four and eight hours of discharge. Beyond that, you're stacking cells you don't need most of the time, and the capex stops penciling.

The pair that does work at that scale is an electrolyzer plus a fuel cell. You use cheap power to make hydrogen during off-peak hours, you store the hydrogen ,and you run it back through a fuel cell when the load needs it. The round-trip efficiency is worse than a battery, but it can scale. Even days or weeks of storage is physically practical. The math doesn't work at residential scale. It starts working at the scale of a datacenter, or a natural gas peaker plant.

Flow batteries are the other version of the same answer. Different chemistry,same idea: store the energy in a form that scales independently of the power conversion hardware. A flow battery's energy capacity is just how big the tanks are. The power conversion is set by the stack. That decoupling is what makes long-duration storage economically tractable.

All three of these — fuel cells, electrolyzers, flow batteries — are membrane devices. They use the same kind of polymer film in different geometries, with different chemistry on either side. A better membrane improves all three simultaneously. The electrolyzer gets cheaper. The fuel cell gets more durable. The flow battery loses less to crossover. The economics of long-duration storage, for a data center load, shifts.

This is most of why we pay attention to the data center buildout. The hyperscalers don't think of themselves as electrochemistry customers, but they're going to be. AI compute demand is pulling on a power infrastructure that can only be built fast enough with on-site generation. On-site generation at gigawatt scale requires storage. Storage at that scale, beyond a few hours is going to be electrochemical.

The mechanics of how this plays out are covered in more depth in the upcoming power piece. The short version: the demand pull from datacenters is one of the largest near-term catalysts for the membrane market, and the customers have the deepest pockets in industry. It's the kind of pull that's easy to underestimate right now because it's coming from a sector that nobody associates with electrochemistry, which is most of why we pay attention to it.

‍