In a recent engagement, I was asked my opinion about redox flow batteries as backup power for data centers. I wasn’t bullish, but that’s partly timing.
I’m reasonably well qualified to answer the question. For a couple of years, I was strategic advisor and Board observer with Agora Energy Technologies, a CO2-based redox flow battery startup, currently around Technology Readiness Level 4 (TRL4). I’ve assessed global grid storage requirements and technology characteristics, and published a scenario through 2060 for the market. I’ve spoken to storage developer CEOs and strategists on a couple of continents, and assessed deployments and opportunities on a couple more.
And in my previous career with a global tech giant, I ended up both visiting multiple data centers (Nortel, CNR, others), former data centers (the echoing vault underneath one of the major bank’s towers in downtown Toronto), and working on proposals and programs that involved data centers (Japan’s national citizen identification program springs to mind). And, obviously while I was more in the software end of the business more of the time, every project I was engaged with was running software in data centers somewhere.
As such, I have a decent appreciation of both sides of the equation.
Let’s talk data centers. You are reading these words because there’s a data center somewhere with a web server in it that’s providing them through the internet to your device. And the odds are that there are multiple data centers involved in the process because ads are being served, and those can be anywhere. And people in different parts of the world are quite likely being served content on their screens from different data centers due to the magic of mirroring. When we use the word Cloud it sounds amorphous, but it comes down to big, raised floor rooms with racks of computers humming away, massive amounts of air conditioning and a fire suppression system that will kill you because it replaces all the oxygen with inert gases.
And data centers are big power consumers. The numbers vary, study to study, but the most credible numbers are that data centers consume about 1%-2% of all electricity globally. Other studies go as high as 4%, but I found that as credible as the claims that emails had massive carbon footprints. But still, 1%-2% is massive, 196 to 400 terawatt-hours annually. For context, I recently worked up the annual additions of low-carbon electricity from wind, solar, hydro, and nuclear for China from 2010 to 2022, part of my ongoing assessment of the natural experiment in low-carbon generation scalability, and the total was 1,658 TWh, only 4-8 times as much as data center demand. Thankfully renewables are accelerating deployments globally, so there’s a good news story there.
Data centers are also major buyers of renewable electricity. They run on electricity, and they don’t need heating, so the vast majority of them don’t consume natural gas at all, except maybe for hot water for the staff where they haven’t just put in a heat pump. Oh, and buyers of Bloom Energy’s Bloom Boxes which are basically fuel cells that almost entirely run on natural gas, meaning that the Californian purchasers of them were actually seeing higher emissions from their ‘clean’ technology than from electricity from the grid. Bit of a head scratcher, that.
This comparison is a bit old, as I developed it in 2019 and big power purchase agreements (PPAs) and virtual PPAs have been signed since then, but it’s indicative. You’ll also note that total TWh consumed were lower than the global total, as the majors are huge, but there are an awful lot of massive data centers out there that aren’t one of the big Cloud providers, and of course time marches on and more data centers keep being built.
Big data centers are up in the 100 MW power requirement levels. That’s about 80,000 US households for a single featureless cube in an industrial park that glows like a grow op in infrared imaging. And there’s a hint about one of the major draws: HVAC. The range is 21% to 61% of demand being for air conditioning, with 38% being the average. Obviously there are a lot of heat pumps and a good case for ground and water sourcing for them involved, but that’s still an awful lot of heat.
But air conditioners don’t really care if they lose power for a few minutes. The racks and racks of computers, however? That’s a different story. They not only love perfectly managed MW whose frequency and voltage are rock solid, they don’t like losing power for a second. These aren’t laptops. The blade servers and GPUs in data centers don’t come packaged with lithium-ion batteries good for 10 hours.
No, all of the power backup and power management is external to them. And here’s where the case for batteries and data centers emerges.
The first thing to know is that every data center already has battery backup, mostly lithium-ion these days, although it wouldn’t surprise me to find some lead acid in an old and small data center somewhere. The reason is that requirement for no interruptions in power. Back when people used to put servers in closets in their offices, something that had already become archaic in 2000, they would mostly buy uninterruptible power supplies (UPS) to go with them. A UPS would provide a couple of minutes of power instantly in the event that there was a blip, allowing a somewhat graceful shutdown of servers.
Fine for non-mission critical stuff, and one of the components in data centers. Every few racks of servers has a UPS in modern data centers, cleansing frequency and voltage and ensuring that if the grid decides it’s uninterested in supplying electricity for any reason, the servers keep getting it. For maybe a minute or two.
What UPS’ provide is instantaneous power, but not longer duration storage. if the grid goes away, then another power source has to turn on to provide power for minutes or hours. Given that the USA’s grids have an average of two hours per customer per year barring major weather or fire events, it’s clear that data centers need more backup than they likely do in Germany or Denmark, where the average per customer is around 13 minutes. (As a note, both Denmark and Germany are approaching 50% of electricity demand annually from renewables, and both experience peaks of greater than 100% regularly, so grid reliability is empirically not impacted by high penetrations of renewables.)
And it’s what data centers use for longer duration backup that leads to some back flips and irrationality. Yes, mostly they use diesel generators. These are fast responding, automatically igniting generators, but they take 45 seconds to get to multiple MW of power output. The UPS’ only have to survive that long, so the batteries are sized accordingly. Engineers are involved, so they’ll provide more than that to guarantee 45 seconds and to deal with generator startup delays and failures, but they aren’t sized to run servers for more than minutes at a time. And there are the economics of mission critical loads and non-mission critical loads. Some banks of servers just stop working because service level agreements don’t demand that they be up. (I’m pretty sure CleanTechnica isn’t paying for mission-critical levels of hosting, for example.)
And that HVAC load? No batteries for it. If the power goes out, the temperature goes up a bit until the backup generators kick in, and that’s all designed into the mix. Brilliant people have spent their entire lives working to optimize this balance, and data centers are increasingly efficient at ensuring that the majority of electricity runs computing, not cooling, lighting, and coffee makers.
So the requirement is for reasonably quickly responding — tens of MW of power for potentially hours. Data centers will test that at least monthly to ensure it’s working, but may never need it in a given year. It will probably be used more in testing than it ever will in earnest. Diesel generators need to be run monthly for a few minutes just to keep them operational, and then there are major failover exercises that occur as well. A data center in Europe might not need the backup for five years. A data center in North America might not need it for a year or two.
But when they need it, it has to be there, it has to work, and it has to provide rather absurd amounts of power for hours.
Is the existence of diesel generators at data centers a climate or pollution problem? No, no it isn’t. Diesel generators that run a few hours a year are the least of our problems globally. Trucks that run hours every day are a much bigger problem, and natural gas and coal generators that run up to 90% of the time are the problem.
But backup for data centers that are being completely powered by renewables and that have dug deep shafts for ground source heat pump cooling still have diesel on site and it tends to stand out to them. And it tends to stand out not for rational reasons, but for irrational reasons. People point at them and pretend that they are meaningful, and the corporation which is trying really hard to be green asks, “Well, what are our options?” and appoints someone to go find out. Employee and consulting dollars are spent.
So what are the battery options? Well, lithium-ion battery packs certainly exist, and would certainly work. But if you want 10 hours of backup power, you need a lot more batteries at a lot more cost, because in cell-based battery technologies power and energy are tightly coupled. Power is like horsepower in a car engine and energy is the gas in the tank. Need more energy in a car? Add gas. Need more energy from cell-based batteries? Add more batteries with both power and energy. And cell-based batteries respond in milliseconds, which is far more than fast enough when paired with a UPS.
These attributes makes cell-based batteries useful for UPS’, aspects of grid storage, and good for behind-the-meter storage for a lot of applications, but makes them a poor fit for data centers. Far more cost than diesel generators, speed of response that’s unnecessary.
So what about redox flow batteries? In theory, they are clearly a better fit than cell-based batteries for data center backup, and I think they’ll dominate the field, eventually. But not today.
Why are they good? Well, they decouple power and energy. Want more energy? Make the tanks on either side of the fuel-cell like power conversion components bigger. Put more vanadium, zinc, or CO2 along with bromine on one side. Make the high-energy liquid tanks on the other side bigger. Want more power? Add more power conversion components in the middle. It’s easy to size power and energy to requirements.
And while redox flow batteries get going faster than diesel generators, they aren’t as quick as cell-based batteries. Pumps have to turn on, liquid has to flow through a bunch of relatively small tubes through cells that end up with ions going back through them membrane and producing electricity. That takes more time than getting juice from a cell where nothing moves and less time than turning on a diesel generator.
So in the future, redox flow batteries are nicely complementary as backup for data centers. But not today, in my opinion. Why? Because redox flow battery firms are almost entirely startups with very short track records of deployment. The mean time between failures for redox flow technologies is poorly documented in field conditions.
And a bunch of redox flow technologies require climate control of their own right now. Remember, data centers have to work in -45° Celsius to +45° Celsius weather, and some deal with greater extremes. Diesel generators don’t much care, although the diesel fuel itself will be different in cold climates than hot ones, but local providers will provide the right mix for local conditions without any prompting.
The electrochemistry and operational characteristics of redox flow technologies are advancing rapidly. Some of the current chemistries will fall by the wayside, most likely, due to low efficiencies and high cost of raw materials (iron air and vanadium respectively, put your hands up), as well as other failure conditions like unavoidable generation of unwanted hydrogen in the cells (iron air, put your other hand up). But machine learning is being used to identify combinations of chemistries and pH balances that have high probabilities of being better in multiple ways, and automated characterization is speeding up the testing by orders of magnitude. Advances in battery technology are leaping ahead much more quickly than they did over most of the development of lithium ion technologies.
Even then, while I strongly posit preserving biofuels for the hard to decarbonize areas of longer haul aviation and shipping, there’s no harm in putting a few hundred gallons of the stuff in tanks where it is consumed in vanishingly small amounts. Data center diesel consumption is a rounding error on a rounding error on a gnat’s thorax. Heck, even synthetic diesel, as absurdly expensive and higher-carbon as it will always be compared to biodiesel, is good enough here.
If you are only burning a few hundred gallons a year for a data center drawing 100 MW 24/7/365, the cost and carbon debt of fuel for your generators is not material. Fill your boots with the foolishness that is synthetic fuel if you run a data center and are being pushed to be virtuous and green. Or just get some good biodiesel and wait 5-10 years for redox flow batteries to be fit for purpose, and your generators to age out.
Source: Clean Technica