Data centers will be powered by solar and gas.
Plus, a new calculator for data center energy use
I’ve been working on a new data center energy calculator. It was published a few days ago, and you can use it here.
I’ll explain (1) what the calculator can do, and (2) what the numbers show.
#1: About the calculator
The calculator is an extension of the model developed by the offgridai.us team in their recent whitepaper: How off-grid solar microgrids can power the AI race.
The calculator pulls weather data from any location, simulates solar & battery power flow for 20 years, then runs a financial model to determine the Levelised Cost of Electricity (LCOE).
The code is open-source, and you can run simulation sweeps for different locations and generation mixes.
#2: The takeaway: data centers will use a lot of sun and gas.
AI needs power today. Developers cannot wait 5+ years for new grid connections, so off-grid data centers will supply a lot of future compute.
AI data centers want to juice their GPUs 24 hours a day. Power supply needs to be continuous at 100%.
Running a 100% renewable data center from solar and batteries is prohibitively expensive. You need too many batteries to guarantee reliability.
Adding full gas backup (capable of supplying 100% of demand) makes this workable.
In other words, you start by building a 100% gas powered data center, then add solar & batteries to reduce costs. The resulting cost vs renewables graph looks like this (for west Texas):
A 100% gas-powered data center is more expensive than a 40% to 90% renewable one. Good news!
Data center developers have been slow to internalise this fact; grid connections are appealing for reliability, and the off-grid paradigm is new. But grid connection queues have never been longer, and demand for quick-compute has never been higher.
The speed and economics are now being clearly born out by future projects. Stargate — the $100 billion megaproject by OpenAI, Oracle and SoftBank — is being built in Abilene, Texas, and will sport vast arrays of solar, as well as ten open-cycle gas turbines. The fact that the turbines are not combined-cycle gives a good indication that they expect to run at low load-factor. Meanwhile, Google is bankrolling 900MW (!!) of solar in Milam County, Texas, for its nearby data centers.
Take a look.
Enjoy the calculator, and let me know what you think. Thanks to Duncan and the rest of the offgridai.us team for developing the underlying model for this project, and helping me get this calculator live.
If you haven’t read the original whitepaper, you should - it has awesome analysis. For example, after filtering private US land for solar potential, distance from natural gas pipelines, and ease of construction, almost all the best land for datacenters is in West Texas.
Thanks for reading!
Ben
Ben, this tool is excellent - thank you for putting it together.
One question on the outputs:
When I put in 200 MW of Solar PV, the peak daily output is only ~150 MW, even in the best parts of the county like the example for Phoenix below.
How are you modeling solar output - are you assuming fixed tilt or 1-axis tracking panels? Something seems off because the the solar farm should be able to hit its nameplate capacity of 200 MW during peak solar hours in the middle of the day.
Cheers,
Kevin