AI Applications
Does NASA Use Supercomputers? Yes, and Here's What They Actually Do
NASA runs multiple supercomputers, mainly at Ames Research Center in California and Goddard Space Flight Center in Maryland. They simulate rocket launches, Mars parachutes, hurricanes and galaxy formation. Here's what each system does, why cloud computing can't fully replace them, and how NASA's compute is shifting towards AI foundation models.
CSEWhy ·
NASA runs multiple supercomputers, mainly at Ames Research Center in California and Goddard Space Flight Center in Maryland. They simulate rocket launches, Mars parachutes, hurricanes and galaxy formation. Here's what each system does, why cloud computing can't fully replace them, and how NASA's compute is shifting towards AI foundation models.
Yes, and not just one
NASA operates several supercomputers, not a single machine. The biggest cluster of them sits at the NASA Advanced Supercomputing facility at Ames Research Center in Moffett Field, California, which has housed systems named Pleiades, Electra, Aitken and Endeavour. A second major site is the NASA Center for Climate Simulation at Goddard Space Flight Center in Maryland, home to the Discover system and GPU resources used for Earth science.
This is not new. NASA has been in the supercomputing business since the 1970s, when Ames ran the Illiac IV, one of the earliest massively parallel machines, to do computational fluid dynamics before wind tunnels became too slow and too expensive for certain problems. The Numerical Aerodynamic Simulation programme followed in the 1980s with Cray hardware. In 2004 NASA stood up Columbia, named for the lost shuttle crew, and it briefly ranked second fastest in the world.
So the honest answer to the question is that NASA has been a serious high-performance computing shop for roughly fifty years, and aerodynamics was the reason long before climate and astrophysics joined the queue.
The main systems, and what each one is for
Names change as hardware is retired and replaced, so treat this as a snapshot rather than a permanent list. NASA refreshes nodes continuously, which means a single named system can be several generations of hardware stitched together over a decade.
| System | Location | Roughly when it came online | Main use |
|---|---|---|---|
| Pleiades | Ames Research Center, CA | 2008 | Long-running workhorse for aerospace CFD, astrophysics and mission modelling |
| Electra | Ames Research Center, CA | 2016 | Expanded capacity in a modular, highly power-efficient facility |
| Aitken | Ames Research Center, CA | 2019 | Current flagship at Ames, housed in the Modular Supercomputing Facility |
| Endeavour | Ames Research Center, CA | 2013 | Large shared-memory system for jobs that need huge single-node memory |
| Discover | Goddard Space Flight Center, MD | 2006 onwards | Weather, climate and Earth system modelling for the GEOS family of models |
What they actually compute
The work falls into a few buckets, and the examples are more specific than people expect.
Launch and vehicle design is the oldest one. Before the Space Launch System flew, engineers simulated the acoustic shockwaves bouncing off the launch pad, because sound energy at liftoff can shake hardware apart. Entry, descent and landing is another. The supersonic parachute that slowed Perseverance down over Mars was modelled on NASA systems, simulating fabric, cords and turbulent air interacting at Mach speeds, a problem you cannot test honestly anywhere on Earth.
Then there is Earth science. The Goddard systems run global atmospheric models at resolutions fine enough to resolve individual storm systems, plus reanalysis products like MERRA-2 that reconstruct decades of past climate from scattered observations. Astrophysicists use NASA allocations to simulate black hole mergers and the gravitational waves they emit, which is what gives LIGO something to match its detections against.
And a lot of it is unglamorous data processing. Kepler and TESS produced light curves for hundreds of thousands of stars that had to be searched, repeatedly, for dips of a few hundred parts per million.
Why NASA doesn't just rent cloud compute
It does, for some things. Earth observation archives sit in commercial cloud storage, and bursty, embarrassingly parallel jobs run fine there.
But the simulations above are tightly coupled. Thousands of processors have to exchange data with each other at every timestep, which makes the interconnect between nodes more important than the raw speed of any single chip. Standard cloud networking adds latency that wrecks performance on those jobs, and running them for months at a time on rented hardware costs more than owning the machine. That is the whole argument, and it applies equally to national labs in India and elsewhere.
The interesting part: NASA's compute is turning into AI compute
If you asked this question in 2015, the answer would have ended with simulation. It doesn't now.
NASA and IBM released Prithvi in 2023, a geospatial foundation model trained on Harmonized Landsat and Sentinel-2 imagery, and put it on Hugging Face for anyone to fine-tune. Flood mapping, burn scar detection and crop classification that previously needed a bespoke model per task now start from a pretrained base. The same partnership produced INDUS, a set of language models trained on Earth science, astrophysics, heliophysics and planetary science literature, because general purpose LLMs handle scientific terminology poorly. A weather and climate version of Prithvi followed with NVIDIA.
This changes what the hardware is for. Classic supercomputer nodes were CPU-heavy because fluid dynamics wanted double-precision maths. Training foundation models wants GPUs and lower precision. NASA's newer installations reflect that shift, and so does almost every other large scientific computing centre on the planet. Simulation is not going away. It is increasingly being paired with models that learn the shortcuts, so you run the expensive physics once and let a trained surrogate approximate the next ten thousand variations.
The part that matters if you're the one asking
Most people searching this are students, or curious engineers, and somewhere behind the question is a quieter one: do I need access to that kind of machine to do anything serious with AI?
No. That was true a decade ago and it is not true now. Prithvi is public. So are thousands of other pretrained models. The expensive part, the training run that burned through a GPU cluster for weeks, has already been paid for by someone else. What separates people who build useful things from people who read about them is whether they can take an existing model, fine-tune it on a specific problem, wire it into a workflow and ship something that works. A laptop and a free Colab notebook cover more ground than a 2010 supercomputer did.
The gap is almost never hardware. It is practice. If you want to move from understanding this to actually building with it, the AI Fellowship is eight weeks of hands-on project work, which is roughly how long it takes to stop reading about AI and start producing things with it. Tell us what you want to build and our team will tell you whether the next cohort fits.
NASA needs supercomputers because it is simulating physics nobody can test. You probably aren't. Act accordingly.
FAQs
1. Is NASA's supercomputer the fastest in the world?
No. The top of the TOP500 list is dominated by US Department of Energy machines like Frontier and El Capitan, along with systems in Japan and Europe. NASA's systems are large and well engineered but are built for sustained mission workloads rather than headline benchmark scores.
2. Where are NASA's supercomputers located?
The main high-performance computing facility is at Ames Research Center in Moffett Field, California, which hosts the NASA Advanced Supercomputing division. A second major site is the NASA Center for Climate Simulation at Goddard Space Flight Center in Greenbelt, Maryland.
3. Can students or outside researchers get time on NASA supercomputers?
Access is generally tied to NASA-funded or NASA-relevant research, requested through allocation processes run by the High-End Computing Capability project. It is not open to the general public, though many of the resulting datasets and models are released publicly.
4. Does NASA use artificial intelligence?
Extensively. Examples include the Prithvi geospatial foundation models built with IBM, the INDUS science language models, and onboard autonomy like the AEGIS software that lets Mars rovers pick their own targets for instrument scans without waiting for Earth.
5. What makes a supercomputer different from a powerful regular computer?
The interconnect, mostly. A supercomputer links thousands of nodes with very low latency networking so they can work on one problem together, exchanging data constantly, which ordinary servers and standard cloud networking cannot do efficiently.
6. Do you need a supercomputer to train AI models?
Only to train very large models from scratch. Fine-tuning an existing open model, running inference, or building an application on top of an API needs nothing more than a decent laptop and some cloud GPU time.