In plain words
A supercomputer is a computer far more powerful than an ordinary one, usually thousands of processors working together. For decades they mostly ran scientific simulations. AI supercomputers are built for a different job: training and running huge AI models.
Inside, they are mostly GPUs or other AI chips, tens or hundreds of thousands of them, wired so tightly that they work on one model at once. That is why people also call them training clusters or AI data centers: the whole building is the computer.
They are growing fast. Epoch AI, a research group, found that the performance of leading AI supercomputers doubled about every nine months from 2019 to 2025. By February 2026 it estimated the largest, Colossus 2 in Memphis, USA, had the computing power of about 1.1 million Nvidia H100 chips.
Everyday example
A board exam marking centre
Picture a huge hall where thousands of teachers mark board exam papers. Each takes a bundle, everyone works at once, and the totals are tallied together. A good centre needs more than teachers: trolleys to move papers, lights, fans and a fast tally desk. An AI supercomputer is the same. Chips do the marking, but networks, power and cooling decide how fast the whole hall finishes.
Where the example breaks: Exam papers are separate jobs. In AI training, all the chips work on one model, so they must share results constantly, not just once at the end.
How it actually works
An AI supercomputer is a system, not a chip. Every part has to keep up with the chips.
Chips
The engines are GPUs or custom chips such as Google's TPUs. Epoch AI found the performance of top systems grew about 2.5 times a year: roughly 1.6 times from using more chips and 1.6 times from better chips.
Network
Chips are linked in two layers: very fast links inside each rack, and a network between racks. If the network is slow, chips sit idle waiting for each other's results.
Power and cooling
The largest sites now approach a gigawatt of power, Epoch AI estimates, enough for roughly a million US homes. The newest racks are cooled with liquid instead of air. See AI energy use.
Not the same race as TOP500
The famous TOP500 list ranks supercomputers on 64-bit maths, the precision most science needs. AI gets good results at 32-bit, 16-bit or lower precision, so the two kinds of power are measured differently. In June 2026 the TOP500 leader was LineShine in Shenzhen, China, which runs only on ordinary processors.
Who owns them
Companies held about 80% of AI supercomputer power in 2025, up from about 40% in 2019, and the US about 75% (Epoch AI). India's government-run AIRAWAT in Pune offers 200 AI petaflops and ranked 75th on the TOP500 in 2023.
- Growth
- Leading AI supercomputers doubled in performance about every 9 months from 2019 to 2025 (Epoch AI)
- Largest known, early 2026
- Colossus 2 in Memphis: about 1.1 million H100-equivalents (Epoch AI estimate, February 2026)
- Cost and power
- The largest approach a gigawatt; a 1 GW site costs roughly $38 billion to build (Epoch AI)
- TOP500, June 2026
- No. 1 LineShine (China), 2.2 exaflops at 64-bit; No. 2 El Capitan (US), 1.8 exaflops
Questions people ask
What is the difference between a supercomputer and an AI supercomputer?
Classic supercomputers are built for science and measured on 64-bit maths. AI supercomputers are built from AI chips tuned for lower-precision maths, and they spend their time training and running AI models.
What is the most powerful AI supercomputer?
As of Epoch AI's February 2026 estimate, the largest known was Colossus 2 in Memphis, with computing power equal to about 1.1 million Nvidia H100 chips. The leaders change quickly, and companies do not always publish details.
Does India have an AI supercomputer?
Yes. AIRAWAT, run by C-DAC in Pune, offers 200 AI petaflops and ranked 75th on the TOP500 list in 2023. The IndiaAI Mission also rents out more than 38,000 GPUs through a shared portal (March 2026).
Is an AI supercomputer the same as a data center?
Often, in practice. A data center is the building with its power and cooling. An AI supercomputer is the linked set of chips inside that works as one machine. The biggest ones fill whole buildings or campuses.
How much electricity does an AI supercomputer use?
The largest now approach a gigawatt, according to Epoch AI, which is roughly what a million US homes use. Their power needs roughly doubled every year between 2019 and 2025.
Was this clear?
1 is confusing, 5 is crystal clear. One tap, no account needed.
Spotted something wrong, or know a better example? Suggest a change