A Data-center traditionally consists of storage (disk space), compute (CPU), and networking infrastructure. Sometimes they would also contain GPUs to perform some complex math faster, but for the most part these weren’t computationally intensive.

Storing data on disk requires very minimal power to keep them running. These machines weren’t spinning the CPUs too much most of the time.

The software that use up the CPUs were also generally “stateless” in the sense that you could move the data to a different data-center, install the software on the machines and then point the machines at this data and run them. An end user wouldnt be able to tell the difference.

This means that, the most important “asset” in a data-center, is actually the “data”. Software running in them can be recreated by various means to various degrees.

For this reason, calling the data-center a “data” center made perfect sense.

AI Data-center runs AI but is fundamentally a different kind of workload.

When companies want to build data-centers to “run AI”, they are doing a fundamentally different thing than what was happenning inside data-centers previously.

“Running AI” means running complex math - which are typically modelled as matrix multiplications - as fast as possible. A CPU is very slow for such kind of multiplication at scale, that they run it on GPU.

GPU stands for Graphics Processing Unit and was used for drawing stuff on the screen - hence the prefix Graphics. CPU stands for Central Processing Unit and is the more general purpose processor between the two and can be used to do anything on a computer.

An AI data-center therefore consists of rack after rack of GPU clusters continuously spinning, cracking numbers.

A typical AI data-center is mostly the GPU, some amount of AI model weights and a very small amount of user data.

An AI model weights is just trillions of numbers in a file that needs to be loaded into the GPU onto which computations happen. All AI models (GPT, Claude, Kimi, Qwen, Mistral etc) are all essentially just a large collection of numbers.

Most of the user context while you chat is stored directly on the GPU for the duration of the session. And then it is cleared. Any questions, or summary of the conversation will be stored on disk for later retrieval.

The user data generated by AI data-centers are not business critical

For example, suppose you buy something from Amazon, and amazon stores an invoice, that is business critical data.

On the other hand, if you asked something cats, and if your next conversation with AI does not know about that, then it is not the end of the world.

This is not to say that these centers do not store any business critical data, but the amount of such data they store is minuscule compared to what they actually need to do to make sure they continue providing the service - which is performing matrix multiplications and generating an answer to your next question in the process.

Compute is power hungry

Even the large data-centers created by humanity so far, have been consuming amount 100-200MW power. The largest of them would be around 600MW. (Source: trust me bro! - or you can search it up yourselves).

The commitments for building AI data-centers are all for a minimum of 1GW power. With the most ambitious projects reaching about 10GW of power.

The reason they need this much power (and not storage space to store “data”) is because a fully utilized GPU that is crunching numbers would need enormous amounts of power. They also generate lot of heat in the process.

The primary asset in an AI data-center is not “Data”. It is “GPU” and “Power delivery”. One cannot simply relocate a data center and resume operations from a different place because these “GPUs” are expensive to buy and install in a new place, or to move it from one place to another at such a massive scale when they are already installed.

Further, getting the “power delivery” and cooling solutions set up in the new location might be next to impossible in the near future.

Hence, I propose we call computation heavy centers such as these as Compute-centers.

What difference does it make?

It provides some clarity while discussing AI. Saying no to AI data-center is not the same thing as saying no to a data-center that serves emails, or videos. People who are against AI data-centers are not against regular data-centers.

Why would people be against these Compute-centers?

• They use up lot of power in the area - could cause blackouts.
• They need lot of water to cool. Air cooled systems also exist.
• They generate a constant hum (as they are constantly crunching numbers) that will be heard for upto 8 kms around the place.
• They require lot of space.
• They do not generate any new employment opportunities for the people living nearby. These data-centers require expert technical engineers to set them up.
• Once they are created, the maintenance doesn’t require too much man power. So compared to setting up a conventional factory, even employment is not generated.
• Most importantly, as of today, they do not generate any return on investment for whoever invests in the data center itself.

Okay, I’m sold!

Data-center GOOD. Compute-center BAD.