Finding the Azure GPT-5 SuperCluster. Serious Replies ONLY

Posted by Crazyscientist1024@reddit | LocalLLaMA | View on Reddit | 9 comments

As we know according to multiple leaks, GPT-5 has most likely began training in Azure Datacenters. I wanted to start this thread to see if we can find as a group effort: When GPT-5 started training, Where is it trained, and how long is it training for. OpenAI is based in San Francisco so I am going to guess that they will train the model on US-West or the US-West 2 datacenter for minimum latency (If there are any other possibilities comment here). We know that datacenters takes lots of water and electricity. I believe the plan here is to find the water and electricity usage peaks around the datacenter areas to find the data we need.

Reply to Post

9 Comments

[-]

byebrew@reddit

It doesn't seem too easy to find water and electricity usage peaks when you don't have access to that information. Also, why? There's no point in finding it

[-]

Due-Memory-6957@reddit

OP is going to be the first person here to say they hate OpenAI and do something about it.

[-]

Open_Channel_8626@reddit

Data centres often don’t even record water usage data

[-]

Everlier@reddit

I also don't think that training this model would significantly stand out from other workloads of a large datacenter. I've read that their largest datacenter is in Washington, so might be there.

[-]

CartographerExtra395@reddit

Just clarifying some terms, a region (ex us west) is fed by many data centers

[-]

danerchri@reddit

That's a needle in a hay stack. Even if your assumptions are right, you won't find peaks. These facilities have extremely high utilization rates. Everything is load balanced and dynamically allocated. That means that hypothetically if they wanted those locations, other lower tier loads would move to other data centers. They'd get displaced and the overall load would remain unchanged. There's an AI silicon shortage so everyone not serving a live service where response time is important will run their load off peak when it's cheaper. Open AI doesn't care how much horsepower they use THIS hour, they care that they get X many hours of processing time in to compete their training before their chosen delivery date distributed across whatever hardware gives them the lowest cost to satisfy their goals.

[-]

CartographerExtra395@reddit

I could nit pick - allocation part is a rabbit hole - but your thesis is more or less right.

[-]

FlishFlashman@reddit

The latency between OpenAI and the datacenter where the training is going on doesn't matter because all the important and latency sensitive stuff is going to be happening at the data center. I doubt you'll be able to get water and electricity usage information with sufficient precision and granularity. You might be able to "find" data on spot pricing on the largest GPU instances Azure has available. I'm sure someone besides Microsoft collects it.

[-]

multiverse_fan@reddit

I have a lot of serious questions. Number one - how dare you?