[ 3 / biz / cgl / ck / diy / fa / ic / jp / lit / sci / vr / vt ] [ index / top / reports ] [ become a patron ] [ status ]
2023-11: Warosu is now out of extended maintenance.

/biz/ - Business & Finance

Search:


View post   

>> No.57831049 [View]
File: 353 KB, 1600x1067, 1684192052609442.jpg [View same] [iqdb] [saucenao] [google]
57831049

Here's the truth about AI driven chip demand that everyone is missing. It's all a mirage.

Training the models is computationally very demanding. But running them is extremely cheap by comparison. To train a model you need billions or trillions of iterations. To simply run the model as a user, to get an output from inputs, you run it once.

This is fundamentally different from the software paradigm that has driven demand for chips in the traditional software era. Software development is extremely cheap by comparison to running it at scale to serve users and consumers, which is where the sustainable source of massive compute demand comes from.

Now as the AI dev ecosystem matures you will have base models developed by a few large companies that are licensed out and fine tuned by specialists which are then used at scale, similar to the current OS and software ecosystem.

The issue for chip demand is that training the base models is largely a 1 time thing. Sure new architectures will come out that will require training from scratch, but that won't happen often. Currently we're in the phase of the cycle where many companies are training many base models. But, over time the number of companies participating in the creation of base models will decline and since the dominant base models have already been trained, all that will be needed will be to fine tune them or give the occasional update.

When that happens, compute demand for training models will collapse. Yes you will have many people using the models, but I can't stress enough how cheap they are to run. You could serve millions of people's queries to ChatGPT on a few GPUs. This doesn't even take into account the GPT tailored hardware that will make things even more efficient.

So basically, chip demand for AI is development driven not consumer driven which is the inverse of the current software dev paradigm and which is not sustainable for long term chip sales. Current chip sales are unsustainably hype driven.

>> No.54978704 [View]
File: 353 KB, 1600x1067, 1684125710654379.jpg [View same] [iqdb] [saucenao] [google]
54978704

>THAT'S HIM!
>THAT'S the guy was bragging about holding Chainlink tokens!

Navigation
View posts[+24][+48][+96]