Hacker News Reader

I don't see how that math works? This is a $15k rig under benchmark and per the results it competes very acceptably against... one consumer GPU.

I really don't see who buys this, except people who want the Studio for some other reason. But nothing in the story says you want to fill racks with these instead of Blackwell or TPU parts; it's not even close.

View on HN

Your math is correct, but it’s math based on today’s economics.

Think of a company like Apple moving onto your turf. They’re not going to cede AI to the cloud. They want their part of the pie.

So in 7 years, how much AI will be handled locally on your iPhone. And will you have repaid all the debt on your balance sheet before Apple eats your lunch

I think most important thing is that Nvidia doesnt want to give 100% of market to frontier AI labs either.

It's way too easy for 1T+ frontier labs to ditch Nvidia. So Nvidia will also put effort to make sure there are open weights models and local hardware available.

And Apple will benefit from this too.

There is zero chance that an LLM approximating a modern frontier model is going to be running on a phone in the next decade. Even if you grant that you could stack enough DRAM dies on top of each other in the package, that would be a three order of magnitude improvement in power efficiency just for the compute.

Reply on news.ycombinator.com