Hacker News Reader

Comment by minimaxir | Hacker News Reader

Neither Google nor OpenAI's infra could handle the potential scale of Apple's userbase, given the scaling issues both companies are having.

Even Google doesn't have enough GPUs.

View on HN

The scale of 30% of world's mobile phone versus Google's 70% Android market share, or the 10% of desktop/laptop market versus 72% from Windows?

But currently only a small subset of Android devices come with all of Google’s current phone AI features out of the box, Pixel 8 Pro and the Samsung S24 lineup.

You can install the Gemini app on other phones, but last time I checked the install counts, it’s only a small percentage of the Android user base.

(I’d also be surprised if a large population would be willing to pay the $20 monthly Google One tier to use Gemini for private use.)

I can't even use text summarization on my Pixel 8 Pro.

do you use Chrome browser? It doesn't work for me either but I use firefox mobile, so I suspect that may be why.

No, afaik, it's still a region locked feature. US only. Pretty convenient that all popular product reviewers are US based and advertise the Pixel 8 as an AI powered device.

Gah, that's obnoxious. Actually that might explain why it didn't work for me. I tried it when I was in Pakistan :-P

Gemini doesn't run on GPUs, it runs on Google's custom TPU silicon.

The TPU v4 was already competitive with Nvidia, and the new v5p is 2.4 times faster and more power efficient than v4.

That may be true: the economics of TPUs are more complicated but they certainly aren't as much in demand.

> aren't as much in demand.

Google gets to manufacture that demand at scale.

That's primarily just because Google is doing a poor job offering them to people. Bad documentation, and few trust Google enough to want to tie themselves to Google's trusted hardware.

Also, Google is using all these TPUs for their own training and inference at the moment.

Do you have a solid source for this claim?

Everything I have read/heard tells me that companies would much, much rather have access to Nvidia chips than Googles TPUs.

I think there are different size Gemini models and some of them can fit on a phone with TPUs or similar and since each user will use their own phone, it offloads the computation.

Because TPUs are aimed at Google workloads so probably not as easy to work with outside of Google.

NVIDIA chips are better for training due to the sheer prevalence of CUDA and NVIDIA being ahead of the curve in providing data centre hardware to scale up. They’re also not keeping stuff for just themselves.

When it comes to inference, most of that isn’t really relevant. TPUs or any other specialized hardware will have a higher ROI. But Google aren’t giving them to others either

I know you’re asking for a link, but it’s not really something other than Google can prove. However it is a reasonable extrapolation for anyone familiar with the product space to make.

"Google will quite literally have more TPUv5's than OpenAl, Meta, CoreWeave, Oracle, and Amazon will have GPUs combined"

https://www.semianalysis.com/p/google-gemini-eats-the-world-...

Reply on news.ycombinator.com