[cm2187]

What moat? I see people reimplementing open source models in a matter of weeks. These are not huge code bases. And when you want to run those models at scale, every cost saving will make big differences given their enormous computing requirements. It think it will become a cut-throat market.

[reaperman]

NVidia doesn’t sell models so it doesn’t matter to them if individual models have moats.

[cm2187]

It's like saying that intel doesn't sell OS, so it doesn't matter to intel if windows is married to x86

[reaperman]

Windows is also compiled for ARM and Microsoft literally sells their own Surface Pro computers with ARM processors[0]. So this is a terrible analogy, for a lot more reasons than that. Rather than go through all the reasons it’s a bad analogy I’ll just get straight to NVidia’s long term moat:

NVidia has 6x the cash-on-hand (>$30B) than their closest competitor (AMD) ($5B), and all of that can go towards GPU R&D rather than being split between CPU+GPU R&D (AMD has 10% of the $65B GPU market share but 33% of the $50B CPU market share — their CPU business is almost 75% of their total business). If there are changes needed to stay competitive, NVidia has the warchest they need to adapt. Just their cash on hand would allow them to fumble an entire release cycle and still catch back up.

Sure, if ROCm ever achieves parity with CUDA then Nvidia’s margins will finally decrease but the only way they’re facing any existential threat is by massively fumbling internally. If you have reason to believe that Nvidia will start executing poorly, then that would potentially be a valid concern — but so far they’re continuing to do great work despite their massive lead and their prices reflect their technical lead over their competitors.

0: https://www.theverge.com/2024/6/26/24186432/microsoft-window...

[cm2187]

On windows, you forget that it wasn't the case historically that windows could run on ARM, and it took years for MSFT to make it work (hence it was a real moat), and now that it works it is a real threat to Intel's grip on the PC market. So it exactly proves my point that you want to know how locked in your customers are to your product.

And as Bezos says, your margin is my opportunity, and nvidia makes humongous margins right now. It will attract investments in competitors. And if a competing GPU is half as powerful but a third of the price, it will eat market shares and compress margins. Unless customers cannot switch, but I don't believe that to be the case.

[totallywrong]

CUDA with all its developer mindshare.

[madaxe_again]

And pre-booked fab capacity, which is the major bottleneck.

[cm2187]

Training those models costs 10s of millions. Running them at scale some order(s) of magnitude that. At one point those AI initiatives will need to return a profit. You really think developers will not have to learn a new API? Users inertia works for retail users, I don't think it works so much for people who get paid to do it or have a financial incentive to switch. Provided there is a suitable alternative.