Super insiteful piece on the hardware layer! The framing of custom silicon as "hedging" rather than competing really resonated - i've seen so many ppl misunderstand this dynamic. Memory bandwidth being the actual constraint hits home; most debates focus on FLOPS when the real bottleneck is data movement.
GPUs are currently very important for AI as outlined in the article above, but it isn’t set in stone that this will be the case forever and especially for NVIDIA. Things like Google’s TPUs could potentially cut a share of NVIDIA GPUs - mostly in Google Cloud and for existing models.
Custom silicon from other companies could potentially also replace the very expensive NVIDIA GPUs for certain tasks. So if AI becomes boring infrastructure business where margins shrink and efficiency count beyond hype, competition for both NVIDIA and especially Nebius might become fierce.
Thanks again for your comments everyone, I've included Nebius and NVIDIA in my latest article about TPUs: https://techfundamentals.substack.com/p/why-google-built-tpus-and-why-investors
Super insiteful piece on the hardware layer! The framing of custom silicon as "hedging" rather than competing really resonated - i've seen so many ppl misunderstand this dynamic. Memory bandwidth being the actual constraint hits home; most debates focus on FLOPS when the real bottleneck is data movement.
Amazing, glad you liked it. Thanks for taking the time to read this!
What will this dependence on Nvidia mean for nebius?
Thanks for your comment!
In my opinion Nebius acts like a typical AI wrapper on a NVIDIA hardware level (I wrote about AI wrappers here: https://open.substack.com/pub/techfundamentals/p/the-ai-boom-has-a-graveyard-no-one). This means Nebius whole existence depends on prices, margins and demand for NVIDIA GPUs.
GPUs are currently very important for AI as outlined in the article above, but it isn’t set in stone that this will be the case forever and especially for NVIDIA. Things like Google’s TPUs could potentially cut a share of NVIDIA GPUs - mostly in Google Cloud and for existing models.
Custom silicon from other companies could potentially also replace the very expensive NVIDIA GPUs for certain tasks. So if AI becomes boring infrastructure business where margins shrink and efficiency count beyond hype, competition for both NVIDIA and especially Nebius might become fierce.
Big Tech isn't trying to beat the GPU. They are constructing an escape hatch to cap their exposure to a single supplier's leverage. 😬
Exactly. I'll write another post about Google's TPU technology next week. That'll also show a bit more where big tech is heading with their efforts.