Sooner or later, lots of competitors including Cerebras are going to take apart Nvidia's data center market share and it will cause many AI model firms to question the unnecessary spend and hoarding of GPUs.
OpenAI is still developing their own chips with Broadcom, but they are not operational yet. So for now, they're buying GPUs from Nvidia to build up their own revenue income (to later spend it on their own chips)
By 2030, eventually many companies will be looking for alternatives to Nvidia like Cerebras or Lightmatter for both training and inference use-cases.
For example [0] Meta just acquired a chip startup for this exact reason - "An alternative to training AI systems" and "to cut infrastructure costs linked to its spending on advanced AI tools.".
There's so much optimization to be made when developing the model and the hardware it runs on, most of the big players are likely to run a non-trivial percentage of their workloads on proprietary chips eventually.
If that's 5 years into the future, that looks bad for Nvidia, if it's >10 years in the future, that doesn't affect Nvidia's current stock price very much.
OpenAI is still developing their own chips with Broadcom, but they are not operational yet. So for now, they're buying GPUs from Nvidia to build up their own revenue income (to later spend it on their own chips)
By 2030, eventually many companies will be looking for alternatives to Nvidia like Cerebras or Lightmatter for both training and inference use-cases.
For example [0] Meta just acquired a chip startup for this exact reason - "An alternative to training AI systems" and "to cut infrastructure costs linked to its spending on advanced AI tools.".
[0] https://www.reuters.com/business/meta-buy-chip-startup-rivos...