[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]  Rihard Jarc [@RihardJarc](/creator/twitter/RihardJarc) on x 51.7K followers Created: 2025-06-13 13:44:06 UTC An interview with a current high-ranking $META employee working in their silicon unit hints that the industry is starting to shift more towards custom silicon vs $NVDA GPU: X. He thinks GPUs have a lot of overhead that is wasted, but up until this point, the industry has not had a very good handle on what type of compute was needed for any AI/ML use case. According to him, we are now finally reaching a point where the industry has a better idea of what data movement, model, and compute look like in different types of AI models. This understanding for him means that the industry will start to lean towards something different than a traditional GPU. X. He believes people will begin to adopt a more customized approach. He also thinks we don't need a GPU to have flexibility for AI model training. X. He predicts that in X years, the split between GPUs and ASICs in training will be 60%/40%, and for inference, the ratio 50/50. X. The biggest hurdle for custom silicon vs $NVDA is CUDA. Additionally, the biggest problem for $AMD is again CUDA, as $AMD for him is caught between a rock and a hard place.  XXXXXX engagements  **Related Topics** [a very](/topic/a-very) [nvda](/topic/nvda) [meta](/topic/meta) [gpu](/topic/gpu) [silicon](/topic/silicon) [$meta](/topic/$meta) [meta platforms](/topic/meta-platforms) [stocks communication services](/topic/stocks-communication-services) [Post Link](https://x.com/RihardJarc/status/1933520758375931994)
[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]
Rihard Jarc @RihardJarc on x 51.7K followers
Created: 2025-06-13 13:44:06 UTC
An interview with a current high-ranking $META employee working in their silicon unit hints that the industry is starting to shift more towards custom silicon vs $NVDA GPU:
X. He thinks GPUs have a lot of overhead that is wasted, but up until this point, the industry has not had a very good handle on what type of compute was needed for any AI/ML use case. According to him, we are now finally reaching a point where the industry has a better idea of what data movement, model, and compute look like in different types of AI models. This understanding for him means that the industry will start to lean towards something different than a traditional GPU.
X. He believes people will begin to adopt a more customized approach. He also thinks we don't need a GPU to have flexibility for AI model training.
X. He predicts that in X years, the split between GPUs and ASICs in training will be 60%/40%, and for inference, the ratio 50/50.
X. The biggest hurdle for custom silicon vs $NVDA is CUDA. Additionally, the biggest problem for $AMD is again CUDA, as $AMD for him is caught between a rock and a hard place.
XXXXXX engagements
Related Topics a very nvda meta gpu silicon $meta meta platforms stocks communication services