There are a few big companies investing in TPU architectures (afaik, mostly Meta and OpenAI). All their stuff is planned to be exclusively manufactured by TSMC. Not because they want that choke point, but because it’s the only fab on the planet with the ability to produce them
If China invades Taiwan and TSMC blows up the fabs to deny them the chips, the whole world loses the ability. For years, not months.
(Edit: FWIW the TPUs are 3nm chips with crazy advanced HBM, China is not making them any time soon)
China’s not interested anymore. The PRC is “encouraging” companies to buy locally developed chips.
They’re not going to develop GPUs
GPUs are not the target. Purpose-built AI chips are.
There are a few big companies investing in TPU architectures (afaik, mostly Meta and OpenAI). All their stuff is planned to be exclusively manufactured by TSMC. Not because they want that choke point, but because it’s the only fab on the planet with the ability to produce them
If China invades Taiwan and TSMC blows up the fabs to deny them the chips, the whole world loses the ability. For years, not months.
(Edit: FWIW the TPUs are 3nm chips with crazy advanced HBM, China is not making them any time soon)
Lol what? They already did.
Not ones to compete with NVidia Not ones that will train GPT5/6/7
The… That’s not
Enough data to train a hypothetical gpt6 literally does not exist.
Most frontier models since GPT4 actually use synthetic pre-training data , before slurping up the Internet and after along with RLHF
too many fingerssausage fingersAlso training data is totally independent of chip feature size, interconnects, memory bandwidth, and all the things actually relevant to chips