Mad, China, and the Semiconductor Showdown (Part 2)
eastwind.substack.com
Mad, China, and the Semiconductor Showdown (Part 2)
1–3 of 3 posts
Re: Mad, China, and the Semiconductor Showdown (Part 2)
#2I recently analyzed some fascinating financing trends as the US and China took very different paths in advancing their AI ecosystems (Part 1 of my Substack post [here](https://eastwind.substack.com/p/mad-china-and-the-semiconduc...))
In Part 2 of my Substack mini-series, I highlight why access to semiconductors is crucial to maintaining the pace of innovation in AI, and why up until the past several years, access to leading-edge onshore fabrication capabilities has been less of a strategic focus.
I contend that geopolitical tensions in Asia and our reliance on the supply of AI accelerators manufactured in Taiwan creates a dangerous reflexive loop, and how that affects the ability for companies in the space (like OpenAI) to continually innovate. I also explore some interesting areas of opportunity that entrepreneurs may look to exploit.
Would love thoughts & comments!
Re: Mad, China, and the Semiconductor Showdown (Part 2)
#3I recently analyzed some fascinating financing trends as the US and China took very different paths in advancing their AI ecosystems (Part 1 of my Substack post [here]( https://eastwind.substack.com/p/mad-china-and-the-semiconduc... )) In Part 2 of my Substack mini-series, I highlight why access to semiconductors is crucial to maintaining the pace of innovation in AI, and why up until the past several years, access t…
Since AI training is highly parallelizable one can look at it as a question of cost for training, that is given enough electric power, any nation state can train the model they want. And it would be nice to know how much cost reduction is due to fab process advance and how much is due to architecture and software advances. My suspicion is that the fab node is not strategic. It is a commercial issue to integrate more cores as fab advances, but training can be done over more chips and boards if integration isn't there, maybe for higher cost.