Some thoughts about AI investment – difference of the East and the West

Happy New Year. Hope all of you and your family had a good one.

In the East and the West, there seems to be a different preference due to the scope of available candidates. In Asia, quite many investors argue AI infrastructure investment could fade out soon after the upstream chip CoWoS capacity bottleneck is eased, as they believe Microsoft CoPilot subscribers of 1M+ (up 40% QoQ) is not satisfying enough, meaning AI hard to be monetized.

Meanwhile, in the US, investors understand Microsoft CoPilot is barely one of the many applications. GitHub also sees rising subscribers due to help on productivity gain (up 55%) and >200K developers contributing to AI projects on Github, with >300K models being published on Hugging Face. Not to mention Stable Diffusion and other rising AI pre-trained models open source.

Indeed we are seeing AI self-evolving themselves - LLM evolving fast to the level that training time has shortened for the same amount of benchmarks. But new real-time data are added to the training data pool. The difference between GPT and BERT is that GPT has more expertise in “completing” tasks in natural language, while BERT has more expertise in “understanding” tasks.

At the time of writing, ChatGPT 5.0 is coming soon, and soon the others within Magnificent 7 or startups will strive to catch up to provide more efficient models. In terms of model, we will likely see both LLM model and the smaller edge model. For LLM model, currently, performance matters the most. For smaller edge model, it would be fair to predict that both power efficiency and cost might matter more than performance, just as our mobile computing devices.

In terms of fundamentals, AI infrastructure investment is still at the beginning for the majority. Stock market reaction is fast due to supply chain check, and the earliest beneficiary (Nvidia) might find its peak earlier than others, but for many of the downstream names, it is risky to follow some Asian financial community's more bearish conclusion that infrastructure peak is in 1H24.

Any thoughts welcome!