A confusion in the evening triggered a systematic reflection, and in the end, I realized those bastards on Wall Street are either truly stupid or truly evil 😂 But I think it's quite likely they are just short-sighted and profit-driven 🥲

I originally wanted to write out the entire deduction process, but I really don't want to write that much. Here are my personal conclusions (don't blame me if I'm wrong).

1. China's current token cost is indeed much cheaper than the US's, partly due to algorithmic savings, and partly due to cheaper infrastructure and electricity costs. However, in the future, the US's token cost will drop exponentially with technological iteration. This then poses a question for China: can it maintain a token advantage? After all, infrastructure and electricity advantages are linear, while technological iteration is exponential.

2. Large language models will truly be a highly profitable business model in the future. Clearly, not many more large models will emerge. Existing large model companies will also form a certain scale of oligopoly. As the practicality of large models increases (practical application scenarios are already visible), token costs will drop exponentially, and operating revenue will also continue to drop exponentially. This business model is simply a cash cow. When token costs drop exponentially to one percent of what they are now, becoming almost negligible, the profit margins will be similar to those of software and internet companies.

3. Judging chip companies

(1) For training chips, NVIDIA is clearly the king. Google has its own unique strengths, and some ASIC chips might also rise unexpectedly, but only NVIDIA is the king.

(2) In the inference era, when large models become similar, competition will be about who has lower costs.

From a macro perspective, GPUs are naturally inferior to ASICs. This is also why Jensen Huang is truly insightful, decisively acquiring Groq. Otherwise, the future would only belong to ASICs. Currently, because NVIDIA acquired Groq, combined with NVIDIA GPU's versatility, the CUDA ecosystem, and NVIDIA's continuous iteration, NVIDIA can still maintain a significant market share. However, with the emergence of other major companies using their own ASICs, it will inevitably lose a lot of share. But considering NVIDIA's versatility, it will certainly retain some share. However, it's really hard to see hope for AMD; it's probably destined to remain a follower.

(3) In the mature era of large models, with the emergence of lobster-like applications, globally renowned companies will obviously deploy large models in their own data centers. This market clearly belongs to NVIDIA and AMD, with NVIDIA likely taking the lion's share and AMD a smaller portion. I judge this global market to be at the trillion-dollar level. Clearly, both have a bright future 😃

(4) For individual users, obviously wealthy people will deploy locally. For those with average conditions, will data centers in the future start to value individual users, offering encrypted services to deploy a personal Jarvis in the data center? This would give data centers a bright future too. Or could large model companies themselves encrypt each person's applications? Or could the lobster itself be encrypted? I really haven't figured it out 🤔

In summary, those bastards on Wall Street are really 🐶

$NVIDIA(NVDA.US)$Alphabet - C(GOOG.US)$Broadcom(AVGO.US)$AMD(AMD.US)$Nebius(NBIS.US)$Coreweave(CRWV.US)$Microsoft(MSFT.US)$Amazon(AMZN.US)