The AI bust is here

For anyone who doesn't realise how similar this is to the dotcom bubble, here is a thought experiment.

In 1998, you would have spent 2 years hearing about Altavista and the new search engine paradigm with curated hyperlinks. They are buying lots and lots of Cisco gear to serve their network. You get excited and invest $100B into Altavista and Cisco.

Then all of a sudden, a startup called Google shows up with a cheaper and more efficient way of presenting search, with much less networking equipment.

What do you think happens to the $100B invested with Altavista and Cisco? What do you think happens to OpenAI and Nvidia?

I think a big difference between dotcom bubble and ai bubble is:

where is the business/how it profit?

there're solid new-born business with dotcom, like amazon and microsoft

but who are the customers of AI(or LLM, to be specific)? for now i can only see limited user such as programers

how will business gain profits by LLMs

and it feels like users are already consumed totally by existing internet and apps, it's a stock market rather than an incremental market

This is not like a gold mine and shovels, it's more like a cliff and ropes

Open source and/or low precision LLMs have been around for years and people have been hacking on them and doing all sorts of things with them. AI is incredibly overhyped and was due for commoditization. If a Chinese company releasing yet another decent but somehow this time incredibly overhyped LLM is what it takes to get investors to see the writing on the wall, then good riddance.

This may not affect the total demand, but the distribution of demand. What's also affected, in my opinion:

- individual company valuations

- capacity planning models ("what would it take to do x?") and related funding

- overall competitive landscape (yes you can compete with OpenAI after all without a trillion dollars in funding)

- it's not the coveted high-end GPUs anymore, more affordable/available hardware may also work

It would be well-deserved and helpful towards culling many AI services, which are essentially snake oil. For example, the science and research on lie detection using image and audio analysis is tenuous at best, but there companies purporting to offer services such as these. On the other hand, such services are perhaps useful for milking money and resources from unscrupulous actors whose primary goal is “cover your ass”. So, essentially the McKinsey business model.

The fact that o3 exists moots any point that relies on AI becoming cheap. When/if deepseek

At most this will force OpenAI to hire some real engineers and not just new grads who are easy to exploit.

I do consulting for multiple GPU hosting companies. DeepSeek means more things can be done with less and the demand for access to GPU has only increased.

The analysts, analysts who do not code, analysts who don't run GPU clouds, are wrong.

The anals are wrong.