[彻夜难眠●不敢相信] Nvidia, Super Micro Computer all 惨重大跌.... whats happening to AI theme?

[Jirachi]

Vix now in the fear zone confirm market gg later

These type of VIX needs a few days to settle down. Likely after FOMC then.

 

[(o.o)y]

So deepseek ish real? 山寨版 really better than original?

Better, much better.

 

[focus1974]

The DeepSeek thing is likely overblown.

There are a lot of LLMs on Ollama with FP8 quantisation and comparable parameters size. I really think this is a China operative.

yes, but the key thing is it is threatening the AI providers (like AMZN, MSFT, GOOG, OpenAI, etc etc) profit margin.
Imagine last time can charge $20 monthly or even $200 monthly for some.
Now , you have to charge $2 monthly to compete with Deepseek.

Profit margin evaporated 90% at the minimal .. means your profit projections on the AI part of the companies all GG.

Now analysts should be busy readjusting the forward earnings and funds are just throwing first buy later.


The scary thing is .. Deepseek has gone viral today. The Deepseek API is GG now.
https://status.deepseek.com/

Have been using it to try out the coding using API and it is performing as good as openai or sonet.
but at a fractino of the cost.

I used so long ... now only used up usd$1.20 for 10 millions tokens(input/output)!!
This would not be possible with OpenAI or Sonet.

 

[LoUsyGamER]

Is this what we called the butterfly effect? Something happened so far away from America also can cause a huge tsunami on the stock market...

This is called timely correction.

There needs to be some news to trigger correction.

 

[mooseolly]

Ho seh liao, big red ang bao from market.

 

[Jirachi]

yes, but the key thing is it is threatening the AI providers (like AMZN, MSFT, GOOG, OpenAI, etc etc) profit margin.
Imagine last time can charge $20 monthly or even $200 monthly for some.
Now , you have to charge $2 monthly to compete with Deepseek.

Profit margin evaporated 90% at the minimal .. means your profit projections on the AI part of the companies all GG.

Now analysts should be busy readjusting the forward earnings and funds are just throwing first buy later.

Yes. I understand where you are coming from

 

[artncraft]

So it's really possible? It's like a few young chaps in China can trump all the geniuses in Google/ms/meta...etc.

Because it uses the synthetic (derived) data from ChatGPT instead of mining its raw data.

 

[Clarence NMT]

So deepseek ish real? 山寨版 really better than original?

Thisnkind of thing keep on have best, thebthing only is free open source. Is like u charge some amount, ppl charge free, is like spoil market

 

[Nevereatrice]

Later market open will have another round of sell off.

Then after lunch time the real sell off will come.

I hope you all have already sold all your stocks when I warn about this crash 1 month ago.

I already off hands nvda after i saw the monthly chart coming off last year. Thought this year would be year of small-mid caps but now market will be pulled down as a whole by heavy weights

 

[metabo]

So how? Buy the dip? Or bubble burst?

 

[coyote]

This is called timely correction.

There needs to be some news to trigger correction.

So wall street is thanking China?

 

[ThisIsSparta]

bought when they announce earnings together with share split that time
sold when edmw got that thread about nvda to the moon
personally for me, when stupid people start to buy, it is indicator to sell

anyway, time for the know-it-all edmw bears to spam "i told u so" and celebrate

 

[AuraKUPO]

I already off hands nvda after i saw the monthly chart coming off last year. Thought this year would be year of small-mid caps but now market will be pulled down as a whole by heavy weights

That's good bro. No loss is a win already.

 

[inf|uencer]

deepseek is from hangzhou i have china friends from hangzhou i will go there one day

 

[cosmothecat]

WL WL WL SM D WL

Shorting now!

 

[selemuse]

Talking about scale, it certain can scale with more hardware, develop more sophisticated model in shorter time.

it's still important but it's overvalued since deepseek proven that you can do more with less

You are the one who has no idea what is going on.

US do not want Nvidia's better chips to go to china already and with China making their AI open sourced means many people know they buy Nvidia's lower end chips can do as good as open AI etc US models.

So why people will buy Nvidia's better chips?

Unless US has major breakthrough, so short term demand for Nvidia chip sure go down

Of course, I know what i talking. I work in this industry, for your info.

 

[dontwastetime]

All i know is this

Stock dropped 12% = 700k loss

1% = 700/12=58k

100% = 5.8m

What did u do man ? Leverage ? But I dun think so.. pls explain

 

[mooseolly]

Remember when TikTok came out and everyone is like Meta, Google is doomed?

 

[joshwong11]

yes, but the key thing is it is threatening the AI providers (like AMZN, MSFT, GOOG, OpenAI, etc etc) profit margin.
Imagine last time can charge $20 monthly or even $200 monthly for some.
Now , you have to charge $2 monthly to compete with Deepseek.

Profit margin evaporated 90% at the minimal .. means your profit projections on the AI part of the companies all GG.

Now analysts should be busy readjusting the forward earnings and funds are just throwing first buy later.

lmao, wad talking u. deepseek approach is same as openai o1 onwards just that it costs less since they found way to optimise nvidia amd hardware. there is nothing preventing openai n other companies from doing the same n they can do it with such scale and abundance that make them even more profitable. openai still have o3-o4 and gpt5 in the pipeline. anyway openai from o1 onwards also the same minimizing costs making their model more efficient so they can sell their API at profit since there are no challenger at that time.
anyway all these companies u listed will be the main beneficiaries instead since they require less compute/costs to offer AI solutions to existing real customers. all can scale up faster.

 

[focus1974]

Because it uses the synthetic (derived) data from ChatGPT instead of mining its raw data.

yup. Now most LLM shoudl be moving to the next phase of sythentic and time-scaled approach.

The three main phases of LLM training are still generally recognized as:

1. Pre-training (self-supervised learning)
2. Instruction fine-tuning (supervised learning)
3. Reinforcement Learning from Human Feedback (RLHF)

These phases continue to be the foundation of LLM development1. However, research is ongoing to improve and expand upon these methods.Regarding synthetic data and time-scaled approaches:

1. Synthetic Data: While LLMs can generate synthetic data, this is typically used for fine-tuning other models rather than for training the LLM itself. The primary training data for LLMs still consists of real-world text from various sources4.
2. Time-scaled Approach: Recent research has explored scaling test-time computation to improve LLM performance. This involves optimizing how the model uses computational resources during inference, rather than changing the training process itself39.