NVIDIA’s reason why I have a bloody sea (very easily explained)

Technology
SEMLCONDUCTORS
AVGO
-18.36
Nvda
-17.22
AMD
Txn
Mu
-7.34
0.68
-13.9
ADI
Nxpi
Qcom
1.45
Mr. B;
1U va
-1.6
INTC
Nchif
2.96
SUFTWARE INFRASTRUCTURE
Orcl
-14.37
MSFT
ADBE
Pltr
-2,52
0.06
-7.01
PANWIISNPSIFTNT |
~ 7.22
0.3
CRWD
-1.72
Sorare

Dropbox
Vice President Morgan Brown
1 First describe the background. Currently the top one AL model
The cost of training is very expensive.
Openal
Companies like Anthropic
dollar
A large -scale that requires thousands of $ 40,000 GPUs
Eater center operated to operate a factory
The same is true of the whole situation:
2 By the way, DEEPSEEEK appeared L and said:
‘If you are a Kyu -guk, this will make this for $ 5 million.
There is a ‘decision?’
And only in words
It’s not actually Hannet.
Their models are in the GPT-4 and Claude framework.
And respond: AL industry is shocked.
How did it be possible?
They thought everything from the beginning.
All the traditional A -fields, up to 32 digits of all numbers
Highly
Same:
DeepSeek said, “How do you record only 8 digits?
I approached it! “
5%
Reduced:
And their “far away
The system is also noteworthy
As the general A -non -elementary school student reads “the
cat
SAT
Read.
On the other hand, DeepSeek is a whole sentence
Read it at once. result
2 times-
Fast
90% accuracy is proud:
Dozens of
When dealing with a dog’s word, this efficiency is very important
do
But the really quirky thing
Expert System “We have built all
Will:
Instead of making a huge poem all (eg:
A person is a doctor lawyer; As if all the engineers play the role of an engineer)
DeepSeek is necessary
To call only experts only
Designed.
61 Existing models
. 8 trillion parameters are always activated
It must be:
DeepSeek is 37 billion of 670 billion parameters
Sanctuary:
As if it is operating a muscular team, only calling only the required experts
Same:
The result is amazing:
Training cost: $ 100 million
$ 5 million
Number of GPUs required: 100,000 units
2,000 units
API cost: 959 savings
Data Center Hardware Instead of the Gay Network GPU,
8/ “But” someone says
there is. “Obviously the disadvantage
Catge! “
The amazing thing is that everything is a right source:
Anyone can do their work. The code is open
And the technical papers explain all the processes:
It’s not magic, it’s simply a very clever engineer:
9/ Why is it important?
Because of this, only large technology companies
Well “Eyes
John’s model is broken.
Now, billions of dollars of data centers
Not required
Good GPU
If you are Taiwan, it is I:
10/1 NVIDIA is a fearful story:
Their business model is 90% margin of super high -cost GPU frame.
It is based on selling.
However, everyone turns into a general gay network GPU
Ahead
Plane
The problem is clear.
11/
And the important point is that DEEPSEEKOL for less than 200 people for two days
It’s a team
META’s team, on the other hand, if it is more than the overall training budget for DeepSeek
Silver is getting paid, but their model is De
Not as much as EPSEEK:
12/ Inlon’s typical destructive student:
Existing companies are optimizing existing processes.
Focal
On the other hand, destructive school companies are a fundamental approach.
I think again
DeepSeek
More than many hardware
What if you approach it well? ”There is no water:
131 The impact is in a hurry:
AL development is more
Accessory
Competition increases rapidly
The “entry barrier” of large technology companies is like a small luck
Hardware requirements (and cost) drop
141 Of course, large companies like Openai and Anthropic
It will not be said:
Are they already implemented this school?
However, the camp of efficiency is now out of the illness,
‘The desire
Let’s put the GPU frame “
I am
15/ Last Thought:
The moment is likely to be remembered as an inflection point later.
:
It’s as if the PC is the mainframe
Make it important or plow
As if the computing has changed everything.
Sinon more accessible and much more:
Change is the current players
The rice paddies to affect
It’s just a problem.
Rocky
Enemy 9

Pryingfan
Follow
@Prying_fan
NVIDIA is more serious than I thought …
China uses 2000 NVIDIA low -performance restaurants
Making an aldush chic bowl of high -performance house
=>
Use of low -priced saliva in the city industry
A framework is possible
=>
NVIDIA is reduced due to the decrease in the necessity of NVIDIA
Occupation in the city industry falls
Translate post
9.40am
1/27/25
34K views
10
t
318
398
193
MOST RELEVANT Replies
Pryingfan
@Prying_fan
5H
Simply loosened with the right sauce right sauce
Whether it’s used by any company’s mid-priced house (H-800)
It is possible to make a frame
01
T 10
18
ILIL 4.2K