$ETH The joke has gone too far, this company in Hangzhou has no
thoughts of taking down Nvidia and OpenAI, because they know they
are a computing power company standing on the shoulders of giants,
after all, this company’s A100 graphics card inventory was built
before the sanctions with 10,000 Nvidia graphics cards.
Innovation is just consolidating 32-bit algorithms into 8-bit algorithms.
After all, 0 to 1 innovation is something others have open-sourced!
The biggest contribution comes from Meta, the former Facebook!
This company in Hangzhou admits that the road ahead is still
extremely long.
But ultimately they have reduced the costs in the real world,
and domestic manufacturers have now started to lower prices,
entering an era of low prices, and the ultimate beneficiaries
are still the consumers.
The following text serves as a unified response; I hope the gentlemen
who have replied to me possess a certain level of reading ability!
According to the Deepseek technical paper, the $6 million does not
include "costs related to previous research and ablation experiments
on architecture, algorithms, and data." This means that if the
laboratory has already spent hundreds of millions of dollars on
preliminary research and can utilize a larger cluster, then it is
possible to train an r1 quality model with an operational cost of
$6 million. Deepseek clearly possesses more than 2048 H800s;
their early paper mentioned a cluster composed of 10,000 A100s.
A similarly smart team could not possibly establish a 2000 GPU
cluster and train r1 from scratch with just $6 million.