1. Cost of training is 6million: many new outlets have reported on the following facts but mainstream news kept spreading this misinformation

Fact check: Deepseek did not reveal the full cost

- the only cost of training provided was the final training run not the full research cost

- they did not reveal cost of training R1 only for V3.

2. “They use old chips to build”: Mainstream media (CNN, CNBC, Bloomberg, etc) love to keep saying “they are using an older chip”.

Fact check: H800 is not an older version 

- H800 is not an older version but a China-version of H100 with reduced chip-to-chip data transfer rate (around 300 GBps compared to H100's 600 GBps).

- exact performance impact is not officially confirmed by Nvidia

- performance difference is not even close  to using “an older chip". The tone that the media love to use makes it sounds like they were using A100 to train.

3. Cost of running free AI service is not a few million dollars affair

Fact check: serving the mass number of users that ChatGPT is currently serving is not going to cost $6million only. Nvidia was very clear in its statement.

- Deepseek probably want to prove that model validity and use that to raise more money.

-"Inference requires significant numbers of Nvidia GPUs and high-performance networking," Nvidia said in its statement.

We think that the constant spread of misinformation is causing irrational fear that is blowing this whole thing way out of proportion.