|
EC SD | The AI model is legit, in that you can ask it questions and it produces decent responses, somewhat similar to ChatGPT. How did that model get developed? No one knows for sure it seems...
Claims are that they used 50,000 NVDA Hopper GPUs. These are the same Hopper chips as the US AI companies are using, just one year older. The cost of these "claimed" chips is well over one billion dollars, so installed into a datacenter, likely double of that.
Claims go on to say the compute time to train the model was just $5.57 million. This is the number that put the AI stocks into a head spin yesterday. So is that a factual number? How did they come up with that number? It looks like complete fiction to me unless you fudge the numbers a lot. For instance, if you say the $2 billion asset will last 100 years, then cost per day is just $55,000. If you assume it lasts for 5 years (much more reasonable), then the cost per day is over $1 million. There is no way to develop, train, and validate a large AI model in just a few days of compute time. Did they steal it then? Maybe. Perhaps they did develop it from scratch, but it cost a lot, lot more to do that then they are claiming.
My view is this was a massive bear attack on the US market-leading AI stocks. Yes, some big money stands to benefit from a drop in stock prices. And yes, this happens all the time on the stock market. The bears cannot turn the market direction without a massive catalyst (like a recession or a war), and these don't happen frequently. This is why they make up a catalyst a few times per year, drive the news cycle with disinformation, and try their best to make some money from the ensuing panic. | |
|
Deepseek legit?
