RE: DeepSeek28 Jan 2025 10:38
Towards the end of the day yesterday, I was reading along these lines (from various senior industry people)
- DeepSeek didn't report on the full training costs, such as how many times previously they attempted and failed.
- DeepSeek builds upon and leverages the $Billions spent on other AI systems, without that expenditure, DS wouldn't exist. So, it could be argued DS is actually the MOST EXPENSIVE AI model yet produced.
- US laws bans the export of the latest technology to China, however, it's reported that DeekSeek dev actually has access to 50,000 H100 GPUs (Nvidia's latest generation) i.e. significant cost.
- Wulf's CFO was on a podcast last night (pre-arranged before DS announcement), he didn't seem that phased by the announcement or that it would affect their HPC plans going forward.
- Whilst it maybe 'OpenSource', people have said the most advanced models are not fully shared so giving more doubts about the claims.
No idea if any of the above is true/false, however, personally, I think it's a great development which will help move AI and the mag 7 forward, I think the market has over reacted and will recover.