Is it Copium or just calling out deceptive claims? This article is two weeks old. By that time we already knew DeepSeek made exaggerated claims about their hardware (https://semianalysis.com/2025/01/31/deepseek-debates/). Now their post yesterday about profits also just has made up and misleading claims like “545%” with a bunch of asterisks around the numbers. They know these misleading claims will be repeated without the fine print.
As for nationalism and protectionism - no one wants to see the CCP, an aggressive authoritarian dictatorship, have access to any powerful technology. The world is correct to recognize that risk and do something about it.
Probably both. I'll admit to be layperson in all this, but deepseek is pretty impressive. Even if they have used more compute than they claimed, this part of article you linked didn't age well judging by reviews of 4.5 :D
>Many have compared V3 to GPT-4o and highlight how V3 beats the performance of 4o. That is true but GPT-4o was released in May of 2024. AI moves quickly and May of 2024 is another lifetime ago in algorithmic improvements.
DeepSeek didn't make exaggerated claims about their hardware. They stated the number of GPU hours for their V3 training run and multiplied by the hourly rate for GPU rentals on the open market to arrive at a dollar figure. That's explicitly not a claim about their hardware! It's a notional value to make comparison with other published training runs easier, and it serves just fine for that purpose.
The article you link isn't debunking DeepSeek's claims, but rather a rebuttal to people who, months later, seized on that notional dollar figure to retroactively explain why Nvidia stock tanked after DeepSeek rose in the US App Store download charts. Who knows what institutional investors' actual reasoning was when they used that news as a catalyst to unload their positions.
Are you sure? They stated that 2048 H800s participated in the V3 training run, but it's unlikely they didn't have additional clusters in use for other experiments at the same time, and they don't absolutely have to use H800s for inference, and they don't all have to be colocated in a single cluster like for distributed training, so it would make sense for them to rent additional inference capacity piecemeal on the open market as long as the unit economics work out.
So barring an explicit statement from DeepSeek to that effect, it sounds like the kind of misunderstanding that would result from a game of telephone.
Statistics of DeepSeek's Online Service
All DeepSeek-V3/R1 inference services are served on H800 GPUs with precision consistent with training. Specifically, matrix multiplications and dispatch transmissions adopt the FP8 format aligned with training, while core MLA computations and combine transmissions use the BF16 format, ensuring optimal service performance.
Additionally, due to high service load during the day and low load at night, we implemented a mechanism to deploy inference services across all nodes during peak daytime hours. During low-load nighttime periods, we reduce inference nodes and allocate resources to research and training. Over the past 24 hours (UTC+8 02/27/2025 12:00 PM to 02/28/2025 12:00 PM), the combined peak node occupancy for V3 and R1 inference services reached 278, with an average occupancy of 226.75 nodes (each node contains 8 H800 GPUs). Assuming the leasing cost of one H800 GPU is $2 per hour, the total daily cost amounts to $87,072.
6 comments
[ 3.2 ms ] story [ 17.3 ms ] threadAs for nationalism and protectionism - no one wants to see the CCP, an aggressive authoritarian dictatorship, have access to any powerful technology. The world is correct to recognize that risk and do something about it.
>Many have compared V3 to GPT-4o and highlight how V3 beats the performance of 4o. That is true but GPT-4o was released in May of 2024. AI moves quickly and May of 2024 is another lifetime ago in algorithmic improvements.
The article you link isn't debunking DeepSeek's claims, but rather a rebuttal to people who, months later, seized on that notional dollar figure to retroactively explain why Nvidia stock tanked after DeepSeek rose in the US App Store download charts. Who knows what institutional investors' actual reasoning was when they used that news as a catalyst to unload their positions.
So barring an explicit statement from DeepSeek to that effect, it sounds like the kind of misunderstanding that would result from a game of telephone.
Statistics of DeepSeek's Online Service All DeepSeek-V3/R1 inference services are served on H800 GPUs with precision consistent with training. Specifically, matrix multiplications and dispatch transmissions adopt the FP8 format aligned with training, while core MLA computations and combine transmissions use the BF16 format, ensuring optimal service performance.
Additionally, due to high service load during the day and low load at night, we implemented a mechanism to deploy inference services across all nodes during peak daytime hours. During low-load nighttime periods, we reduce inference nodes and allocate resources to research and training. Over the past 24 hours (UTC+8 02/27/2025 12:00 PM to 02/28/2025 12:00 PM), the combined peak node occupancy for V3 and R1 inference services reached 278, with an average occupancy of 226.75 nodes (each node contains 8 H800 GPUs). Assuming the leasing cost of one H800 GPU is $2 per hour, the total daily cost amounts to $87,072.