DeepSeek – Maths Matters

DeepSeek can’t count.

• In a bizarre statement made on GitHub over the weekend (see here), DeepSeek has laid claim to 545% profit margins which is a mathematical impossibility.
• This leads me to wonder whether this is another attempt to generate headlines to promote the excellence of Chinese AI instead of informing everyone of the true state of affairs.
• On day 6 of its “Open Source Week” DeepSeek released an overview of its V3/R1 inference system where it outlined how the inference of its own hosted services work with a few details of the economics.
• In the blog post, it outlines how its inference service works and then goes on to estimate what the revenue and cost would have been for R1 during a 24-hour period from 12 pm 27^th February 2025 to 12 pm on 28^th February 2025.
• During that period, it processed a total of 608bn input tokens of which 342bn hit the on-disk cache (different price) and generated a total of 168bn output tokens.
• The price for R1 tokens is $0.14/m tokens (hit cache), $0.55/m (miss cache) and $2.19/m output tokens.
• To support all of this inference, DeepSeek used 43,538 H800 GPU hours where it used a price of ($2.00/hour) for an H800 to calculate its total compute cost of $87,072.
• This means that gross profit was $475,028 representing a gross margin of 85% not 545%.
• 545% is a mathematical impossibility because if a business has no costs at all, then the revenue is all profit meaning that margins are 100%.
• Margins of 545% would mean that costs are negative and, in this instance, 5.5x larger than revenues, which is impossible.
• This is the kind of mistake one would expect a model to make not a human.
• This is surprising as DeepSeek is a subsidiary of a hedge fund and if there is something that hedge funds understand well it is profits and how they work.
• This means that these calculations were either done by someone who has no understanding of finance or by someone who wanted to generate as big a figure as possible for marketing purposes.
• Either way, this means that these figures have no credibility whatsoever, and are not worth wasting column inches on.
• RFM Research has concluded that DeepSeek has made some meaningful innovations and has found substantial savings in both training and inference but especially in inference, it depends on the task at hand.
• RFM Research estimates that R1’s inference is between 0x – 11x more efficient than OpenAI depending on the task meaning that far more details need to be given before one can start to make a true estimate of DeepSeek ‘s economics.
• RFM Research and Alavan Independent have also concluded that the main aim of the Chinese state with these revelations is to generate press and public interest in its AI to push forward the notion that it is a leader in AI despite the West’s efforts to contain it.
• The main impact is that questions will now begin to be asked of other AI providers in terms of the efficacy of their models and I have long suspected that the Western companies who have had endless money thrown at them could be far more efficient than they are.
• This means that the cost of AI is likely to fall, which in turn makes it more accessible and for the use cases that it has, usage will rise.
• It also means that if AI is more efficient, more of it can run on edge devices which I have long argued is better for the service provider and offers the user a better service.
• The main beneficiaries are Apple, Samsung, Qualcomm and MediaTek who will benefit from using DeepSeek’s efficiencies (if proven) to deploy more advanced models at the edge where they will consume fewer resources.
• China is clearly a force to be reckoned with in AI where it will put up much more of a fight than it has in semiconductors.
• This looks certain to set the tone of the ideological struggle for 2025.