Artificial Intelligence – DeepSeek, The Honeymoon Spoiler

DeepSeek’s timing undermines its credibility.

• DeepSeek’s latest model is surprising the AI industry, but the timing of the model’s release damages its credibility as, at face value, it looks like Chinese propaganda aimed at spoiling the new US administration’s honeymoon.
• DeepSeek is a Chinese AI company that was spun out of a hedge fund (High-Flyer) that was set up in 2015 to use machine learning to recommend trades for generating index-beating returns.
• In 2019 a subsidiary called High-Flyer AI was established to conduct research into AI algorithms which I presume was initially focused on the transformer architecture and then rapidly added LLMs to its focus.
• In 2023, High-Flyer spun out High-Flyer AI, renaming it DeepSeek which has released a number of LLMs into the open-source community with its latest model attracting some attention.
• On 20^th January 2025, DeepSeek released (to open source) DeepSeek-R1 which it refers to as a “reasoning model” which is scoring extremely well on independent testing.
• This testing can be verified as anyone can download the model and run the tests themselves, meaning that the real questions are how it was created and what its economics really are.
• DeepSeek-R1 is a 671bn parameter Mixture of Experts (MoE) architecture which means lots of mini-models each trained in a discipline stuck together to operate as one.
• This is what I suspect OpenAI is doing with its current models although the company has long since stopped following its namesake and being open about anything.
• DeepSeek R1 scores extremely well against almost all of the benchmarks (see here) and comfortably rubs shoulders with the likes of OpenAI o1, Claude 3.5-Sonnet and GPT4o.
• This comes as a big surprise as China is supposed to be cut off from advanced semiconductors that are needed to train a model of this size and complexity.
• Combining this with the release’s timing gives the Chinese state the appearance of thumbing its nose at the new administration, which the CCP has proved very adept at doing over the last few years.
• Hence the real questions are:
  • First, silicon access where the only publicly available data indicates that DeepSeek has a 10,000 cluster of Nvidia A100 that it purchased in 2001 prior to any sanctions being put in place.
  • Given the size and complexity of DeepSeek-R1, it appears that this array is insufficient to train DeepSeek-R1 with speculation that DeepSeek has managed to access an array of 50,000 Hopper H100s.
  • There are two main ways by which this could have been achieved.
  • Either, DeepSeek acquired the silicon through re-exports from a non-sanctioned country (which the 2024 rule updates seek to close out (see here)) or it has rented the capacity in a non-sanctioned country.
  • China’s national security laws prevent algorithms from leaving China meaning that without some form of special dispensation, DeepSeek would have risked getting in very hot water had it taken this route.
  • DeepSeek appears to have regular dialogue with the Chinese state and its CEO, Liang Wenfeng, has publicly met China’s Premier Li Qiang to discuss the development of AI in China.
  • Hence, while DeepSeek has links to the Chinese state, it is unknown how closely it is working with the CCP and so I think the most likely answer is that it has been able to either acquire the H100s or rent them locally from someone that has.
  • This makes the sanctions put in place by the US Department of Commerce look ineffective which is what I suspect was behind the tightening of the rules announced last week.
  • It also helps China fight the propaganda war although this current round has not worked well as it has yet to register on the horizon of the mainstream media and tech press.
  • Second, economics which I think is by far the most important question.
  • DeepSeek’s earlier models have been offered very cheaply in the local market which forced Baidu, Alibaba, Tencent and ByteDance to cut their prices which will almost certainly helped adoption but pressured economics.
  • DeepSeek-R1 is very likely to follow this precedent and given that it is available outside of China on Hugging Face, it could put pressure on everyone else as well.
  • DeepSeek claims to make small but positive margins on its very low price for its service, but given how much everyone else is burning, I find this quite hard to believe.
  • Hence, I wonder if DeepSeek is being financed indirectly by the Chinese state which if executed in certain ways would allow DeepSeek to show a positive return despite its price leadership.
• The economics question is by far the most important one because the real winners and losers of the AI race will eventually be determined in the free market where uneconomical business models will be resoundingly punished.
• Furthermore, the governments of the world including both China and the USA have crippling debt and spending problems and can not endlessly fund AI companies and projects that burn through billions of dollars every quarter.
• Hence, I continue to believe that at some point economic reality is going to raise its head leading to a correction in both expectations and valuations.
• It is at this point that we will really find out who is floating on a bed of hot money and who has a proper business that will stand the test of time.