The Chinese company of Deepseek artificial intelligence has emerged as a potential challenger for US artificial intelligence companies, demonstrating innovative models that claim to offer services comparable to the main offers to a fraction of the cost. The company’s mobile app, released in early January, recently reached the summit of the App Store charts in the main markets including the United States, the United Kingdom and China, but did not escape doubts about the truthfulness of its statements .
Founded in 2023 by Liang Wenfeng, former head of the High-Flyer Quantum Hedge Fund based on artificial intelligence, the Deepseek models are open source and incorporate a reasoning feature that articulates its thoughts before providing answers.
Wall Street’s reactions have been conflicting. While the brokerage company Jefferies warns that the efficient approach of Deepseek “Mina part of the capex euphoria” following the recent spending commitments of Meta and Microsoft – each greater than 60 billion dollars this year – Citi wonders if These results have actually been achieved without advanced GPUs.
Goldman Sachs sees larger implications, suggesting that development could reshape the competition between the established technological giants and the startups by lowering the barriers to the entrance.
Here’s how Wall Street analysts are reacting to Deepseek, with their own words (our underlining):
Jefferies
The powerful implications of Deepseek for training on artificial intelligence perforated part of the capex euphoria which followed important commitments by Stargate and destination last week. With Deepseek offering performances comparable to GPT-4O for a fraction of the calculation power, there are it potential negative implications for manufacturersSince the pressure on the actors of artificial intelligence to justify ever -growing capex plans could ultimately lead to a lower trajectory for data center revenues and the growth of profits.
If the smaller models can work well, it is potentially positive for the smartphone. We are bearish on smartphones to as artificial intelligence has not gained land among consumers. Another hardware update (ADV PKG+Fast Dram) is required to perform larger models on the phone, which will increase costs. The AAPL model is in fact based on the MOE, but 3 billion data parameters are still too small to make services useful for consumers. So Deepseek’s success offers some hope, but there is no impact on the short -term perspectives of smartphones AI.
China is the The only market that pursues the efficiency of the LLM due to the bond of the chip. Trump/Musk probably recognize that the risk of further restrictions is to force China to innovate more quickly. Therefore, we believe that Trump will allocate the policy of spreading artificial intelligence.
Quotes
Even if the result of Deepseek could be revolutionary, us question the notion that his companies were made without the use of advanced GPUs to develop it and/or build the underlying LLM on which the final model is based through the distillation technique. Although the dominance of US companies on the most advanced artificial intelligence models could be potentially questioned, having said that, we estimate that in an inevitably more restrictive environment, the access of the United States to more advanced chips represents an advantage. Therefore, we do not expect that the main artificial intelligence companies move away from the most advanced GPUs that provide the most interesting $/TFLOP relationships on a large scale. We consider recent Capex ads on artificial intelligence as Stargate as a nod to the need for advanced chips.
Bernstein
In short, we believe that 1) Deepseek I have not “built Openi for 5 million dollars”; 2) The models seem fantastic but we Don’t think they’re miracles; and 3) the resultant Twitterverse panic over the weekend seems exaggerated.
Our initial reaction does not include panic (quite another). If we recognize that Deepseek may have reduced the costs to obtain equivalent performance of the model, let’s say, 10 times, we also notice that the current trajectories of the costs of the model increase by about the same every year (the notorious “laws of scale … “) which cannot continue forever. In this context, we need innovations like this (Moe, distillation, mixed precision, etc.) if we want artificial intelligence to continue to progress. And for those who seek the adoption of artificial intelligence, as semi-analyzes we firmly believe in Jevons’ paradox (i.e. that the efficiency gains generate a clear increase in demand) and we believe that any new unlocked calculation capacity has many more likely to be absorbed due to the use and increase in demand compared to the impact on the long -term expenditure perspectives at this point, since we do not believe that the elaboration needs are close to achieving the limit in artificial intelligence. It also seems exaggerated to think that the innovations implemented by Deepseek are completely unknown to the vast number of high -level artificial intelligence researchers in the other numerous artificial intelligence laboratories of the world (frankly we do not know what the large closed workshops have used to develop and unfold their own models, but we cannot believe that they have not considered or even used similar strategies).
Morgan Stanley
We have not confirmed the truthfulness of these relationships, but if they are accurate and it is possible to develop an advanced LLM for a fraction of the previous investment, We could see the artificial generative intelligence work on ever smaller computers (reduction from supercomputer to workstation, office computers and finally to personal computers) and the SPI sector could benefit from the consequent increase in the demand for related products (chips and sPocations) as the generative IA demand is spread.
Goldman Sachs
With the latest developments we also see 1) potential competition between the internet giants rich in capital and start-upsGiven the reduction of the barriers to the entrance, especially thanks to the recent new models developed to a fraction of the cost of the existing ones; 2) From training to inferenceWith greater emphasis on post-information (including reasoning and reinforcement capabilities) which requires significantly lower computational resources compared to pre-formation; and 3) The potential of further global expansion for Chinese operators, given their performance and competitiveness costs/prices.
We continue to expect that the race for applications/agents to the continuous in China, in particular among the To-C applications, where Chinese companies have been pioneer in mobile applications in the internet era, for example with the creation by Tencent of the Super-Interface Weixin (WeChat). App. Among the to-c applications, Bytedance has been at the forefront by launching 32 applications to the last year. Among these, Doubao has so far been the chatbot to the most popular in China with the highest Mau (about 70 million), which has been recently updated with its Doubao 1.5 Pro model. We believe that the flows of incremental revenues (subscription, advertising) And the final/sustainable path towards monetization/positive unitary economy among applications/agents will be fundamental.
As for the infrastructure level, the attention of investors has focused on the event of short -term misalignment between market expectations in terms of capex on artificial intelligence and computer demand, in the case of significant improvements in the efficiency of the Cyber costs/models. For Chinese cloud/data center operators, we continue to believe that attention to 2025 will focus on the availability of chips and on the ability of the CSP (cloud service providers) to provide a contribution in improvement to the revenue deriving from the growth of revenues of the cloud led by artificial intelligence and beyond the rental of infrastructure/GPU. , how the workloads and related services could contribute to growth and margins in the future. We remain positive on the growth of the long -term computer demand as a further reduction in processing/training/inference costs could encourage greater adoption of the AI. See also theme n. 5 of our report on key themes for our basic/reduction scenarios for BBAT Capex estimates depending on the availability of chips, where we plan that the aggregate growth of the Bbat capex will continue in 2025E in our basic case (GSE: +38% a/a) even if at a slightly more moderate rhythm than a strong 2024 (GSE: +61% a/a), led by the continuous investments in artificial intelligence infrastructures.
JP Morgan
Above all, there is a lot of talk about the research documents of Deepseek and the efficiency of their models. It is not clear to what extent Deepseek is taking advantage of the Hopper GPUs of ~ 50,000 of High-Flyer (of size similar to the cluster on which it is believed that Openi is training GPT-5), but what seems likely is that they are drastically reducing costs (Inference costs for their V2 model, for example, would be 1/7 of those of the GPT-4 Turbo). Their subversive (even if not new) affirmation – which has started to hit the American names of artificial intelligence this week – is that “more investments do not equivalent to more innovation”. Liang: “At the moment I don’t see new approaches, but large companies do not have a clear advantage. Large companies have existing customers, but also their cash flow activities represent their burden, and this makes them vulnerable to interruptions at any time “. And when asked that GPT5 has not yet been released: “Openai is not a god, it will not necessarily be at the forefront.”
UBS
During 2024, the first year in which we attended a massive workload of training on artificial intelligence in China, over 80-90% of the application of IDC was guided by the training on artificial intelligence and concentrated in 1-2 customers Hyperscaler, which has translated into an application of wholesale IDC on a large scale in relatively remote areas (such as training based on artificial intelligence, which consumes a lot of energy, is sensitive to the costs of users rather than the User latency).
If the costs of training and inference of the AI were significantly lower, We would expect a greater number of end users to exploit artificial intelligence to improve one’s business or develop new use casesIn particular, retail customers. This application of IDC implies greater attention to location (since the user’s latency is more important than the cost of services) and therefore a greater power of prices for IDC operators that have abundant resources in level 1 cities and satellite cities . In the meantime, a more diversified customer portfolio would also imply a greater power setting power.
We will update the story as the most analysts will react.
