Chinese AI company DeepSeek has emerged as a potential challenger to U.S. AI companies, demonstrating a breakthrough model that it claims offers performance comparable to leading products at a fraction of the cost. The company’s mobile app, released in early January, has recently topped the App Store charts in major markets including the US, UK and China, but doubts abound as to whether the claims are true. .
Founded in 2023 by Liang Wenfeng, former chief of AI-driven quantitative hedge fund Highflyer, DeepSeek’s models are open source and include built-in reasoning capabilities that clarify thinking before providing answers. are.
Reaction on Wall Street was mixed. Brokerage firm Jefferies believes that DeepSeek’s efficient approach “punches some of the capital spending euphoria” following recent spending commitments by Meta and Microsoft (each exceeding $60 billion this year). However, Citi questions whether such results could actually be achieved without advanced GPUs.
Goldman Sachs is considering broader implications, suggesting the development could reshape competition between established tech giants and startups by lowering barriers to entry.
Here’s how Wall Street analysts are reacting to Deep Seek, in their own words (emphasis ours).
jeffries
DeepSeek’s impact on AI training punctures some of the capital spending euphoria following last week’s massive commitments by Stargate and Meta. Because DeepSeek offers comparable performance to GPT-4o with a fraction of the computing power, the pressure on AI players to justify ever-increasing capital spending plans could ultimately lead to lower data center revenues and profits. Because of its nature, there are potentially negative implications for builders. growth.
If the smaller model performs well, it could potentially be a positive for smartphones. We are bearish on AI smartphones as AI is not gaining traction with consumers. Running larger models on the phone requires more hardware upgrades (adv pkg + fast DRAM), increasing costs. AAPL’s model is actually based on MoE, but the 3 billion data parameters are still too small to serve consumers well. Therefore, while DeepSeek’s success provides some hope, it does not affect the short-term outlook for AI smartphones.
China is the only market that pursues LLM efficiency due to chip constraints. Trump and Musk likely recognize that further restrictions risk forcing China to accelerate innovation. Therefore, we believe that President Trump is likely to ease AI penetration policies.
city
DeepSeek’s work may be groundbreaking, but the idea that the feat was accomplished without the use of sophisticated GPUs to fine-tune the distillation technique or build the LLM that forms the basis of the final model I have doubts. While US companies’ dominance in cutting-edge AI models could potentially be challenged, US access to more advanced chips could be an advantage in an environment that will inevitably become more restrictive. We estimate. Therefore, we do not expect major AI companies to move away from more advanced GPUs that offer more attractive $/TFLOPs at scale. Recent AI capital investment announcements like Stargate can be seen as a nod to the need for advanced chips.
bernstein
So, 1) I believe that DeepSeek did not “build OpenAI for $5 million.” 2) The model looks great, but we don’t think it’s a miracle. 3) The resulting Twitterverse panic over the weekend seems overblown.
Our own initial reactions do not include panic (far from panic). If we accept that DeepSeek may have reduced the cost of achieving comparable model performance by, say, a factor of 10, then the current model cost trajectory is increasing by about the same amount every year anyway. Also note (the infamous “Law of Scaling…”). Lasts forever. In that context, innovations like this (MoE, distillation, blending precision, etc.) are needed if AI is to continue to advance. And for those considering AI adoption, as semi-analysts we strongly believe in the Jevons paradox (i.e., efficiency gains create a net increase in demand), and we believe that newly liberated We believe that the compute power gained is much more likely to be absorbed by usage and consumption. At this point, we don’t believe our computing needs are close to the limits of AI, so we can weigh the impact on demand growth and long-term spending prospects. It also seems unreasonable to think that the innovations being deployed by DeepSeek are completely unknown by a vast number of top AI researchers in numerous other AI labs around the world (frankly). (I don’t know what the large, closed labs are using for development, as they are deploying their own models, but I’m not sure if they themselves have considered similar strategies or , I can’t believe you probably didn’t use it).
morgan stanley
We have not confirmed the veracity of these reports, but if they are accurate and it is indeed possible to develop advanced LLMs at a fraction of the previous investment, then generative AI could ultimately may be run on increasingly smaller computers (downsizing from supercomputers to workstations). , office computers, and finally personal computers) and the SPE industry are likely to benefit from the accompanying increase in demand for related products (chips and SPEs) as the demand for generative AI expands.
goldman sachs
With the latest developments, we also believe that 1) the relationship between well-capitalized Internet giants and start-ups is increasing, especially given the recent lower barriers to entry for new models developed at a fraction of the cost of existing models; We acknowledge that there may be potential competition between. 2) From training to further inference. The focus is on post-training (including inference and reinforcement capabilities), which requires significantly fewer computational resources than pre-training. 3) Potential for further global expansion of Chinese companies given their performance and cost/price competitiveness.
We expect competition for AI applications/AI agents to continue in China, especially among To-C applications. Chinese companies continue to pioneer mobile applications in the Internet era, for example with Tencent’s creation of the Weixin (WeChat) super service. App. Among To-C applications, ByteDance has led the way with 32 AI applications released in the past year. Among them, Doubao has been the most popular AI chatbot in China to date, boasting the highest MAU (approximately 70 million), and was recently upgraded to the Doubao 1.5 Pro model. We believe that increasing revenue streams (subscriptions, advertising) and an eventual/sustainable path to app/agent-to-app/agent monetization/positive unit economics are key.
For the infrastructure layer, investors will focus on whether a significant increase in cost/model computing efficiency will result in a short-term mismatch between AI capital spending and market expectations for computing demand. I’m concentrating. For Chinese cloud/data center companies, the focus in 2025 will be on chip availability and increasing revenue contribution from AI-driven cloud revenue growth and CSP (cloud service provider) expansion beyond infrastructure/GPU rental. I keep thinking to focus on ability. , how AI workloads and AI-related services can contribute to future growth and profitability. We remain positive on long-term growth in AI computing demand, as further declines in compute/training/inference costs are likely to accelerate AI adoption. See also Theme #5 of the Key Themes Report for a base/bearish scenario for BBAT’s capex estimates depending on chip availability. BBAT’s total capex growth is expected to continue into 2025E in the base case (GSe: +38% YoY). This is a slightly slower pace compared to 2024 (GSe: +61% YoY), which is strong due to continued investment in AI infrastructure.
JP Morgan
In particular, much has been made of the DeepSeek research paper and the efficiency of its models. It’s unclear how much DeepSeek leverages High-Flyer’s roughly 50,000 hopper GPUs (the same size as the cluster that OpenAI is believed to be training GPT-5 on), but it’s likely that the cost (inference costs ) is considered to have been significantly reduced. For example, the V2 model is claimed to be 1/7th the GPT-4 Turbo). Their disruptive (though not new) argument, which started making headlines for AI in the US this week, is that “more investment does not lead to more innovation.” Liang: “I don’t see any new approaches at the moment, but big companies don’t have a clear advantage. They have existing customers, but their cash flow business is also a burden, so ”And when asked about the fact that GPT5 hasn’t been released yet, he said, “OpenAI is not God. It’s not always at the forefront.”
UBS
Through 2024, the first year of large-scale AI training workloads in China, over 80-90% of IDC demand was driven by AI training and concentrated among 1-2 hyperscaler customers . This has led to the demand for large-scale hyperscale IDCs in relatively remote locations. Power-intensive AI training is more sensitive to utility costs than user delays).
If the cost of AI training and inference is significantly lower, we expect more end users, especially retail customers, to leverage AI to improve their business or develop new use cases. This demand for IDCs means more emphasis on location (as user latency is more important than utility costs) and therefore for IDC operators with deep resources in tier 1 and satellite cities. , greater pricing power. On the other hand, a more diversified customer portfolio may result in greater pricing power.
We will update the article as more analysts respond.