DeepSeek shakes Silicon Valley, and its cost-effective training technology has attracted widespread attention in the market.

In the latest published Research Reports, Citigroup Analysts Atif Malik, Asiya Merchant, and others detailed the potential impacts of DeepSeek on various segments of the AI Infrastructure Industry Chain, revealing which segments are likely to benefit and which may face challenges.

Citigroup stated that DeepSeek's R1 model is expected to promote the widespread adoption of AI models, especially in the Consumer and enterprise markets, as the computational costs decrease, the potential ROI of AI will significantly improve.ROI(ROI) will significantly improve.

At the beginning of the report, Citigroup first defined scaling laws and pointed out that the latest view currently holds that scaling laws have three stages: pre-training, post-training, and test time scaling. Among them, DeepSeek is an excellent example of 'test time scaling.'

Pre-Training: The process of training machine learning models on large datasets to generate general features.

Post-Training: A series of techniques to optimize the model through reinforcement learning, human feedback, and other technologies after pre-training.

Test-Time Scaling: During the inference phase, extending the model's thinking process through multi-step reasoning.

For the AI Infrastructure Industry Chain, the innovative technologies brought by DeepSeek have multiple impacts. Citigroup believes that although certain links may face short-term challenges, overall, the construction of AI infrastructure will continue to see strong growth. The prevalence of the inference phase will bring new opportunities for GPUs, ASICs, DCI, and other aspects, while the decrease in training demand may exert some pressure on elements such as Retimers and optical modules.

With the continuous advancement of AI technology, the construction of AI infrastructure will become one of the core driving forces of the Global Technology Industry. DeepSeek's innovations not only promote the proliferation of AI models but also bring new opportunities and challenges to various links in the AI Infrastructure Industry Chain.

The impact on various segments of AI infrastructure

GPU: Neutral.

As the core hardware for AI training and inference, the market demand for GPUs has remained strong. Citigroup believes that despite the potential reduction in demand for large-scale training due to DeepSeek's technology, the computational demand during the inference phase will increase, thus keeping the GPU market overall neutral.

ASIC (Application-Specific Integrated Circuit): Neutral to Positive.

ASICs perform particularly well in the AI inference stage. Citi expects that as the demand for inference increases, the market share of ASICs will gradually expand. Although the reduction in training demand may have some impact on the ASIC market, in the long term, the growth in inference demand will offset this negative impact. ASICs may ultimately be more associated with inference, so the potential reduction in the training stage may be compensated by future growth in inference.

Retimers: Neutral to negative.

Retimers are mainly used for high-speed data transmission, especially in the AI training stage. Citi points out that as AI computing shifts from training to inference, the demand for Retimers may decline due to the lower computational intensity in the inference stage and relatively reduced demand for high-speed data transmission.

Optical Modules (Intra Server/DC): Neutral to negative.

Similar to Retimers, the demand for optical modules is higher in the AI training stage. Citi believes that as the inference phase becomes more prevalent, the market demand for optical modules may be affected, especially in terms of internal connection needs within datacenters.

DCI (Datacenter Interconnect): Positive.

DCI is relatively unaffected by training and inference, as it is not closely related to the specific details of AI models and workloads. Citi believes that the demand for DCI is relatively insulated from changes in the proportion of training and inference, with growth in the inference phase presenting new opportunities for DCI.

Switch: Neutral to positive.

Switches, as core components of Datacenter networks, have a market demand that is closely related to the popularity of AI computing. Citibank noted that although the decrease in training demand may have a short-term impact on the Switch market, the growth in the inference stage will drive higher network bandwidth demand. In the long term, the Switch market will remain neutral to positive.

Connectors: Neutral to positive.

Similar to switches, connectors are considered unrelated to training and inference, and their market demand is closely tied to the overall growth of AI Infrastructure. Citibank believes that as AI computing becomes widespread, the demand for connectors will remain stable, especially driven by distributed computing, and the connector market is expected to welcome new growth points.

Storage: Neutral to positive.

Storage OEMs have not yet been affected by AI spending, and discussions about AI's impact are still far off. Citibank pointed out that while the impact of AI on storage devices has not fully manifested, the accelerated inference stage will drive demand for data management, data movement, and data security, which will boost the storage market.

Server OEM (Original Equipment Manufacturer): Neutral.

Citibank believes that although the demand for AI training Servers may decrease, the demand for inference Servers will increase, and overall, the Server OEM market will remain neutral.

PC/Smart Phone: Neutral to positive.

With the optimization of AI models, the AI computing power on local devices will be improved. Citibank pointed out that this will drive the upgrading of PCs and Smart Phones, especially on devices with higher computing power, where market demand will remain neutral to positive.

Editor/danial