According to the Star Daily on February 6, as the popularity of DeepSeek skyrockets, the trend of local deployment of large models in edge AI follows suit. Some have already made their "first bucket of gold" from this—by searching for "DeepSeek" on Taobao, the second most searched term is "DeepSeek local deployment," with related tutorials priced from 1 yuan to 20 yuan, and in some cases, over 2000 people have already paid for the tutorials.

At the same time, multiple companies related to AIPC have officially announced their connection to DeepSeek:

$Microsoft (MSFT.US)$ They will soon directly integrate the NPU-optimized DeepSeek R1 version into Windows 11 Copilot+ PC, allowing developers to build AI applications that run on the device using AI models;

$Intel (INTC.US)$ It was stated that DeepSeek can currently run on Intel products and can also achieve offline usage on AI PCs.

$LENOVO GROUP (00992.HK)$ On the 5th, the marketing department for China announced that Lenovo's AIPC personal intelligent agent "Xiao Tian" has integrated with DeepSeek, enhancing its AI capabilities in areas such as professional language processing, code generation and programming, mathematical reasoning, and multimodal processing.

From the stock price trend, LENOVO GROUP rose over 4% today, accumulating nearly 23% increase over the past four trading days.

GTJA's latest report indicates that compared to other terminals, PCs possess productivity tool attributes, with users pursuing a better performance experience; they are the preferred terminals for carrying larger local models.

It is worth noting that according to data released by IDC recently, global PC shipments are expected to reach 0.2627 billion units in 2024, up 1% year-on-year, the last time for positive growth can be traced back to 14.8% in 2021. Although the growth is small, this trend shows signs of recovery in the PC market.

While AIPC hardware products have been launched for some time, the previous limitations of pre-trained large model capabilities and restrictions based on overseas vendor API calls have resulted in insufficient motivation for users to upgrade. Recently, the hardware compatibility costs for Deepseek's open-source models are lower, and their inference performance is excellent. The distilled models can be deployed locally on PCs through AnythingLLM and Ollama, not only protecting data privacy but also allowing for customized optimization according to demand. The parameters of the distilled models cover 1.5B/7B/8B/14B/32B/70B, and based on evaluation results, models with parameters of 32B and above show significantly better performance.

Canalys estimates that AIPC shipments are expected to reach 0.1 billion units by 2025, accounting for 40%; by 2028, shipments are expected to reach 0.205 billion units, accounting for 70%, with a CAGR of 44% from 2024 to 2028.

GTJA further points out that local deployment demands higher requirements for PC hardware, especially for models with parameters of 32B and above, which require GPU configurations of 24GB or more, as well as larger memory and higher heat dissipation/electromagnetic shielding requirements. With the explosive demand for higher specification local model deployments, AIPC upgrades will become an inevitable trend.

Previously, Futu News compiled a graphic on the AI PC concept; interested mooers can take a look:

Specifically regarding the Industry Chain, analysts at GTJA believe that the bargaining power of Terminal Manufacturers is strengthening, with both the quantity and price of core components rising. Among them, Terminal Manufacturers are evolving from traditional Hardware manufacturers to ecosystem organizers, with AIPC placing higher demands on the overall delivery and iteration of software and hardware from Terminal Manufacturers. In terms of core components, the local deployment of AIPC models further increases the demand for high-capacity, high-speed video memory and high-bandwidth memory DDR5. Additionally, due to upgrades in CPU and GPU, the corresponding IC substrates and PCBs will also see rising quantities and prices. The high computing power of AI PCs leads to higher power consumption and electromagnetic interference, which will further necessitate upgrades in thermal materials, cooling systems, and electromagnetic shielding materials. As for structural components, the application of Lightweight materials such as magnesium alloy and Carbon Fiber is expected to expand further.

Editor/Somer