Another bad set of interim results was reported by SenseTime with flat YoY revenues and heavy net losses of RMB3.1bn. However, GenAI related and AIDC revenues accounted for over 20% of sales driven by booming enterprise GenAI demand. The race for the throne of China’s GPT-4 and Midjourney is still ongoing and SenseTime is competing with huge resources including capital, data, knowhow and computational capacity. We believe investors should mainly focus on Company’s progress of GenAI base model SenseNova and relevant chatbot, digital human or text to image model suites while traditional AI business should be written off in terms of valuation. We maintain our HOLD rating for now and key stock catalyst will be SenseTime’s GenAI competitiveness among all peers in upcoming beta tests. We lower our TP from HK$2.85 to HK$1.75 based on 7.5x 2024E EV/Sales.
　　Key Factors for Rating
　　1H23 result reviews: Revenue was flat YoY at RMB1.4bn, below BOCIe/consensus by 23%/31%. GPM decreased 21ppts YoY to 45.3% with OPM down 3ppts YoY to -164% due to higher revenue mix related to lower margin hardware and AIDC. Net loss was another massive miss at RMB3.1bn.
　　Only GenAI matters now: SenseTime decides to do its best to become China’s SOTA LLM and text-to-image AI with its “SenseNova” foundation Model Sets and relevant model suites. In 1H23, generative AI related revenue recorded a YoY growth of 670.4%, and its contribution to the Group’s business rapidly increased from 10.4% in 2022 to 20.3%. Management targets to generate over 40% sales from GenAI by 2025. In other words, we believe Company’s current business segments (Business, City, Life and Auto) will be completely reshaped by the ever-growing GenAI.
　　China’s largest AI infrastructure: SenseCore increased the number of GPUs from 27k in 1Q23 to approximately 30k in 2Q23. The computational capacity scaled from 5 ExaFLOPS in 2H22 to 6 ExaFLOPS in 1H23 and 10 ExaFLOPS based on current GPU purchase orders placed. In data collection, SenseTime has a monthly output of around 2tn high-quality tokens and targets to accumulate 10tn by 4Q23 (the training of GPT-4 used 13tn tokens) to support the training of its latest 100bn parameter base models.
　　Key Risks for Rating
　　Potential breakthrough of SenseTime’s large model; operating loss narrows faster than our forecasts; change in access to Nvidia GPU.