SENSETIME(20.HK):LARGE LANGUAGE MODEL COMPETITION INTENSIFIES
SENSETIME(20.HK):LARGE LANGUAGE MODEL COMPETITION INTENSIFIES
We project SenseTime's generative AI related revenues should exceed 50% mix in 2024 and believe SenseTime's full transition to GenAI will pay off in the long term. However, we also see significantly intensified domestic LLM competition this year with more domestic competitors achieving good rankings in common international benchmark systems where SenseTime lacks enough presence. ByteDance's token price war since May may also weigh on generative AI growth and margin prospect of SenseTime. Based on the operating losses, token price war and huge AI investments required, we believe SenseTime may need new fundraising by 2025. We maintain HOLD and the key stock catalyst will be the outperformance of SenseNova's model intelligence over domestic and overseas peers. We lower our TP from HK$1.75 to HK$1.45 based on 7.5x 2025E EV/Sales.
我們預計，到2024年，商湯的生成式人工智能相關收入組合將超過50％，並相信商湯向GenAI的全面過渡將在長期內獲得回報。但是，我們也看到，今年國內法學碩士競爭明顯加劇，在商湯缺乏足夠影響力的通用國際基準體系中，越來越多的國內競爭對手取得了不錯的排名。字節跳動自5月以來的代幣價格戰也可能打壓商湯的生成式人工智能增長和利潤前景。根據運營虧損、代幣價格戰和所需的巨額人工智能投資，我們認爲到2025年，商湯可能需要新的籌款。我們維持持有，關鍵股票催化劑將是SenseNova的模型情報表現優於國內外同行。根據7.5倍的2025年電動汽車/銷售額，我們將目標從1.75港元下調至1.45港元。
Key Factors for Rating
評級的關鍵因素
Business momentum resumed upon GenAI: Sensetime has restructured its business into three key categories including Generative AI, Traditional AI and Smart Auto, with respective rev. mix of 35%, 54% and 11% in 2023. We are positive on Company's strategy of phasing out Traditional AI involving Smart City projects and fully embracing Generative AI.
GenAI恢復了業務勢頭：商湯已將其業務重組爲三個關鍵類別，包括生成式人工智能、傳統人工智能和智能汽車，2023年分別增長率爲35％、54％和11％。我們對公司逐步淘汰涉及智慧城市項目的傳統人工智能並全面採用生成式人工智能的戰略持積極態度。
Domestic LLM competition intensifies: SenseTime targets itself as China's SOTA LLM with SenseNova but we also see impressive progress made by domestic LLMs including Qwen by Alibaba, Ernie by Baidu, KimiChat, ChatGLM, and Baichuan, who have made remarkable publicity on various international LLM benchmark leaderboards including LMSYS Chatbot Arena, MT-bench, MMLU, GLUE and SuperGLUE. So far, SenseTime showed up in benchmark leaderboard competition much less often besides SuperCLUE, a Chinese language based leaderboard. We think with SenseTime's early and huge investment in pre- trained transformers since 2017, showing up in such public leaderboards will be key for investors to reinstall confidence. From the monetisation perspective, competition also intensifies as Bytedance initiated token price war, slashing price to 99.3% lower than the industry average for business users since May, which should also add near-term pressure to SenseTime's GenAI revenue and margins.
國內法學碩士競爭加劇：商湯將自己定位爲SenseNova的中國SOTA LLM，但我們也看到包括阿里巴巴的Qwen、百度的Ernie、KimiChat、ChatGLM和百川在內的國內LLM取得了令人印象深刻的進展，他們在包括LMSYS聊天機器人競技場、MT-Bench、MMLU、GLUE和SuperGlue在內的各種國際LLM基準排行榜上大放異彩。到目前爲止，除了基於中文的排行榜SuperClue之外，商湯科技在基準排行榜競賽中出現的頻率要低得多。我們認爲，自2017年以來，商湯科技對預訓練的變形金剛進行了早期的巨額投資，因此出現在此類公開排行榜上將是投資者重拾信心的關鍵。從獲利的角度來看，隨着Bytedance發起代幣價格戰，自5月份以來將商業用戶的價格下調至比行業平均水平低99.3％，這也將增加商湯的GenAI收入和利潤率的短期壓力，競爭也將加劇。
Fundraising may be required by 2025: projecting losses over 2024-26 with a net cash balance of RMB4.4bn, SenseTime may need additional funding by 2025 given our concerns about 1) AIDC and AIaaS price war, and 2) a larger LLM training investment may be required to compete with SOTA models both domestically and internationally.
到2025年可能需要籌款：預計2024-26年將出現虧損，淨現金餘額爲44億元人民幣，鑑於我們對1）AIDC和AiaaS價格戰的擔憂，商湯可能需要更多的資金；2）可能需要更大的LLM培訓投資才能在國內和國際上與SOTA模式競爭。
Key Risks for Rating
評級的主要風險
Potential breakthrough of SenseTime's LLM; operating loss narrows faster than our forecasts; competition with domestic LLMs.
商湯法學碩士的潛在突破；營業虧損的縮小速度快於我們的預期；與國內LLM的競爭。
譯文內容由第三人軟體翻譯。