百度24Q1業績前瞻：我們預計百度24Q1總收入320億元，同比+2.7%； Non-gaap淨利潤55億元，同比-4%。
24Q1E preview: total non-GAAP net profit down 4% yoy; core revenue up 2% yoy
Key metrics:
ahead of Baidu’s 24Q1 results release slated for 16 May, we provide a preview in this report. We are expecting revenue of RMB32bn, up 2.7% yoy; with non-GAAP net profit at RMB5.5bn, down 4% yoy.
百度Core：
我們預計24Q1百度Core收入爲235億元，同比+2%；其中在線營銷收入167億元，同比+1%，我們預計主要受線上廣告投放不及預期影響；非在線營銷收入68億元，同比+6%；愛奇藝收入89億元，同比+6%。我們認爲，24年因宏觀消費恢復不及預期，導致公司廣告業務承壓，但輕量模型有望帶動雲業務發展，公司通過發展AI雲基礎設施、AI企業大模型平臺等細分業務有望推動雲業務增長。
Core revenue:
our numbers place core revenue at RMB23.5bn in 24Q1E, up 2% yoy; and a breakdown shows:
•Online marketing contributed RMB16.7bn (up 1% yoy), likely due to a weaker online advertising impact than we had originally anticipated
•Non-online marketing generated RMB6.8bn (up 6% yoy)
•iQIYI generated RMB8.9bn, up 6% yoy.
We reckon the advertising business faced pressures this year from a slower-than-expected recovery in macro consumption. However, Baidu’s lean business model would have driven cloud business development. The company would have pushed cloud business growth by developing lucrative segments such as AI cloud infrastructure and the AI large language model (LLM) enterprise platform.
毛利&利潤:
我們預計24Q1 Non-gaap運營利潤65億元，其中百度Core爲50億元，運營利潤率21%。我們認爲，百度在AI領域持續投入，圍繞“MoE、小模型、智能體”思路，發展輕量模型，推理成本顯著降低，調用量有望提升，未來中長期會有增長。
Profitability:
we expect Baidu’s total non-GAAP operating profit at RMB6.5bn in 24Q1E, including RMB5bn from core businesses, and total operating margin at 21%.
Outlook:
we reckon Baidu will keep up with investments to drive its three key AI development trends: Mixture of Experts (MOE), small models and agents. We reckon this would continue to slash inference costs and expand query volumes, and help drive medium to long-term growth.
智能雲&文心一言：
我們預計，百度AI雲業務營收達45億元，實現9%的同比增長。Create 2024百度AI開發者大會，文心一言用戶數和API日均調用量突破2億，服務客戶數達8.5萬，利用千帆平臺開發的AI原生應用數超19萬。相較於一年前，文心大模型算法訓練效率提升到了原來的5.1倍，周均訓練有效率達到98.8%，推理的成本降到了原來的1%，客戶調用成本顯著下降。公司總結大模型AI應用開發思路爲：MoE、小模型、智能體，介紹智能體開發工具AgentBuilder，AI原生應用開發工具AppBuilder，以及各種尺寸的模型定製工具ModelBuilder，並在會上正式發佈文心大模型4.0。我們認爲，受益於AI對雲驅動影響和公司文心大模型系列推理成本顯著降低，疊加輕量模型相較開源模型優勢，雲業務額外收入增量重心由模型訓練過渡到模型推理，推理成本顯著降低帶來的增量反哺雲業務，有望帶動雲業務發展。
Ernie and AI cloud: we estimate revenue of RMB4.5bn in 24Q1E, up 9% yoy
Performance:
•Ernie Bot has over 200m users and gets more than 200m API queries per day on average, Baidu stated at its Create 2024 conference for AI developers last month.
•The Qianfan platform has served 85,000 clients and helped them develop over 190,000 AI native applications.
•The training efficiency of Ernie’s LLM algorithm has lifted 5.1x compared with a year ago, while average weekly training efficiency is at 98.8%.
•Inference cost has reduced to 1% of the March 2023 version while client query costs have also fallen sharply.
Developments:
•Baidu summarized its LLM AI application developments into three key trends at the conference: MOE, small models and smart agents.
•It rolled out three toolkits accordingly: AgentBuilder (a development tool to build and deploy AI agents), AppBuilder (AI native applications) and ModelBuilder (a customizable curated mix of models of various sizes).
•And it released Ernie 4.0 LLM as well.
Outlook:
In view of AI boosting cloud business developments and sharp cuts in inference costs at Ernie LLMs, and Baidu’s lean model advantages over open-source models, the driver of incremental cloud revenue growth is transitioning from model training toward model inference. We believe the company will invest additional revenue from sharp inference cost cuts into the cloud business; this could spur cloud business growth.
智能駕駛：
基於超過1億公里的、中國複雜城市道路測試里程數據，百度訓練出來了Apollo視覺感知大模型。蘿蔔快跑在武漢目前已覆蓋3000多平方公里、770萬人口，是全球範圍內最大規模的自動駕駛運營區域，23Q4提供83.9萬次乘車服務，同比+49%。截至24年1月2日，蘿蔔快跑在開放道路的累計單量超過5百萬。於23Q4，無人駕駛訂單佔武漢整體訂單量的比例突破45%（23Q3爲40%）。我們認爲，伴隨着公司道路測試里程數據的進一步豐富，蘿蔔快跑業務覆蓋的區域和人口越來越多，公司的智能駕駛技術預計將會迎來範式突破時刻。
Autonomous driving: Apollo VTA tested more than 100m km of city mileage
The training of Baidu’s Apollo VTA (Vision Takes All) visual perception LLM has gobbled up masses of data in more than 100m km of test mileage on complex urban roads in China. Apollo Go’s operations currently cover more than 3,000 sq km and 7.7m people in Wuhan, the largest autonomous driving operation by area in the world. It fulfilled 839,000 orders in 23Q4 (up 49% yoy), while cumulative orders on open roads have surpassed 5m as of 2 Jan 2024. Driverless orders accounted for more than 45% of Wuhan’s total order volume in 23Q4 (vs 40% in 23Q3).
Outlook:
As the company’s road test mileage data builds up, along with Apollo Go’s expanding coverage area and population, we believe that Baidu’s smart driving technology will bring about a paradigm shift.
