I'm Gai Bang Milk Can: Quick review of Baidu's Q3 earnings, AI drives LLM growth

$Baidu(BIDU.US)

A few days ago, I attended Baidu World 2024, which was packed with people. I’ve become so known for telling Baidu-related jokes that whenever Baidu is mentioned, people think of me. After the event, a reporter interviewed me, saying, "Hey Ada, tell us about Baidu World 2024." I replied, "Three things stood out to me. First, I didn’t realize how hard it is to monetize large language models—it’s a long and tough road. The biggest issue with large models right now is actually the lack of commercialization. Second, there are many fine details to tweak in large models, like the 'image generation hallucination' problem, and Baidu has made some progress here. Third, 'Miaoda' is pretty impressive—it lets you create software without writing code." But when the interview was published, the headline turned bizarre: 'Miaoda—Even Ordinary People Can Write Programs!'—implying that those who usually write code aren’t ordinary people.

In any case, the World Conference set the tone for this earnings report, and the core of this report revolves around the development of large models. Let’s first look at the financials:

Total quarterly revenue reached 33.6 billion yuan, with Baidu Core revenue at 26.5 billion yuan, which at least exceeded my own expectations. I had anticipated Baidu Core (all Baidu businesses excluding iQiyi) revenue of 26 billion yuan for Q3 2024. Online marketing revenue came in at 18.8 billion yuan (I expected 18.5 billion yuan, assuming a significant YoY decline due to ad business pressure), down 4% YoY. Non-online marketing revenue was 7.7 billion yuan (I expected 7.5 billion yuan), up 12% YoY. Overall, Baidu’s core business numbers were solid. Net profit attributable to Baidu Core was 7.54 billion yuan, up 17% YoY, with a net profit margin of 28%, which was well above my expectations.

My logic is that Baidu’s main business in 2024—online marketing—is under pressure from macroeconomic consumption trends and will continue to face challenges into Q4. I believe the turning point won’t come until next year. Meanwhile, cloud services and large model businesses are steadily progressing, offsetting the decline in ad revenue and driving growth in non-ad businesses.

Two key points from the earnings report: First, the usage volume of large models—as of November 2024, ERNIE Bot has reached 430 million users, with daily calls to ERNIE large models hitting 1.5 billion. Compared to the 50 million calls disclosed in Q4 last year, usage has grown 30-fold in a year, showing a classic exponential growth curve. This proves there’s no shortage of demand for large models.

I didn’t have a clear sense of what 1.5 billion daily calls meant until someone in the industry explained that, from an engineering perspective, the higher the call volume, the more room there is to optimize inference costs. Traditional models used single-machine inference, while large models employ distributed inference. If underlying computing power is fully utilized, inference costs can be significantly reduced. This is one positive externality of high call volumes.

Beyond call volume, Baidu’s large models have made two other advancements: 1) In Q3, they launched two new models, expanding the product line from the flagship 4.0 and faster Turbo to enhanced lightweight models like Speed Pro and Lite Pro—a so-called marketing mix. 2) Model hallucinations have been greatly reduced. As I mentioned in the interview, RAG has largely eliminated text hallucinations, while iRAG reduces image-generation hallucinations (e.g., if you ask for an image of the Temple of Heaven, the model might generate a four-tiered version when it actually has only three tiers). Greater accuracy means better usability, which is crucial.

The current state of large models is a land grab—everyone’s trying to attract seed users now to lock in future adoption. The bigger problem now is actually too many choices. For typical consumer products, two options cause hesitation, three create frustration, and more than three lead to decision paralysis. Car sales, for example, stick to low, mid, and high trims because offering 38 configurations would make buyers walk away. Too many choices can paralyze decisions. Personally, I currently use two large models: for generating a world-famous painting like "The Great Rooster Day," I’d choose GPT for international needs and ERNIE-ViLG for domestic ones—prioritizing speed, convenience, and accuracy, especially accuracy that aligns with my aesthetic.

The other key point in the earnings report is AI-powered cloud services. Non-online marketing revenue of 7.7 billion yuan, up 12% YoY, was mainly driven by cloud growth. The report hasn’t yet disclosed specific cloud numbers, but I estimate around 5 billion yuan, with at least 15% growth. In the first half of 2024, Baidu AI Cloud ranked first in China’s "MaaS market" and "AI large model solutions market," with shares of 32.4% and 17%, respectively. It’s also been the leader in the AI public cloud market for years. Another telling stat: 60% of central state-owned enterprises use Baidu Cloud.

To sum up, Baidu’s long-term AI-driven strategy, with application-driven development as the main path, is becoming clearer. The 1.5 billion daily calls prove demand is strong, and the steady growth of AI Cloud shows business synergy. As large models become more commercialized, I expect this to gradually reflect in user metrics and financials. As for which model will win over consumers, it’ll likely come down to usability rather than some arbitrary factional divide. It’s like asking whether the bold or delicate school of Song Dynasty poetry preferred beggar’s chicken—the school doesn’t matter; what matters is that the chicken tastes good.

———————

Disclosure: The author holds long positions in $Baidu(BIDU.US) $BIDU-SW(9888.HK) and has always been a Baidu bull.