Baidu – China AI Ecosystem

Baidu leads the way.

• Baidu has launched a series of tools that allow developers to create generative AI services based on its ERNIE foundation model in yet another sign that one of the key themes of 2024 is going to be the emerging battle for the AI ecosystem.
• Baidu is holding its Create 2024 developer conference which I suspect will serve as a cutout for all of the other developer conferences from the AI players this year.
• Baidu has disclosed statistics for ERNIE indicating that it is being widely used and added to its offering with a series of tools that Baidu says make it very easy for developers to create generative AI services using its foundation model.
  • First, usage and performance: where Baidu claims 200m users and that ERNIE handles 200m queries per day.
  • Baidu has not said whether these are daily or monthly users but the usage data strongly suggests that this is either a monthly active or a cumulative user number.
  • On average this would mean around 67m DaU each of whom are making around 3 queries each per day which sounds about right to me.
  • This means that Baidu’s user base is about 1/3^rd of the size of ChatGPT probably making it the number 2 in the world in terms of size although almost all of these users are in China.
  • Baidu also disclosed that 10,000 companies are using Baidu Comate its AI coding assistant while 85,000 enterprises are using its cloud service (Qianfan) to train and run their own models.
  • Baidu also made some bold claims about the performance of its systems with a 5.1x improvement in training efficiency and a 105x improvement in its capacity to respond to queries every second.
  • It is also claiming that its cost of inference has been reduced to just 1% of what it was in March 2023 which, if true, is a staggering improvement.
  • These performance statistics come with no frame of reference and so it is impossible to judge how they compare to the latest Western systems using Nvidia’s H100, but I think it is relatively safe to say that it is a lot better than it was last year.
  • I suspect that improvements from here will begin to significantly slow down as Baidu will be able to buy neither the latest chips nor last year’s chips from Nvidia and AMD creating a limit to how much further it can progress.
  • Second, developer tools: where Baidu launched three new tools.
  • These are names AgentBuilder, AppBuilder and ModelBuilder all of which pretty much do as their names suggest.
  • AgentBuilder provides tools to build chatbots based on the ERNIE foundation model and AppBuilder is aimed at companies who want to create generative AI services that they can use for business intelligence functions.
  • ModelBuilder is a tool that makes it easy to fine-tune 77 different models including Baidu’s variants of ERNIE as well as 3^rd parties to create generative AI services for specific use cases.
  • Third, AI trends where Robin Li, founder and controlling shareholder of Baidu, called out three trends that he expects to shape the AI industry going forward.
  • Mr Li said that AI services in the future will be an amalgamation of a number of specialised models working together to solve complex problems and that the focus will shift to smaller models over time.
  • These are trends with which I tend to agree.
  • Mr Li also suggested that chatbots would significantly improve their ability to reason which is something I think remains unlikely given that the empirical evidence available indicates that the machines cannot reason even the simplest of tasks.
• Despite abandoning the battle for the digital ecosystem on smartphones in China some time ago, the AI ecosystem is a battle it should be able to win.
• Although other companies such as ByteDance and SenseTime are better than Baidu when it comes to specific AI tasks such as facial recognition and video categorization, Baidu is by far the leader everywhere else.
• This includes in the area of generative AI where Tencent and Alibaba have made some noise but not done an enormous amount.
• By contrast, Baidu is on the 4^th generation of its foundation model and is launching the tools necessary to build an ecosystem based on its foundation models.
• This is precisely what OpenAI launched in November and what I am expecting Google to launch at I/O next month which will be similar to what Meta launches when it next holds an event aimed at developers.
• This puts Baidu in an excellent position to be the AI ecosystem of China and at the moment, there is nothing other than government intervention to challenge its ascendancy.
• This is where the uncertainty lies as the last few years have seen a brutal crackdown on privately owned technology companies to ensure that they remember who is in charge in China.
• Given the importance of AI, there is a risk that Baidu will receive more of this treatment if it becomes very successful and the dominant player in AI in China.
• This is the risk that one would take when looking at Baidu which for an AI company that looks unchallenged in the world’s 2^nd largest economy is ridiculously cheap.
• Baidu is trading on less than 10x 2024 PER and less than 2.0x 2024 EV / Sales which is a tiny fraction of what is being demanded for companies involved in generative AI outside of China.
• The problem is that no one wants to know and the level of depression in China has reached a point where valuations almost do not matter.
• I am sitting on a position in 3 Chinese state-owned banks which have done nothing in 3 years except churn out a 7% dividend yield which is fine, but dull.
• Switching to Baidu greatly increases the risk because it means switching from state-owned enterprises into the private sector, but the valuation of Baidu would seem to pretty much be pricing in this risk.
• Hence, I think Baidu is worthy of serious consideration as the best way to play generative AI for anyone with a strong stomach.