Baidu Spent Fifteen Years Building New Engines. It's Still Running on the Old One.

On the morning of March 16, 2023, Baidu's founder stood on a stage in Beijing to unveil China's answer to ChatGPT, a large language model the company had been building since 2019.⁵ The world was watching. And as the demos rolled, investors noticed something: they were pre-recorded. No live questions, no live answers - just a tape of the future, played back. By the time the livestream ended, Baidu's Hong Kong shares had fallen about 10%.⁵ The company had spent more than a decade trying to become something other than a search engine, and on the day it most wanted to prove it, the market reminded it what it still was.

The official story is that Baidu transformed itself - from a search box into a sprawling AI conglomerate spanning streaming video, self-driving cars, cloud computing, and generative AI. The truer story is that Baidu added engines without ever replacing the one it was built on. After fifteen years of expansion, advertising on search still pays the bills, and the new businesses mostly cost money.

Four bets, one bill-payer

Baidu was incorporated in January 2000, grew out of a search engine its founder had been tinkering with since 1996, and reached NASDAQ in 2005.¹² Then it went hunting for adjacencies. In January 2010 it co-founded a streaming video service - first called Qiyi, later iQIYI - alongside Providence Equity Partners, buying out the minority partner in November 2012 to take full control.³ (The common shorthand that Baidu 'acquired iQIYI' gets the sequence backwards; it built the thing first.) Then came Apollo and its robotaxi arm, Apollo Go. Then AI Cloud. Then ERNIE. Each expansion borrowed the same logic: search produces data and a massive audience, and that surplus should fund the next platform. The video was, in the founder's own framing, 'a key strategic vertical.'⁴

It is a textbook adjacency play. The problem is what the numbers show fifteen years in. In the first quarter of 2024, Baidu Core brought in RMB 23.8 billion - and RMB 17.0 billion of it, more than seven dollars in ten, was still online marketing.⁸ The non-marketing slice, the part meant to represent the new Baidu, was RMB 6.8 billion.⁸ The core engine had not been replaced. It had simply been asked to tow more trailers.

Why the new engines never took over

The mechanism is the part worth slowing down on. Adjacency expansion works when the new business eventually grows faster than the old one shrinks - the new line crosses over and becomes the load-bearing one. Baidu's adjacencies never crossed over, for two reasons that compound. First, the base. AI Cloud grew 26% year-over-year in the fourth quarter of 2024 - genuinely strong - but from RMB 7.1 billion against a Baidu Core of RMB 104.7 billion for the full year.⁷ A business one-fifteenth the size growing fast still can't move the total; that is just arithmetic. Second, the drag. Each adjacency is capital-intensive in a way search never was. iQIYI competes for streaming rights and subscribers and saw revenue fall 14% in Q4 2024.⁷ Apollo Go runs a fleet of driverless cars on public roads. ERNIE burns compute. The result is that the very businesses meant to diversify Baidu away from its core also bleed the cash the core throws off - so 'transformation' becomes a treadmill, where running faster just keeps the company in place.

None of this means the adjacencies are vanity. Some are real and impressive. By the third quarter of 2024, fully driverless Apollo Go rides accounted for more than 70% of its total rides nationwide, climbing to 80% in October.⁶ In Wuhan, the service reached fully driverless ride-hailing across effectively the entire municipality.⁹ Baidu AI Cloud was ranked the No. 1 AI cloud provider in China for the fifth consecutive year - a position, notably, it held before the generative-AI gold rush even began.⁶ And the flagship app still drew 704 million monthly active users in September 2024.⁶ These are not the metrics of a failing company. They are the metrics of a company whose new businesses are real but not yet large enough to matter to the income statement.

Read that sentence again. It is the whole thesis in the founder's own words: the old engine holds; the new one grows. What it does not say - because it cannot - is that the new engine has taken over. By late 2024 the entire non-marketing line was RMB 7.7 billion in a quarter, up 12% and driven mostly by cloud.⁶ A healthy number. Not a replacement.

Isn't this just what an AI transition looks like before it pays off?

The honest counter is strong. Every platform shift looks like cost drag before it looks like a new business; you don't get the harvest before you plant. ERNIE went from a tape-recorded demo to general public availability on August 31, 2023, and past 100 million users by December.⁵ AI Cloud is compounding from a leadership position. Apollo Go's unit economics improve every time a safety driver comes out of the car. Maybe 2024 is simply the awkward middle, and the crossover is coming. That is a fair read, and it might be right. But notice what the bull case quietly concedes: it is a case about the future, not the present. The defensible point of view is about today's structure, and today's structure is that a search-advertising business still carries a company that has spent fifteen years and enormous capital trying to need it less. Until one of the adjacencies grows large enough to survive the loss of the ad engine, 'transformation' is not a completed move. It is a narrative the company tells while the cash register at the front keeps ringing the same way it has since 2005.

Baidu has built genuinely hard things - driverless fleets that work, a cloud that leads its market, a language model with a hundred million users.⁵⁶ What it has not built is a second business large enough to carry the company. The search box it incorporated around in January 2000 still pays for everything that has come since. The adjacencies are real; the dependence is realer. And the lesson is the one the market priced in the day the demos turned out to be pre-recorded: you can stand on a stage and show the world your future, but until the new engine can pull the train without the old one, you are still the company you were trying not to be.