Baidu Is Racing Out of a Burning Building. The Question Is Whether It Built the New One First.

In the third quarter of 2024, a car with no driver - and in the newest model, no steering wheel - picked up a passenger in Wuhan, completed the trip, and added one to a count that reached roughly 988,000 rides that quarter, then about 1.1 million the next, up 36% year over year.⁵ In the same building, the part of Baidu that pays for all of this - the search ads that have funded the company for two decades - shrank 7% year over year.³ Both numbers are true at once, and together they tell the whole story. Baidu is not so much pivoting to AI as it is running out of a building that is slowly burning, hoping it finished building the new one first.

The official story is that Baidu boldly pivoted to AI - that it was the first major Chinese tech company to ship a ChatGPT rival, and that it is now leading China into robotaxis and the cloud. The truer story is quieter and more precarious. The old engine is structurally declining, the new engines are real but unproven, and almost every flattering number about the new ones comes from Baidu itself.

The pivot isn't a turn. It's a relay race against a clock.

Start with the thing the bullish framing gets wrong: 'pivoted to AI' implies a recent, clean directional break. It wasn't. Baidu drove into autonomous technology in 2013 and open-sourced its Apollo platform in 2017, building a community of more than 45,000 developers and 210 industry partners over the following years.⁸ The generative-AI surge of 2023 was an acceleration of a decade-old program, not a fresh idea. So the real question is never 'will Baidu pivot?' - it's a relay race. Can the new revenue lines grow fast enough to outrun the decline of the old one before the cash that funds the experiments runs thin? That is a question about timing and arithmetic, not vision.

And the arithmetic is tight. AI Cloud is the strongest leg: Baidu Core's non-online marketing revenue — mainly AI Cloud — grew 10% year over year in Q2 2024; it then accelerated to a 26% year-over-year gain in Q4 2024, reaching RMB 7.1 billion.⁹¹² That is genuinely good. But in the same Q4, Baidu Core's online marketing revenue - still the majority of the company - fell 7%.³ A larger base shrinking at 7% can swamp a smaller base growing at 26% for a long while. The relay handoff hasn't happened yet.

First to ship, first to be forgotten

The chatbot story is the cautionary half of the pivot. Baidu is widely credited as the first major Chinese tech company to launch a ChatGPT-style chatbot, ERNIE Bot - though 'first major' is a category Baidu drew around itself.¹⁰ And being first did not protect it. ERNIE Bot opened to invited users on March 16, 2023, but full public access waited on regulatory approval and didn't arrive until August 31, 2023 - nearly half a year of the most important early window spent behind a velvet rope.¹ The launch event itself reportedly leaned on pre-recorded demos rather than live ones, and Baidu's stock dropped roughly 10% that day.² By early 2025, CNBC reported that ERNIE had been eclipsed by rivals from Alibaba and ByteDance and by startups such as DeepSeek.¹⁰ First out of the gate, lapped before the first lap was done.

The lesson here isn't that Baidu was slow. It's that in generative AI, being first means almost nothing if the product can be matched within months by a better-resourced rival or a cheaper startup. The moat people assumed Baidu had - a decade of AI investment, a head start, a Chinese-language edge - turned out to be a head start on a track everyone could enter at once.

The robotaxi numbers are great. They're also all Baidu's.

Apollo Go is the most photogenic part of the pivot, and the operational momentum is genuine: in Q3 2024 fully driverless rides crossed 70% of total rides nationwide, rising to 80% in October.⁴ Ride volume hit about 988,000 in Q3 and 1.1 million in Q4, the latter up 36%.⁵ Those are the kind of curves that make a story. But read the footnotes. The ride counts, the driverless share, and the per-unit vehicle economics are Baidu's own figures, with no independent third-party audit cited. And the geography is narrow - the showcase Robin Li keeps pointing to is Wuhan, one municipality. A bet that looks national is, today, largely one city deep.

Watch the one number every robotaxi case rests on: the cost of the car. Baidu unveiled the RT6 at Baidu World in July 2022 at RMB 250,000 per unit, roughly 48% below its predecessor.⁶ In May 2024 at Apollo Day, it produced an invoice claiming the cost had fallen further to RMB 204,600.⁷ Then, at a Wuhan trade event in November 2024, Robin Li stated the figure had come back to RMB 250,000 per unit — framed this time as the cost after mass-production scale-up.¹¹ Three numbers in three years, every one self-reported, none independently audited. That isn't necessarily dishonesty. But when the single variable that decides whether robotaxis ever turn a profit keeps moving and is never checked by anyone outside the company, a careful reader treats the unit economics as a claim, not a fact.

But isn't a real cloud business and a real robotaxi fleet exactly what a pivot looks like?

The honest objection is that this read is too harsh. Plenty of pivots are messy and self-reported in the early innings, and Baidu's are at least anchored in things that exist: a cloud business growing 26%, a robotaxi service that genuinely runs without drivers across a whole municipality.³⁹ That's far more than vaporware - it's operating infrastructure with improving metrics, and dismissing it because the audited 20-F hasn't landed would be its own kind of error. Fair. The point is not that the pivot is fake. The point is that it is unfinished and load-bearing on Baidu's own credibility. The cloud growth is real but smaller than the ad decline. The robotaxi growth is real but unaudited and concentrated. The chatbot was real and got passed. None of these is a failure; all of them are unsettled. A pivot is only complete when the new engine can carry the company without the old one - and Baidu isn't there yet. It's mid-air, between trapezes.

Baidu did the hard thing most incumbents never manage: it started building the new building before the old one finished burning. The cloud is rising, the robotaxis are running, and the decade of AI work is finally pointed at products. But it ran out of the gate first on chatbots and got passed anyway, it tells the robotaxi story in one city with numbers only it has seen, and the engine still paying the bills is shrinking. The pivot isn't a triumph yet, and it isn't a flop. It's a relay race where the runner with the baton is fading and the one waiting to take it hasn't proven it can carry it the rest of the way. The whole company is the handoff - and the handoff is still in the air.