CFO Henry He calls Baidu, the Chinese search giant, a 'full stack AI player': in house AI chips, language models, cloud servers, and a commercial robotaxi service under one corporate roof, an arrangement few Western tech platforms have attempted.
Baidu, the Chinese internet company that built its fortune on search in the 2000s, is now attempting an unusually integrated AI business: its own AI chips, its own AI models, its own cloud infrastructure, and its own commercial robotaxi service, all under one corporate roof. On the Odd Lots podcast, Baidu CFO Henry He framed the strategy in a single phrase, calling Baidu a "full-stack AI player" (Odd Lots interview). The question the company's investors, competitors, and regulators will spend the next year answering is whether that integration is a structural advantage or an expensive set of side bets stitched together by a maturing search business.
The chip layer is the most concrete place to test the claim. TechInsights, a firm that reverse-engineers semiconductor designs, has published a floorplan analysis of Baidu's Kunlunxin P800 accelerator, confirming the company is shipping real in-house silicon rather than buying off-the-shelf Nvidia parts (TechInsights floorplan analysis). In an industry where the cost of training and running AI models has become the dominant line item, owning the silicon is the obvious place to compress margin. It is also where Baidu's path diverges from most Western AI labs, which still depend heavily on Nvidia hardware, and from the bigger Chinese cloud companies, which have run their own accelerator programs for years.
The model layer sits on top of the chips. Ernie is Baidu's family of large language models, the company's Chinese-language counterpart to OpenAI's GPT and Anthropic's Claude. He is explicit that the proxy Baidu watches most closely is what he calls "token spend," the volume of text processed by and generated through AI models, because that number translates directly into both the load on Baidu's cloud and the revenue it can book from AI customers.
Cloud is the connective tissue. The same in-house chips running Ernie inference for outside customers are, in effect, the merchant offering, the same basic logic Amazon used to turn internal AWS capacity into a trillion-dollar business.
Robotaxi is the layer where Baidu is most visibly betting that vertical integration pays off. Apollo Go, the company's commercial autonomous ride-hailing service, now operates in more than twenty Chinese cities as of January 2026 (CN EV Post, January 2026). By late February the service was logging roughly 300,000 weekly rides, and was expanding into South Korea (CN EV Post, February 2026). On Yas Island in Abu Dhabi, Apollo Go and K2's AutoGo have launched a fully autonomous service with a phased expansion plan backed by the local government (PR Newswire announcement). A broker commentary from CMC Markets places Baidu alongside Pony.ai and WeRide as the three firms defining China's commercial robotaxi rollout (CMC Markets analysis).
The Wikipedia entry on Apollo Go is useful for the historical spine: the program began as a research effort in the mid-2010s, ran public pilots through Chinese airports and business districts in the early 2020s, and only recently cleared the regulatory and operational hurdles needed to operate without safety operators in many zones (Wikipedia: Apollo Go). That arc matters because it means the robotaxi business is not vaporware. It has paying customers, third-party reported deployment metrics, and a footprint outside mainland China.
What He is implicitly claiming is that the four layers feed each other. Robotaxi fleets generate driving data that sharpens the models. The models get served through Baidu's cloud, which runs on Baidu's chips. The savings either show up as margin on the cloud offering or as a lower cost per ride for Apollo Go. If that loop is real, Baidu has the kind of integrated AI stack that Western search companies, mostly renting OpenAI or Anthropic models on rented Nvidia hardware, do not.
There are reasons to be cautious about the framing. The "full-stack AI" language comes from a CFO explaining the strategy to investors. It is not a third-party verdict. Ride-count and city-count figures originate from industry reporting that closely tracks the company's own announcements rather than from independent regulators. The CFO interview does not give Baidu's current capex split between chips, models, cloud, and cars; sizing how much legacy search cash is actually flowing into each layer would require reading the most recent quarterly results, which are not in this source packet. And the comparison to Pony.ai and WeRide, both of which face Western investor scrutiny and competitive pressure, is a reminder that the robotaxi segment is crowded enough that an integrated incumbent cannot coast.
The honest read of the CFO podcast is closer to a hypothesis than a verdict. Baidu has the rare ability to fund four large AI bets at once because search advertising, the business most analysts had written off a decade ago, still throws off cash. Whether those bets compose into something more than the sum of their parts is the test the next twelve months of operating data, regulatory outcomes, and capex disclosure will give the market.