Baidu's Kunlun says it has begun mass producing second-generation AI chips built on a 7nm process, offering 2-3x the power of the first-gen chips from 2020
BEIJING (Reuters) - Chinese tech giant Baidu said on Wednesday it had begun mass-producing second-generation Kunlun artificial intelligence …
Context & Ripple Effects
Mass production of the second-generation Kunlun chip is the moment Baidu's in-house silicon effort becomes a real product line: five months earlier the unit raised money at a $2B valuation in a round led by CITIC Capital, and this 7nm part — claimed at 2-3x the 2020 first-generation chip — is what that valuation was underwritten on.
The move matters because it predates the supply shock that later defined the unit: by 2023 Baidu was ordering 1,600 Huawei Ascend 910B chips to shift away from Nvidia, and by 2026 Kunlunxin was planning a dual Shanghai-Hong Kong IPO at a $14.69B+ valuation with Baidu holding 58%. This article is where that arc starts.
First-order effects
- Baidu gains a domestic, mass-produced AI accelerator for its own search and cloud workloads at a moment when its core business was already under pressure — Q2 2021 revenue of ~$4.62B marked the start of a five-quarter decline, making cost control on compute directly material.
- The Kunlun unit converts from a research project into a revenue-capable chip business, giving substance to the $2B valuation it raised at earlier that year.
Second-order effects
- Domestic in-house silicon reduces Baidu's exposure to Nvidia's China export posture — a dependency that later showed up concretely in the 2023 Ascend order, and which Nvidia answered with China-specific parts like the Blackwell-based B30A.
- A chip unit with its own valuation creates a spin-off path: the 2021 raise and this production milestone are the first steps toward Kunlunxin's eventual dual IPO, separating chip economics from Baidu's ad-and-cloud P&L.
Third-order effects
- If the pattern holds, China's AI compute stack stratifies into domestic in-house silicon (Baidu Kunlun, Huawei Ascend) and constrained foreign supply, with each major platform owning its own accelerator line rather than renting all its compute from one US vendor.
- Chip units becoming separately listed companies — Kunlunxin at $14.69B+ versus $2B in 2021 — points to AI infrastructure being valued as a standalone asset class in Chinese markets, independent of the parent platform's advertising business.
The trend: Chinese tech giants are building in-house AI silicon as a hedge against Nvidia access, and spinning their chip units out as separately valued businesses along the way.