Baidu is pursuing one of the most ambitious vertical integration strategies in the global AI industry, building its own chips, training its own frontier models on those chips, and deploying the resulting capabilities across its cloud infrastructure, all while simultaneously gaining selective access to restricted Nvidia hardware. The company’s approach, laid out in detail at its Create 2026 conference, represents a calculated hedge against the uncertainty of US export policy and a long-term bet that China’s domestic semiconductor ecosystem will eventually close the gap with Western alternatives.
The timing is significant. After years of strict US export controls that locked Chinese companies out of the most advanced chips, the door has cracked open slightly. Nvidia H200 GPUs have been approved for export to roughly ten Chinese firms between January and May 2026, with Baidu, ByteDance, and Alibaba among the recipients. Rather than treating this as a reason to slow its domestic chip investment, Baidu is treating the current access window as potentially temporary, and pressing ahead with Kunlunxin regardless.
The Kunlunxin Chip Roadmap
Baidu’s Kunlunxin division launched its M100 chip in early 2026, with the more powerful M300 scheduled for release in early 2027. Both chips are designed to handle both training and inference workloads, giving Baidu a domestic alternative to Nvidia’s H100 and H200 for the full range of AI compute tasks.
The financial trajectory of the chip business is striking. Analysts project that Baidu’s chip sales could increase sixfold to approximately RMB 8 billion ($1.1 billion) by the end of 2026. Macquarie has pegged the valuation of the Kunlunxin unit at approximately $28 billion, and Baidu has floated plans for a potential separate listing of the business on both the Hong Kong and Shanghai stock exchanges. A successful IPO at that valuation would make Kunlunxin one of the most valuable AI chip companies outside of Nvidia and AMD, and would provide Baidu with substantial capital to accelerate development of next-generation silicon.
ByteDance Circles the Chip Unit
The strategic value of Kunlunxin has not gone unnoticed by Baidu’s peers. ByteDance, the parent company of TikTok and Douyin, is in active talks to use Kunlunxin chips for its inference workloads, according to multiple reports. ByteDance is simultaneously exploring chips from Shanghai-based Iluvatar CoreX, suggesting it is diversifying its domestic chip sourcing rather than committing to a single supplier. Tencent is already a Kunlunxin customer.
The interest from ByteDance is significant for several reasons. ByteDance operates one of the world’s largest AI inference infrastructures, powering the recommendation algorithms that drive engagement across its platforms. Securing ByteDance as a major customer would validate Kunlunxin’s inference capabilities at scale and provide a substantial revenue base ahead of the anticipated IPO.
ERNIE 5.1 and the Efficiency Imperative
Alongside its chip ambitions, Baidu released ERNIE 5.1, its latest frontier language model, in May 2026. The model achieves a 94% reduction in pre-training costs compared to its predecessor and runs with one-third the number of parameters, while maintaining competitive performance. Critically, ERNIE 5.1 was trained on a Kunlunxin cluster with a 97% training utilization rate, a figure that demonstrates the chip’s readiness for large-scale model training, not just inference.
The cost reduction figure may matter more than the chip roadmap in the long run. If frontier-quality AI can be produced at a fraction of the previous cost on domestic silicon, the economic case for continued dependence on Nvidia hardware weakens considerably, a point that echoes the efficiency-first philosophy that made DeepSeek a global phenomenon earlier this year. For Baidu, the combination of a competitive model trained on domestic chips at a fraction of the cost of Western alternatives is a powerful commercial and strategic argument.
Full-Stack Integration as a Competitive Moat
Baidu’s full-stack approach, controlling chips, models, and cloud infrastructure, is rare even among the world’s largest technology companies. It mirrors the strategy that has made Nvidia so dominant: by owning the hardware and the software stack, a company can optimize across layers in ways that pure-play model developers or cloud providers cannot. For Baidu, this integration also provides a degree of insulation from geopolitical disruption that companies dependent on a single layer of the stack do not enjoy.
The wildcard remains US export policy. The approval of H200 sales to Chinese firms represents a loosening from the strictest periods of the chip embargo, but these decisions are inherently political and subject to reversal. Baidu’s continued investment in Kunlunxin suggests the company views the current access window as a supplement to, rather than a substitute for, its domestic chip strategy, a prudent posture given the volatility of US-China technology relations.