Tencent placed orders for Baidu's internally-designed AI accelerator chips in significant volume.
A significant procurement deal between Tencent and Baidu's chip subsidiary is reshaping the foundational logic that has governed China's internet sector for two decades. According to The Information, Tencent Holdings (00700.HK) has become a major customer for Kunlun Core, Baidu Group's AI chip subsidiary (09888.HK). Simultaneously, Kunlun Core is planning a Hong Kong IPO spin-off targeting a valuation of approximately $50 billion—a figure that already dwarfs Baidu's current market capitalization of roughly $36 billion. Within the same timeframe, Alibaba Group's chip company Pingtouge is also pursuing an independent listing. As China's largest internet platforms begin procuring each other's most critical infrastructure, an era of closed monopolies is drawing to a close, giving way to a new industrial landscape centered on open specialization.
The spin-off news triggered a sharp market rebound. The Hang Seng Tech Index, which had been declining for nearly nine consecutive months, rebounded on June 29, closing up 3.68% at 4,412 points. Baidu Group, Alibaba Group, and Tencent Holdings rose 7.14%, 6.15%, and 3.98%, respectively. However, market participants cautioned that this represents a technical rebound, with medium-to-long-term performance still dependent on fundamentals, as concerns persist over the impact of AI spending on cash flow.
Kunlun Core exemplifies how Chinese tech companies are reassessing the value of AI infrastructure. Historically, the logic behind proprietary chip development was straightforward: cost savings. Baidu developed chips to train search algorithms and large models more cheaply than purchasing Nvidia GPUs. Alibaba built chips to support its massive cloud computing infrastructure and reduce hardware costs. At that stage, chip divisions functioned purely as cost centers—spending money without generating revenue.
The explosion of AI agent applications has fundamentally altered this calculus. The critical cost burden now lies not in large model research and development, but in high-frequency inference. Every AI response, every agent-executed task, and every line of code generated consumes tokens at scale. The foundation rests on GPUs, inference chips, networks, and data centers. Once user volumes cross a critical threshold, API calls transform into tangible revenue streams. Hardware R&D units previously tucked away inside corporate giants as unremarkable departments suddenly possess an independent and compelling business model: chips themselves are an immensely profitable business.
Kunlun Core's P800 has completed scaled validation, delivering multiple 10,000-card clusters since 2025 and completing the training of Baidu's ERNIE 5.1 on a fully domestic cluster. Its customer list has rapidly expanded from Baidu's internal use to include China Mobile, Geely, China Southern Power Grid, China Merchants Bank, and—most dramatically—Tencent. When chips transition from a cost center to a profit-generating business, spinning off and going independent becomes an unavoidable strategic imperative. Kunlun Core officially initiated STAR Market listing guidance on May 7, 2026, with China International Capital Corporation serving as the guidance institution. Earlier in 2026, Kunlun Core had already confidentially submitted a listing application to the Hong Kong Stock Exchange, advancing an "A+H" dual listing simultaneously.
Tencent's procurement of Kunlun Core chips is the most symbolic detail in this industrial transformation. For two decades, China's major internet firms maintained an almost complete separation at the infrastructure level. Alibaba's cloud services were never sold to Tencent, and Tencent's technology could never run on Baidu's infrastructure. Each company repeatedly reinvented its own wheel in the most expensive way possible.
Tencent's decision to purchase Kunlun Core chips marks the first time China's AI industry has begun forming a genuinely mature division of labor. AI-era infrastructure is extraordinarily expensive. Chip development cycles are long, capital investment is heavy, and talent barriers are high. If large firms develop chips solely for internal use, economies of scale can never be achieved, and the cost per chip cannot be amortized. This resembles the relationship between Apple and Samsung in the smartphone industry—the two compete fiercely in the end-user market, yet the iPhone's most critical OLED screens remain dependent on Samsung's factories. When Tencent entrusts a portion of its computing power base to Kunlun Core, it signals that domestically produced AI chips have passed the most rigorous real-world tests. Even competitors deem them sufficient and secure.
During its roadshow, Kunlun Core presented an aggressive condition: those wishing to subscribe to IPO shares must first commit to chip purchases, with subscription amounts being three to seven times the share value. "If you want to be a shareholder, be a customer first." This structure ties chip commercialization to capitalization, reinforcing the transformation from cost center to profit center.
From a global perspective, everyone is engaged in precisely the same endeavor. OpenAI and Broadcom jointly unveiled Jalapeño, their first proprietary inference chip, moving from design to tape-out in just nine months and targeting commercial deployment by the end of 2026 for gigawatt-scale data centers. OpenAI needs proprietary chips because call volume from its monthly active users is so immense that every interaction incurs real costs. Even a 20% improvement in performance-per-watt could save billions of dollars annually. More importantly, the company understands that its destiny cannot be tethered to a single supplier like Nvidia.
Other global leaders have followed suit: Google's TPU has iterated to its eighth generation, representing the world's most mature proprietary chip ecosystem; Amazon has Trainium and Graviton; Microsoft has Maia; Meta has MTIA. The world's leading AI players have all extended their reach to the most fundamental hardware layer.
The logic is straightforward: inference cost is the single largest expense line for any AI company. Whoever drives down hardware costs establishes a truly viable business model. When you simultaneously control both the large model architecture and the chip architecture, the software-hardware co-optimization achievable creates a moat that external, general-purpose GPUs can never match. Kunlun Core was established in 2011, yet its IPO acceleration in 2026 reflects not Baidu's strategy, but the capital market's shifting focus on AI infrastructure as a fundamental, independent, and exceptionally profitable business segment.