Tech giants have their own chip-related woes, caught in a love-hate tangle with NVIDIA

Not only cloud vendors need to reduce their dependence on NVIDIA, but even large model vendors are also developing self-designed chips to lower reasoning costs.

Tencent is not only the largest external shareholder but also the number one customer, which is the reality revealed in the recent IPO filing of Enflame Technology.

As a shareholder, Tencent holds a total of 20.26% of Enflame Technology's shares, while as a customer, Tencent contributes more than 80% of Enflame Technology's revenue. Such a high customer concentration is second to none among the "Four Little Dragons of Domestic GPUs", and the company is almost "raised" by Tencent.

This flaw has sparked widespread public discussion, and also exposed the characteristics of the early commercialization stage of the domestic GPU industry: there are not many customers placing large-scale orders, and more efforts are still needed to accelerate business diversification.

Estimated based on Enflame Technology's proposed market value of 60 billion yuan, the market value of Tencent's holdings exceeds 10 billion yuan.

However, compared with the return on investment, what is more important is that under the AI wave, the roadmap for large internet companies to develop self-designed chips is gradually emerging, and differentiated development has become a trend: Alibaba's chip development is benchmarked against Google; Tencent's strategic intention is closest to META; while ByteDance's chip development is more like a hybrid of the paths of OpenAI and Amazon.

In the era when AI computing power has become a core production factor, not only cloud vendors need to reduce their dependence on NVIDIA, but even large model vendors are also developing self-designed chips to lower reasoning costs.

OpenAI has launched the Jalapeño chip, which outperforms NVIDIA's Blackwell system in two key indicators: "tokens per user" and "throughput per kilowatt", and it only took 9 months from the initial design to the tape-out; Anthropic has set up a self-developed chip team and recruited core early members of OpenAI's custom chip project; DeepSeek was also recently exposed that its self-developed chip project has been launched for about a year, and its chips are dedicated to reasoning rather than training.

This is also the trend of future AI competition. As model capabilities begin to converge, token prices have become a key competitive point, which in turn makes underlying chips a strategic asset for all major model vendors.

1

Tencent Will Not Build a Token Factory

The deep binding between Tencent and Enflame Technology is a move to support and bind a domestic chip startup with equity plus orders.

Tencent's goal of developing self-designed chips is not to build a "token factory", but driven by actual demands. First of all, it aims to achieve customized cost reduction for its own ultra-large traffic scenarios, and the newly added computing power will be prioritized to support the training of Hunyuan large model and applications such as WorkBuddy. Finally, the remaining computing power resources will be used for leasing.

Giving priority to internal use and then external leasing, from the perspective that the purpose of chip development is to serve its internal ecosystem first, Tencent's strategic intention is closest to META. When META released its first generation of MTIA chips, the goal was also to improve the efficiency of content recommendation models, so as to recommend advertisements to users more accurately, and most of the scenarios were used inside META.

In contrast, Tencent established the Penglai Laboratory focusing on chip R&D in 2020, and has developed three chips, including the "Zixiao" series of AI inference chips, the "Canghai" series of video codec chips, and the "Xuanling" series of smart network interface cards.

The most remarkable one is the "Zixiao" 2.0 series, which is the result of joint development between Tencent and Enflame Technology.

In 2021, "Zixiao" was still serving traditional scenarios such as natural language processing and search recommendation, but in the 2.0 era, it shifted to large model training and inference. Official data released by Tencent during WAIC in July this year shows that the training performance of the latest "Zixiao" is 4 times higher than that of the previous generation, and the inference throughput is increased by 3 times.

In actual scenarios, Tencent's internal businesses have taken the lead in deploying the Zixiao 2.0 chip, achieving a 50% performance improvement in the training of the Hunyuan large model. It is planned that the proportion of newly added self-developed Zixiao chips will reach 30% in 2026, and the deployment scale of self-developed chips will be expanded to the level of 100,000 cards by the end of 2026.

According to the disclosure of Caixin, Enflame Technology has been accessed to Tencent Meeting, Tencent Docs and other businesses to meet their AI inference demands.

Compared with the "Zixiao" chip, the "Canghai" series of chips can better demonstrate the logic required by business scenarios. "Canghai" focuses on video codec, and its purpose is to solve the efficiency problem of bandwidth traffic. The "Canghai V2" chip, which was delivered by Tencent in May this year, has entered the mass production cycle and is planned to be fully put into service in the second half of this year.

In the second quarter of this year, Tencent's revenue exceeded 200 billion yuan for the first time, of which the game and advertising businesses contributed more than half of the revenue. But at the same time, the gross profit of the game business increased by 14%, while the gross profit margin of the advertising business decreased by 1 percentage point, the main reason of which is the rising depreciation and operating costs of AI recommendation infrastructure.

Traffic businesses such as WeChat Channels are the key areas that consume a lot of bandwidth traffic. If self-developed chips can save 1% of the expenditure, it will also mean a huge number. Therefore, the "Canghai" series of chips is one of the key points to make Tencent's WeChat Channels run smoothly with lower game live streaming latency, and it is also the underlying infrastructure for Tencent to maintain its profit margin.

2

Tencent's Investment Enters the Harvest Period

The deep binding between Tencent and Enflame Technology is not an isolated case. Yunbao Intelligent, which is about to be listed on the ChiNext market, also has a deep binding relationship with Tencent.

DPU (Data Processing Unit) is regarded by the industry as the third major mainstream chip after CPU (Central Processing Unit) and GPU (Graphics Processing Unit). Founded in August 2020, Yunbao Intelligent is mainly engaged in DPU chips, and Tencent is its largest shareholder with a shareholding of about 19.78%.

In the whole year of 2025, Tencent, the largest shareholder, contributed 95.22% of Yunbao Intelligent's operating revenue, and Yunbao Intelligent was also "raised" by Tencent.

However, Yunbao Intelligent is still in the loss-making period. Its operating revenue in 2025 was 370 million yuan, with a loss of 1.189 billion yuan. The total R&D expenses in the past three years reached 1.758 billion yuan, which is 4.32 times of the total revenue in the same period. As of the last capital increase in December 2025, the post-investment valuation was 14.273 billion yuan, and the company has not made a profit so far.

Following Enflame Technology, Yunbao Intelligent will be another fruit harvested by Tencent in the chip field. Although its scale is much smaller, it will still complete a key capital replenishment: it plans to raise about 3.035 billion yuan for DPU chip R&D and industrialization.

The structure of Yunbao Intelligent and Enflame Technology is exactly the same: Tencent holds shares without controlling the company, binds the enterprise without merging its statements, nurtures a chip startup with orders, and then takes IPO as the exit to realize the return of its investment.

But the difference is that Tencent participates in the joint product definition of Enflame, and Enflame's products carry the genes of Tencent's load, which determines that it is difficult to sell its products to other customers as they are, while Tencent does not intervene in the product definition level of Yunbao Intelligent.

The functions carried by DPU are network acceleration, storage virtualization, and GPU interconnection scheduling. These functions are almost isomorphic for all cloud vendors. Solutions such as those from Alibaba Cloud, China Mobile, and AWS Nitro all solve the same type of problems, and the products can be highly standardized with significant scale effects. Therefore, Yunbao Intelligent's product form is applicable to any customer, and the 95.22% concentration of revenue is only the market result at the current stage, which does not mean customer dependence in the future.

In the final analysis, how do large internet companies divide their work with chip companies?

Yang Zhenyuan, Vice President of ByteDance, once put forward a distinction: for technologies that require high demand grasp, the upper-layer application companies can achieve higher integration efficiency; for technologies with stronger commonality and greater scale effect, they are more suitable for third-party chip companies.

From this perspective, Enflame Technology meets Tencent's demand side, while Yunbao Intelligent meets the common demand of the whole industry.

More importantly, Enflame Technology and Yunbao Intelligent represent the AI inference and network DPU product lines behind them, which form a dual-track configuration with Tencent's self-developed chips "Zixiao" and "Xuanling", that is, moving forward on two legs of self-development and parallel investment in startups. Self-developed chips are designed to obtain phased competitive advantages, drive the invested enterprises to expand their scale, and achieve higher product cost performance.

Therefore, with the listing of Enflame Technology, the "Four Little Dragons of Domestic GPUs" will all gather in the capital market. But for AI chip enterprises such as Enflame Technology and Yunbao Intelligent, the real watershed is not the IPO, but whether they can reduce the dependence on a single major customer after listing.

3

The Future After Core Talent Resignation

On the eve of the IPO of the invested enterprise Enflame Technology, some media broke the news that Gao Jianlin, the head of R&D of Tencent Penglai Laboratory, left his job to start his own business.

Gao Jianlin joined Tencent in 2013, responsible for the hardware team. From the initial development based on FPGA to the start of AI chip design in 2018, and the launch of the first self-developed chip and the establishment of Penglai Laboratory in 2020, Gao Jianlin has led every milestone in the development of Tencent's internal chips so far, and directly led the chip architecture, design, verification, back-end and RISC-V teams.

In fact, Gao Jianlin's new company Hengyi Epoch was registered in April 2026, and the industrial and commercial change was completed at the end of July, with Gao Jianlin serving as the legal representative. During the WAIC in July, Tencent released the Zixiao 2.0 chip. As the person in charge of the R&D of this chip, Gao Jianlin did not appear in any public reports.

This does not mean the contraction of Tencent's self-developed chip business, but only a personnel fluctuation in the self-developed chip team of large internet companies.

After all, Tencent plans to expand the deployment scale of self-developed chips to the level of 100,000 cards by the end of 2026. The subsequent Zixiao 3.0 and the 100,000-card target need to be achieved through organization and process construction.

This is the common situation faced by the chip teams of large internet companies. In February this year, Wang Jian, the former head of ByteDance's chip business, no longer took charge of this sector. Shi Yunfeng, the former head of DPU and AI chip teams who used to report to Wang Jian, and Yu Hongbin, the former head of the CPU team, now report directly to Yang Zhenyuan, the head of Volcano Engine business.

Earlier, Wang Jian was the former superior of Lu Shan, the former head of Byte's server chip business. In 2023, Lu Shan left his job to found Lanxin Computing Power, whose business direction is still the RISC-V and server chips he was responsible for during his time at Byte. Three years after starting his business, in March 2026, Lanxin Computing Power completed hundreds of millions of yuan in financing, with mass production orders exceeding 200,000 chips, and its customer list includes Lenovo, China Mobile and Tencent Cloud.

The departure of Lu Shan and Gao Jianlin essentially reflects the same judgment: CPU is the category with the strongest commonality, and it is also the category that is most suitable to be developed by independent companies.

The difference lies in the time point. Lu Shan left when Byte was in the stage of "cutting redundant resources and improving efficiency", and the server chip project was blocked, while Gao Jianlin left when Tencent's self-development investment was at its peak. What is the same is the direction: both of them chose server CPU. Inside large internet companies, this type of chips can only be used for internal business, and only by leaving to start an independent business can they sell products to more customers.

It is dramatic that three years after Lu Shan started his business, he sold chips to Tencent Cloud; the former chip head who left Byte finally took Tencent as his customer. The business that large internet companies have verified but not chosen to do has finally returned to the doorstep of large companies in this way.

This article is from the WeChat official account Blue Hole Business (ID: value_creation), written by Zhao Weiwei, and authorized for release by 36Kr.