Introduction: Enflame plants the 'spark' in Tencent's AI 'fertile ground'.

The 'Four Little Dragons of Domestic GPUs' are about to gather in the capital market. Recently, Shanghai Enflame Technology Co., Ltd. (hereinafter referred to as 'Enflame Technology') has just passed the deliberation of the SSE Listing Committee and is about to be officially listed. According to the prospectus, Enflame Technology plans to raise 6 billion yuan through the listing, primarily for the fifth-generation AI chip series products and the sixth-generation AI chip series products.

Previously, when Moore Threads, MetaX, and Biren Technology went public, the market's enthusiasm for domestic GPUs had already been vented. Therefore, on the eve of Enflame Technology's listing, the market reaction was somewhat muted.

However, Enflame Technology's listing still holds significant reference value for understanding the domestic GPU market and the development of AI large models. One of the most noteworthy aspects is the cooperation between Enflame Technology and its major shareholder, Tencent.

01 From Major Shareholder to Big Client: Tencent's Growing Importance to Enflame

From the perspective of shareholder structure, Tencent is one of Enflame Technology's major shareholders. The prospectus shows that Tencent Technology and its affiliates hold a combined 20.26% stake in Enflame Technology. ZHAO LIDONG and Zhang Yalin are the co-actual controllers of the company. ZHAO LIDONG, Zhang Yalin, and their concerted parties directly and indirectly control a combined 28.14% of Enflame Technology's shares.

However, the voting rights corresponding to the shares held by Enflame Technology's single largest shareholder are dispersed, with no single shareholder or affiliated parties holding more than 30% of the shares. The company has no controlling shareholder. Enflame Technology also cautions that after this issuance, the proportion of voting rights controllable by the two actual controllers will be further diluted.

Tencent and Enflame Technology began deep business cooperation in 2019, primarily focusing on core areas such as AI chip scenario adaptation, model performance optimization, and supporting software stack improvement.

In 2021, Tencent made its first equity investment in Enflame, leading Enflame Technology's Series C financing that year.

Besides being a major shareholder, Tencent is also Enflame Technology's largest client.

From 2023 to 2025, Enflame Technology's main business revenue increased from 292 million yuan in 2023 to 986 million yuan in 2025, with a compound annual growth rate of 83.76%. Revenue growth was primarily driven by Tencent's orders.

From 2023 to 2025, Enflame Technology's total revenue from direct sales to Tencent Technology (Shenzhen) and corresponding AVAP model sales was approximately 100 million yuan, 273 million yuan, and 830 million yuan, respectively, accounting for 33.34%, 37.77%, and 83.79% of the same period revenue. It is expected that the high proportion of sales to Tencent will continue for some time.

Enflame Technology stated that since beginning business cooperation with Tencent in 2019, the initial collaboration was primarily through the AVAP model. The AVAP model involves Enflame Technology completing assembly and manufacturing through Tencent-designated server manufacturers before delivering the products to Tencent.

Under the AVAP model, Tencent pays the server manufacturers first, who then pay Enflame Technology. In the initial AVAP model, Enflame Technology sold to four other server manufacturers designated by Tencent Technology (Shenzhen).

Starting in 2025, Enflame Technology's cooperation with Tencent shifted to direct sales. All revenue generated in 2025 was from selling third-generation AI acceleration card products, with sales revenue of approximately 768 million yuan and a gross profit margin of 33.95%, comparable to the AVAP model for similar products.

From 2023 to 2025, Enflame Technology's sales to Tencent and its affiliates increased from 100 million yuan to over 800 million yuan, an eightfold growth in three years. The reasons behind this growth are twofold:

On the one hand, it is the future development trend. AI has become a must for the core strategies of major internet companies, and Tencent is no exception. Leveraging AI to empower its extensive business scenarios, such as WeChat, Tencent Meeting, and Yuanbao, to achieve a commercial closed loop , is a necessary path for Tencent's future development.

According to a Bernstein report, Tencent's demand for AI computing power continues to rise, with sustained increases in AI-related capital expenditures in recent years. Its AI-related direct and indirect capital expenditures for 2026-2028 are expected to be $11 billion, $13 billion, and $14 billion, respectively.

On the other hand, it is data security. Given that Tencent's business scenarios are national-level applications, considering data security, it needs to deeply collaborate with domestic AI chip manufacturers in terms of underlying computing power. Tencent's cooperation with domestic chip manufacturers aligns with its interests.

Orders exceeding 800 million yuan annually make Tencent an absolutely major client for Enflame Technology. However, in Tencent's massive AI investments, Enflame Technology may only be a small supplier.

According to Martin Lau, President of Tencent, Tencent's investment in new AI products alone reached 18 billion yuan in 2025, and he plans to double this investment in 2026. The Q1 2026 financial report shows that Tencent's investment in AI-related areas (including new AI products) has already reached 37 billion yuan.

02 Tencent Contributes Over 80% of Revenue, Enflame Says It Will Decline in the Future

With Tencent transitioning from a major shareholder to a big client, Enflame Technology's 'Tencent attribute' has become increasingly evident. Enflame's spark has been planted in Tencent AI's 'fertile soil'.

For Enflame Technology, its cooperation options are limited.

According to CIC data, the market size of AI acceleration cards in China reached 378.39 billion yuan in 2025, with more than half sold to the internet industry. This segment of the market reached 200.55 billion yuan, accounting for 53% of the AI acceleration card market that year, with the majority purchased by three major companies: ByteDance, Alibaba, and Tencent. The market is highly concentrated.

Furthermore, according to Bernstein's 2025 research report, AI capital expenditures by the three leading internet companies, ByteDance, Alibaba, and Tencent, are expected to increase from $45 billion in 2025 to $85 billion in 2028, accounting for more than 50% of domestic AI capital expenditures each year.

Faced with NVIDIA's CUDA ecosystem monopoly in the international market, it is clearly difficult for domestic AI chip manufacturers like Enflame Technology to achieve an ecological breakthrough solely with their own software and hardware systems. Therefore, domestic GPUs must find their own 'divine intervention'.

Looking around, the domestic entities with sustained, substantial demand and sufficient funds are none other than the three major internet giants mentioned above. Leading internet companies like Tencent and ByteDance possess vast developer communities, rich application scenarios, and massive user bases. Collaboration with these major players can effectively drive software adaptation and application deployment of domestic AI chips.

Moreover, for Enflame Technology, cooperation with Tencent is not only about commercialization but also one of the best paths to achieve an ecological breakthrough for domestic cloud-based AI chips.

In its response to inquiries, Enflame Technology repeatedly stated that its existing AI chips are highly compatible with Tencent... Its existing products have been extensively adapted to a large number of AI models at Tencent, supporting multiple business scenarios within Tencent.

But which specific scenarios are they?

After careful examination, 'Alpha' found that Enflame Technology's support for Tencent's business scenarios likely began with its second-generation products and is experienced by all WeChat users: 'WeChat Voice to Text.' Additionally, Enflame Technology's second-generation products support Tencent's traditional AI capabilities in content understanding, Tencent Meeting minutes, and dozens of other business scenarios.

Furthermore, Enflame Technology stated that its third-generation products, which began mass production and shipment in 2024, can better support traditional AI models and also support applications such as AI large models with small parameter counts and AI models for advertising search recommendations, adaptable to hundreds of business scenarios.

In this process, Enflame Technology not only gains billions of yuan in revenue from orders but, more importantly, continuously improves its software ecosystem, including operator libraries and toolchains.

This allows Enflame Technology to significantly reduce migration and deployment costs for other clients when providing computing power support, as they typically do not require complex and time-consuming migration and adaptation work. This also enhances Enflame Technology's ability to explore and deploy for other clients.

Similar to Enflame Technology, which relies on a single major client for 80% of its revenue, Cambricon's revenue in 2024 and 2025 also relied on a single major client for over 80%.

However, Enflame Technology is not without ideas regarding its client structure. It has outlined a three-step client acquisition strategy: 'prioritize single-point breakthroughs, subsequently expand from point to line, and gradually extend from line to surface.'

This means focusing resources on Tencent in the first phase to create scale and demonstration effects within major internet clients. After achieving this, it enters the second phase, applying relevant experience and technical capabilities to other internet and non-internet leading clients, achieving expansion from point to line. In the third phase, it replicates the application scenarios of leading clients from the second phase across various industries, gradually achieving a broad impact.

Enflame Technology stated that it is currently gradually entering the second phase. Non-affiliated leading internet client A has completed preliminary hardware system testing and model matching testing, placed small-volume orders, and is in the cluster gray testing phase. Leading internet client B is negotiating commercial terms and bulk orders with the company. Other potential leading internet clients have all passed hardware system testing and are undergoing model matching testing.

Additionally, among non-internet operators and industry clients, leading operator client D has signed a framework contract with the company, and leading industry client E has begun product delivery.

03 Expected to Turn a Profit as Early as 2026, Relying on AI Acceleration Cards

From 2023 to 2025, Enflame Technology accumulated losses of 4.3 billion yuan and experienced a net outflow of 3.972 billion yuan in operating cash flow. The proportion of overdue accounts receivable surged from 29.56% at the end of 2024 to 82.96% at the end of 2025. The post-period collection rate dropped from 63.47% to 31.37%, and the bad debt provision ratio rose sharply from 6.45% at the beginning of the year to 24.76%, accounting for nearly a quarter.

Furthermore, as of the end of Q1 2026, Enflame Technology's total assets were approximately 11.766 billion yuan, an increase of about 7.244 billion yuan from the end of 2025, representing a 160.17% growth. Total liabilities were approximately 9 billion yuan, an increase of about 7.625 billion yuan from the end of 2025, representing a 554.58% growth.

Enflame Technology expects to achieve consolidated financial statement profitability in 2026 or 2027, stating that the timing of profitability is mainly influenced by its 2026 revenue achievement rate and 2026 gross profit margin level.

However, Enflame Technology's current product revenue structure is relatively singular. Its main products can be divided into AI acceleration cards and modules, intelligent computing systems and clusters, and IP licensing and others. Among the 986 million yuan in revenue in 2025, revenue from AI acceleration cards and modules was approximately 856 million yuan, accounting for a high 86.83%.

This reflects a microcosm of the industry's development to date.

Enflame Technology's intelligent computing systems and clusters have also contributed over 100 million yuan in annual revenue, but this is clearly not on the same scale as AI acceleration cards, representing the current state of domestic AI industry chain development.

Regarding intelligent computing center projects, Enflame Technology claims to 'maintain a cautious attitude' for two reasons: First, Enflame Technology has limited resources and cannot guarantee supply capabilities when facing demand from leading internet clients (AI acceleration cards), requiring priority to meet orders from leading clients.

Another reason is that the domestic intelligent computing center ecosystem is still imperfect, with factors such as the owner's funding arrangements leading to longer payment cycles. Industry insiders told 'Alpha' that the leasing progress of some intelligent computing center projects is not ideal, prolonging the project's capital recovery cycle.

Out of prudence, Enflame Technology has established more cautious business terms for undertaking new intelligent computing center projects. Current revenue from intelligent computing systems and clusters mainly comes from the Chengdu Hi-Tech Electronics Cluster Project, which was awarded in 2022 and completed acceptance in 2025.

In 2025, Enflame Technology's revenue from intelligent computing systems and clusters declined by nearly 98% compared to 2024, a situation similar to that of comparable company Cambricon (according to its annual report, Cambricon's main revenue in 2023 came from intelligent computing systems and clusters, with no revenue from this business in 2024 and 2025).

From an industry perspective, Enflame Technology's choices still offer significant market space, related to the technical route it has chosen since its inception.

In the AI chip field, Enflame has chosen the DSA architecture, which is not an easy path. Its competitors include domestic cloud-based AI chip manufacturers such as Huawei HiSilicon, Cambricon, and Enflame Technology itself, all representing the DSA architecture. Foreign competitors include major internet companies like Google, Amazon, META, and Microsoft, which also choose the DSA architecture to mass production self-developed cloud-based AI chip products.

Besides this route, Enflame Technology also faces competition from another route: GPGPU architecture manufacturers represented by Moore Threads, MetaX, Iluvatar CoreX, and Biren Technology.

Why did Enflame Technology choose the DSA route?

As mentioned earlier, more than half of domestic AI acceleration cards are sold to the internet industry, with ByteDance, Alibaba, and Tencent purchasing the majority. Therefore, entering these leading internet companies is crucial for the performance growth of any domestic AI chip manufacturer.

So, what type of AI chips do these major companies need the most? The answer is DSA architecture (non-GPGPU architecture).

With the emergence of AI conversational applications like Yuanbao, Doubao, and Qianwen, as well as new AI agent applications, there is a massive demand for inference computing power. The cost-effectiveness of inference computing power, i.e., the cost per Token, has become a focus for major companies.

This requires AI chips to be optimized in terms of power consumption, cost, and real-time performance.

DSA architecture chips can be separately accelerated and optimized for target scenarios and core algorithms. By 'hardening' core computing, communication, and storage characteristics into the chip and performing software-hardware co-optimization, they can better meet the practical application needs in the AI field and provide extreme acceleration ratios, achieving superior energy efficiency and efficiency compared to traditional GPGPU architectures in corresponding scenarios.

Therefore, DSA architecture acceleration cards have better adaptability for specific scenarios or models (such as inference scenarios) compared to GPGPU architecture acceleration cards, offering advantages in cost and cost-effectiveness.

Ultimately, the significant increase in procurement of inference acceleration cards by major internet companies has driven up demand for DSA architecture acceleration cards. Enflame Technology's choice of the DSA architecture is driven by both commercialization needs and AI development trends.

In 2025, Enflame Technology's AI acceleration cards have achieved large-scale shipments to Tencent, primarily used in inference scenarios. The products support multiple AI application scenarios and meet the requirements for large-scale commercial use in terms of performance, stability, and cost.",