China’s AI Chip IPO: Why Tencent’s 84% Bet Matters

The 84% Question

There is a specific kind of tension that only exists in China’s semiconductor sector right now. It is the friction between the desperate need for self-reliance and the brutal reality of commercial viability.

On June 15, the Shanghai Stock Exchange’s listing committee approved the initial public offering (IPO) of Zhenyuan Technology. This is a milestone. It means the so-called “Four Little Dragons” of domestic GPU companies are all now public. Moore Threads and MetaX listed on the STAR Market in late 2025, with Moore Threads briefly hitting a market cap of 440 billion yuan (about $62.8B). Biren Technology went public on the Hong Kong Stock Exchange in January, valuing the company at over 130 billion Hong Kong dollars (about $16.7B).

Zhenyuan is the last one standing in this quartet, and its prospectus reveals a story far more complex than just “another chip maker.”

The headline number is its valuation: roughly 20 billion yuan (about $2.86B) after its final private funding round in late 2024. But the real story isn’t the valuation. It’s the customer concentration. In 2025, nearly 84% of Zhenyuan’s revenue came from a single client: Tencent.

This isn’t just a sales statistic. It is a mirror reflecting the entire state of China’s AI infrastructure. The narrative that domestic chips are ready to replace Nvidia across the board is, at best, premature. The reality is a tightly coupled, high-risk partnership between a tech giant and a chip startup, where the startup’s survival depends on the giant’s internal needs.

The Architecture War: DSA vs. GPGPU

To understand why Zhenyuan’s path is so precarious, you have to look at the technical split in China’s AI chip industry. It’s not a monolith. It’s a battlefield of architectures.

Zhenyuan, along with Huawei’s HiSilicon and Cambricon, follows the DSA (Domain-Specific Architecture) route. This means they design chips optimized for specific tasks rather than trying to be general-purpose graphics processors. The competitors—Moore Threads, MetaX, and Biren—largely pursue the GPGPU (General-Purpose GPU) path, trying to mimic Nvidia’s CUDA ecosystem.

Why does this matter? Because the DSA route is currently the only one making money in the inference market.

According to data from CIC, the Chinese market for AI inference accelerators is projected to reach 808.58 billion yuan (about $115.5B) by 2028, accounting for more than 70% of the total market. Inference is the process of running a trained model to generate answers or predictions. It is less computationally intensive than training, but it requires massive scale and efficiency.

DSA chips excel here. They are built for “fixed models and clear application scenarios” like search, advertising, recommendation engines, and OCR (optical character recognition). As one industry analysis noted, if a domestic chip can prove its cost-performance ratio in these specific buckets, it can pry open procurement doors that are otherwise locked by Nvidia’s dominance.

Zhenyuan’s hardware has reportedly adapted to nearly 1,000 AI models across more than 300 application scenarios, according to its prospectus. They aren’t just selling a chip; they are selling a software bridge called “TopsRider”, designed to lower the migration cost for customers moving away from Nvidia’s CUDA ecosystem.

The Tencent Symbiosis

Let’s look at the numbers. They are stark.

Source: Zhenyuan Technology IPO Prospectus

In 2023, Tencent was a significant partner. By 2025, it is the entire business.

This level of concentration is usually a red flag for investors. If Tencent decides to switch vendors, Zhenyuan collapses. But in the context of China’s tech war, this is also a defensive moat. Tencent is one of China’s largest consumers of AI compute. Its WeChat, Video Account, advertising algorithms, content safety systems, and its own Hunyuan large language model all require massive inference power.

Tencent didn’t just invest in Zhenyuan; it co-created its supply chain. Tencent led the Pre-A round in 2018 and has invested in six consecutive rounds. This isn’t passive capital. It’s active integration. Tencent provides the real-world workload, the feedback loop for software optimization, and the guaranteed purchase volume that allows Zhenyuan to scale production.

For Tencent, this is risk mitigation. After US export controls tightened, relying solely on Nvidia became a strategic liability. By backing Zhenyuan, Tencent ensures it has a domestic backup for its inference workloads. It’s a classic vertical integration play, adapted for the semiconductor age.

The Cambricon Contradiction

While Zhenyuan leans on Tencent, its peer Cambricon is trying to prove it can sell to everyone. Or at least, it’s trying to deny that it’s selling to anyone specific.

On June 18, Cambricon’s stock surged over 16%, hitting a record market cap of nearly 965.6 billion yuan (about $138B). The rumor mill was churning. Online “internal meeting minutes” claimed Cambricon was supplying 120,000 to 160,000 units of its MLU580 chip to ByteDance in Q2 alone, with another model, the MLU690, coming later.

Cambricon’s securities representative shut this down immediately. “Please distinguish between official information and online rumors,” they told investors, stating they had not held any internal exchange events recently. Their website still lists older models like the Siyuan 370, not the newer 580 or 690.

This denial is telling. It suggests that while the demand for domestic chips is real (ByteDance is reportedly discussing purchases with other vendors like Enflame for at least 50,000 chips), the suppliers are walking a tightrope. They cannot confirm large deals due to potential export control sensitivities or contractual confidentiality. But the market believes them anyway.

The contrast is clear. Zhenyuan is open about its single-big-client model. Cambricon is hiding its big-client model. Both are navigating the same minefield.

What This Means for the Market

Zhenyuan’s IPO is not just a listing event. It is a stress test for the “domestic substitution” thesis.

For investors, the key question is whether Zhenyuan can diversify. Can it move beyond Tencent? The prospectus claims its hardware is adapted to 1,000 models, but if 84% of revenue comes from one company, the “adaptability” is likely tailored specifically to Tencent’s stack.

For operators and entrepreneurs, the lesson is about ecosystem lock-in. Nvidia’s CUDA is powerful not just because of the hardware, but because of the decades of developer investment. Zhenyuan’s “TopsRider” software is an attempt to replicate this, but it only works if you have a client like Tencent willing to do the heavy lifting of adaptation. For smaller companies, the migration cost remains prohibitive.

The broader implication is that China’s AI chip industry is fragmenting. It’s not a single race to replace Nvidia. It’s a series of walled gardens. Tencent has Zhenyuan. Alibaba likely leans on its own initiatives or Huawei. ByteDance is shopping around. This fragmentation reduces the overall efficiency of the Chinese AI sector but increases its resilience against external shocks.

Zhenyuan’s path is the hardest one. It bet on the DSA architecture, bet on the inference market, and bet on Tencent. If Tencent wins, Zhenyoin wins. If Tencent’s AI strategy shifts, or if Nvidia finds a way to re-enter the inference market with cheaper, more efficient chips, Zhenyuan faces an existential crisis.

The IPO opens the door, but the market will decide if the house is built on sand or steel.

Signals to Watch

For those tracking the sector, the metrics are shifting from “who has the best chip” to “who has the most sticky client.”

1. Diversification Rate: Watch Zhenyuan’s revenue breakdown in its first annual report post-IPO. If Tencent’s share drops below 50%, the thesis of independent commercial viability strengthens. If it stays above 70%, the company remains a Tencent subsidiary in all but name.
2. ByteDance’s Procurement: The rumors about ByteDance and Cambricon/Enflame need verification. If ByteDance, China’s most aggressive AI spender, commits to a multi-vendor strategy, it could break the Tencent-Zhenyuan duopoly and force price competition.
3. Software Adoption: The real moat is the “TopsRider” platform. Track how many third-party developers are building on it. If the ecosystem remains closed to Tencent’s internal tools, the DSA advantage is limited.

The race isn’t over. It’s just entering the commercial phase, where survival depends on more than just engineering talent. It depends on who holds the purse strings.