Summary of Key Points
The financial reports of domestic GPU manufacturers for the first half of 2026 have shown a collective “turnaround”: revenue has doubled, and losses have narrowed. However, behind this lies anxiety over competing for production capacity and securing supply chains (with a surge in inventory and prepaid payments), as well as the urgency to secure orders from major internet companies. The industry is transitioning from the “Self-Reliance Innovation Era” (government/state-owned enterprise orders) to the “Internet Era” (led by companies like ByteDance and Tencent). The essence of competition is a “dual-layer funnel screening process”: first, securing 7nm production capacity from SMIC to gain entry; then, passing the rigorous tests of these internet giants to grow and strengthen. Only 2-3 companies may ultimately succeed in both stages, and it remains the norm to experience high growth alongside significant investment.
I. Bright Financial Reports, but Hidden Concerns: Doubling of Inventory and Prepaid Payments, High Cash Flow Pressure
The revenue and profit figures for domestic GPU manufacturers look promising, but upon closer inspection of the financial reports, two sets of data are particularly striking:
- Surging inventory: Cambricon has 8.2 billion yuan in inventory, and摩尔 Threads has 3.55 billion yuan, which is equivalent to producing and stocking up on chips for several months in advance.
- Exponential increase in prepaid payments: Cambricon has prepaid 2.9 billion yuan, and Moore Threads has prepaid 1.34 billion yuan, paying in advance to foundries (such as SMIC) to secure future production capacity and raw materials.
- Cash flow pressure: Moore Threads experienced a net outflow of 2.169 billion yuan in operating cash flow, indicating that expenses exceeded earnings.
Why do they do this? Because everyone fears a shortage of production capacity—chip production cycles are long, and if they don’t stock up now or secure supply chains, they might miss market opportunities. It’s like bidding for concert tickets; you need to pay a deposit in advance to even get a chance to buy them.
II. Production Capacity: The Critical Threshold
For domestic GPUs to produce AI training chips, 7nm manufacturing is essential (12/14nm can only be used for low-end inference and cannot participate in large-scale model training). Currently, the only foundry capable of producing 7nm chips in China is SMIC. However, SMIC’s capacity is extremely tight:
- Limited actual production: SMIC’s N+2 process (similar to 7nm) has a monthly output of less than 20,000 wafers, with a yield of just over 50% (only 50 out of 100 wafers are usable). After deducting the share for mobile and communication chips, even less capacity remains for AI chips.
- Quota allocation based on background: Huawei has been allocated 43% of the capacity (covering mobile, communication, and AI applications, with the highest priority), followed by Cambricon (9%-11%). The remaining manufacturers receive less than half of the quota.
Companies that fail to secure 7nm production capacity are directly excluded from the large-scale model training market and can only focus on low-end applications. This is like the college entrance exam; 7nm is the “first-tier threshold,” and without it, they cannot enter the elite market.
III. Why the Focus on Internet Giants?
The demand for domestic GPUs falls into two categories:
- Self-Reliance Innovation Market: Orders from governments, state-owned enterprises, and telecom operators offer stability (driven by policy) but have a limited potential for growth due to project-based, customized, and long-payment-term contracts with uncertain repurchases.
- Internet Market: Orders from companies like ByteDance and Tencent are large and continuous—ByteDance’s 2026 chip procurement budget is 85 billion yuan, half of the non-internet market’s total. The upgrade of large models requires more computing power, and repurchases are certain. However, entry is difficult: companies must pass multiple tests (hardware, model compatibility, cluster trials), with a process lasting up to two years, requiring stable hardware, high cost-effectiveness, and good software integration.
Therefore, manufacturers are seeking to transition from the Self-Reliance Innovation market to the Internet market—self-reliance innovation helps them survive, while the internet market allows for growth. It’s like running a restaurant; you can initially rely on regular customers (self-reliance innovation), but to expand, you need to enter the mainstream market (internet).
IV. Current Status of Manufacturers: Who Is Leading? Who Has Concerns?
The current state of the five major domestic GPU companies (Cambricon, TianShiZhiXin, SuYuan Technology, Moore Threads, BiRen/MuXi) varies:
- Cambricon: The leader, having secured production capacity (11% of SMIC’s quota) and internet customer relationships (a core supplier for ByteDance), the only company to have completed the transition from self-reliance innovation to the internet market.
- TianShiZhiXin: A potential leader, in talks with ByteDance for a 50,000-chip order for inference purposes. If successful, it could catch up with Cambricon.
- SuYuan Technology: Focused on one customer (Tencent), accounting for 74.9% of revenue, but this concentration poses high risks due to limited diversification and lower gross margins.
- Moore Threads: Aggressive in strategy, with inventory increasing by 166% and revenue doubling, yet lacking large-scale orders, similar to Cambricon’s situation in 2024.
- BiRen/MuXi: Fast-growing but still in the self-reliance innovation phase, not yet part of the core supply chains of major internet companies.
In summary, Cambricon has already succeeded, TianShiZhiXin and SuYuan Technology are on the verge of success, Moore Threads is struggling, while BiRen/MuXi is just starting out.
V. Future Outlook: Only 2-3 Companies Will Survive; We Are Still in a “Cash Burning” Phase
The domestic GPU industry is undergoing a structural transformation, with the self-reliance innovation market being limited and the internet market as the main driver of growth. To establish a foothold in the internet market, companies must overcome both production capacity and customer acquisition challenges—these hurdles act like a funnel, filtering out most players.
Only 2-3 companies will ultimately survive, similar to how the mobile phone industry evolved from a competitive landscape to a few dominant players. Manufacturers are currently investing heavily (in inventory, research and development, and customer acquisition), but those who successfully pass these challenges will become the core players of the next decade.
Conclusion: The “first half” of the domestic GPU market (self-reliance innovation) is coming to an end, and the “second half” (internet) has just begun. This competition is not about who runs fastest now, but who can secure production capacity and pass the tests of major companies, ultimately staying in the game. For investors, it’s important to assess a company’s production capacity reserves and customer progress, rather than focusing solely on surface-level revenue growth.
*(Disclaimer: This article does not constitute investment advice; risks are borne by the reader.)*