The year 2025 has proven to be a fruitful period for domestic AI chip manufacturers. Cambricon Technologies Corporation Limited achieved full-year profitability, while several leading GPU companies, often referred to as the "Four Little Dragons," completed their public listings. Notably, some of these firms have seen AI training scenarios become their dominant revenue source. This growth is underpinned by ongoing advancements in AI infrastructure development and the continuous strengthening of the domestic computing power ecosystem.
According to research from third-party agency IDC, domestic AI chips accounted for 41% of shipments in the Chinese market during 2025, surpassing the 40% mark for the first time.
Analysis reveals that alongside the sustained rapid growth of accelerated computing chip manufacturers—such as those specializing in GPUs and ASICs—CPU chips are also becoming increasingly scarce. A trend of CPU chips being in short supply has recently emerged in the market.
A senior industry professional emphasized the situation, stating, "The current challenge with CPUs is that supply is very tight. Even after price increases, securing supply is not guaranteed. The CPU is an indispensable component within the entire ecosystem."
Financially, major AI chip manufacturers saw potential orders grow by a certain percentage in 2025, prompting companies to actively build inventory to mitigate supply chain fluctuations.
The performance surge of domestic AI chips in 2025 is not merely a result of market-driven demand but also a key indicator of the further integration and development of domestic computing power. Some perspectives suggest that 2026 could become a year of accelerated development for domestic super-node solutions, which would also support the continuously rising demand for token consumption.
Leading computing power companies are reporting stable profitability, with both general-purpose and AI computing chips experiencing tight supply.
Domestic AI chip companies generally achieved significant revenue in 2025, alongside accumulating robust new potential contracts. Hygon Information Technology Co.,Ltd., a leading CPU chip company, reported annual revenue of 14.377 billion yuan for the 2025 fiscal year, a year-on-year increase of 56.92%. Its net profit attributable to shareholders reached 2.545 billion yuan, up 31.79% year-on-year. The company's announcement indicated that it seized opportunities in the high-end chip market, maintained high investment in product performance R&D and services, and gained widespread customer recognition.
Beyond strengthening R&D, ecosystem collaboration was also a key initiative. The company deepened industry collaboration with manufacturers and partners during the year, expanded product application areas, advanced the market share of domestic products, and achieved significant shipment growth.
However, Hygon's gross margin declined by 5.92 percentage points year-on-year to 57.78%. The announcement attributed this to changes in the product sales mix and price increases across multiple supply chain segments.
A major highlight for ASIC giant Cambricon Technologies Corporation Limited was achieving stable profitability in every quarter of the year, while also maintaining a relatively stable gross margin performance amidst supply chain volatility.
In 2025, Cambricon reported revenue of 6.497 billion yuan, a surge of 453.21% year-on-year. This was primarily driven by the continuous climb in computing power demand from the AI industry, coupled with the company's ongoing market expansion and active promotion of AI application scenarios. Benefiting from revenue growth, the company's net profit attributable to shareholders reached 2.059 billion yuan, a significant turnaround from a full-year loss of 452 million yuan in the previous year. Furthermore, the company maintained positive net profit in every single quarter of 2025, signaling its entry into a phase of stable profitability.
From a product structure perspective, the cloud product line contributed the vast majority of Cambricon's revenue. This business generated 6.479 billion yuan in revenue in 2025, with a gross margin of 55.22%, a slight decrease of 1.47 percentage points year-on-year. Although the gross margin for the edge product line decreased by 6.98 percentage points to 44.52%, its contribution to overall revenue was relatively low, thus not significantly impacting the composite gross margin.
Additionally, it is noteworthy that while GPUs are widely considered crucial in the AI era, the advent of the "Agent元年" has positioned CPU chips as a scarce asset.
The aforementioned industry analyst explained that the core logic behind the current CPU shortage is twofold. First, against the backdrop of global efforts to build AI infrastructure, production capacity for high-process wafers (e.g., 2nm-3nm) at foundries is tight, with a tendency to prioritize GPU chip production, thereby squeezing out CPU capacity. Second, no matter how important GPUs become, CPUs remain an essential component in both AI servers and general-purpose servers, indicating a stable and continuously growing demand space for CPUs.
The heightened importance of CPUs is also reflected in the moves of major international players. NVIDIA recently released a CPU product named "Vera" after a hiatus of many years. Meanwhile, Arm, previously focused on semiconductor IP, announced its entry into the CPU chip design space and has established deep business connections with giants like Meta and OpenAI.
Arm CEO Rene Haas publicly stated that CPU chips are undertaking core computing tasks in the AI era. He highlighted that the recent rise of agents, exemplified by OpenAI's "Claw," involves substantial scheduling and computational workloads that must be handled by CPU chips—a capability that accelerated chips like GPUs cannot replace.
Consequently, it is evident that both general-purpose computing CPUs and accelerated computing chips like GPUs and ASICs are experiencing a period of rapid performance growth.
Beyond impressive revenue performance, domestic AI chip manufacturers are also actively building inventory to prepare for future supply chain volatility and potential contract opportunities.
Hygon Information Technology Co.,Ltd. saw a substantial increase in potential orders. Its financial report showed contract liabilities of 2.019 billion yuan for the year, a surge of 123.42% year-on-year. The proportion of contract liabilities to total assets increased from 3.16% in the same period last year to 5.66%.
The announcement indicated this increase was due to higher customer prepayments following recognition of the company's products, with the company obligated to deliver according to customer-specified milestones, leading to a significant rise in unsettled contract liabilities at period-end.
Correspondingly, the company's inventory and prepayments also grew. Hygon's prepayments reached 2.885 billion yuan in 2025, a sharp increase of 132.62% year-on-year, with their proportion of total assets rising from 4.34% to 8.1%. This was attributed to longer raw material procurement cycles and a further increase in strategic inventory stocking. The company's inventory for the year stood at 6.406 billion yuan, up 18.1% year-on-year.
Generally, the "contract liabilities" item in financial statements indicates a company's potential contract pipeline. Preparing for these potential orders, especially amid ongoing price increases for some supply chain materials, necessitates appropriate inventory buildup. It is therefore unsurprising to see corresponding increases in inventory across these companies.
Cambricon Technologies Corporation Limited's inventory as of the end of 2025 was 4.944 billion yuan, a substantial increase of 178.67% year-on-year, accounting for 36.79% of total assets. In the same period last year, inventory was 1.774 billion yuan, representing 26.41% of total assets. The announcement cited an increase in raw materials at period-end as the reason for the growth.
Other companies like MXIC also reported significant inventory growth, often linked to ensuring timely product delivery to downstream customers and preparing for increased demand and supply chain resilience.
Behind the robust inventory levels and potential orders lies the continuous improvement and capability integration within the domestic computing power ecosystem.
Companies like MXIC highlighted their rapid adaptation capabilities with domestic large language models in their financial reports. For instance, MXIC's software stack achieved high compatibility with the CUDA ecosystem, supporting adaptation for over 13 major models on launch day.
Furthermore, although a proposed merger between Hygon Information Technology Co.,Ltd. and Sugon did not materialize, the two companies continue to collaborate deeply. Sugon previously launched the world's first cable-less box-type super-node, scaleX40, which utilizes an orthogonal wireless primary interconnect architecture enabling direct connection between compute and switch nodes, eliminating performance losses and maintenance risks associated with cables from the source.
According to analysis by Dongwu Securities, Hygon's DCU provides the autonomously controllable computing core for Sugon's super-node. Its chip-level interconnect technology, HSL, works in concert with Sugon's scaleFabric high-speed network to achieve full-stack interconnect synergy, propelling the domestic AI computing platform towards system leadership.
This year is also considered crucial for the deployment of domestic super-node solutions. Huatai Securities pointed out that while market concerns exist that the potential entry of advanced foreign cards could reduce demand for domestic super-nodes, the institution believes that although some substitution may occur, domestic super-nodes are likely the long-term trend. Utilizing both advanced foreign cards and domestic super-nodes is not mutually exclusive for Cloud Service Providers. The reasoning is that foreign card supply itself is limited, only meeting part of the AI computing demand. Foreign cards are suitable for AI training, while CSPs can still use domestic card super-nodes for inference tasks.
Improvements in profitability, optimization of product portfolios, accelerated ecosystem collaboration, and the accumulation of order reserves collectively demonstrate that domestic computing power is steadily capitalizing on key opportunities presented by AI infrastructure construction and the rise of AI agents. Certainly, a particular challenge this year lies in ongoing supply chain volatility testing manufacturers' profitability, while the domestic ecosystem must continue efforts towards compatibility, making the trajectory of the domestic computing power chain in 2026 even more compelling.
Comments