In the intelligent computing exhibition area of the 2026 WAIC exhibition, the real-time operation of the Wanka level domestic cluster is surrounded by industry audiences on the large screen. Behind the jumping token output figures, it is a clear signal for the entire domestic computing industry to bid farewell to the flat label and start a new cycle of independent growth. The breakthrough of domestication penetration rate from 20% to 80% is becoming the core growth mainline of China's AI industry in the next five years.

Triple barriers to collaborative breakthrough

The current large-scale implementation of domestic computing power still faces triple stacking bottlenecks. The first hurdle before large-scale commercialization lies in the inherent shortcomings of insufficient video memory capacity on the hardware side, low inter card interconnect bandwidth, and inadequate fusion operator performance. The problem of software side M × N adaptation is prominent, and the migration of operator framework toolchains between massive models and more than ten domestic chips cannot be independently completed by a single enterprise.

The problem of ecological fragmentation further prolongs the implementation cycle. There are about ten leading GPU manufacturers in China, combined with supporting enterprises such as complete machine switches, optical modules, and liquid cooling. The main participants in the industry chain are highly dispersed, and the software and hardware systems of each manufacturer are incompatible with each other. Customers who change computing power suppliers have to reconstruct the entire business system, resulting in high trial and error costs. Combined with the negative cognitive inertia left by past pilot projects and the short-term cost disadvantage caused by the lack of large-scale shipments, multiple factors have jointly raised the threshold for the landing of domestic computing power.

At present, the industry has explored a feasible breakthrough path, relying on heterogeneous hybrid inference technology, mainstream domestic chip MFU can achieve a performance improvement of 85% to 152%, and the overall inference cost-effectiveness is 1.25 times that of overseas H-series. At the same cost, the token output scale is 2.5 times larger than that of domestic homogeneous solutions. With the mass landing of domestic clusters at the 10000 card level, the economies of scale brought about by the release of chip production capacity will quickly smooth out the cost gap. Coupled with the continuous landing of benchmark cases in fields such as advanced manufacturing and scientific research in government affairs, soft card points will gradually be resolved.

Changing lanes and overtaking open up a whole new space

The new computing power demand brought about by the large-scale outbreak of agents enables domestic computing power to bypass single point performance catch-up and achieve lane changing overtaking at the system level innovation level. Domestic giants have generally concentrated their overseas inventory of high-end cards on a very small number of super large model pre training, fully introducing domestic computing power in inference and fine-tuning processes. The heterogeneous hybrid push mode allows different architecture chips to perform their respective duties, activating idle computing assets of different brand generations and significantly improving resource utilization efficiency.

The business model has also shifted from selling hardware computing power to delivering standardized token services. By 2026, the token market will enter an exponential growth channel. Top scale token factories can dynamically allocate high and low performance computing power resources, distinguish high-value paying customers from testing needs, and avoid idle loss of high-end computing power. The industry's profit model has been implemented. The daily average token service volume of the current Shangtang large device platform has reached 2.42 trillion yuan, and it is expected to exceed 10 trillion yuan per day in the fourth quarter of 2026, an increase of about 25 times compared to the beginning of the year.

The long-term frontier layout has been launched, and space computing, optical computing, quantum computing has become the reserve direction of leading manufacturers. Relying on on orbit satellite fusion communication and on-board computing capabilities, it can quickly meet the needs of overseas AI computing power of countries co constructing the the Belt and Road, and seize the first mover advantage of global AI infrastructure.

From the narrative of catching up with overseas products in the past to establishing an independent value coordinate that is affordable and controllable, domestic computing power is completing the reconstruction of the entire industry valuation system. The process of domestic substitution in the next five years is essentially a comprehensive reshaping of the entire artificial intelligence ecosystem. This path, which has no ready-made experience to copy, will eventually give birth to a new stable state of China's AI industry.Editor/Cheng Liting