Summary of Key Points
At the 2026 World Artificial Intelligence Conference, experts discussed the "changes in the computing power supply side during the Token economy era" and agreed that to support the sustainable development of the AI industry, it is essential to build a new computing power ecosystem that integrates both hardware and software, as well as terrestrial and space-based resources. The Token economy has shifted the demand for computing power from focusing solely on "training" to emphasizing both training and inference, presenting differentiated opportunities for domestic computing power solutions. However, this also brings challenges in areas such as systems engineering and software ecosystems. Space-based computing power (computing in space) offers innovative solutions that can overcome limitations related to energy consumption and heat dissipation on Earth, meeting the needs for low latency and high reliability. Additionally, the development of computing power is moving towards more integrated heterogeneous systems, full-stack optimization, and a focus on operational efficiency.
1. The Arrival of the Token Economy: Changing Computing Power Demands
In the past, large models relied primarily on extensive training, which required massive amounts of computing power. With the advent of the Token economy, the demand for "inference" (real-time calculations when AI systems answer questions or perform tasks) is increasing. There is a growing emphasis on computing power that must be fast (low latency) and reliable (high reliability). This shift poses challenges to the computing power supply side but also opens new possibilities for domestic computing power manufacturers. They no longer need to compete solely with foreign chips in terms of absolute performance; instead, they can optimize their solutions for commonly used models, focusing on cost-effectiveness.
2. Sending Computing Power into Space: Solving Earthly Challenges
Earth-based computing power faces two significant limitations: high energy consumption and complex heat dissipation issues. Space-based computing power leverages the advantages of space, such as abundant solar energy (no need for terrestrial power grids) and a vacuum environment that facilitates efficient heat dissipation (no need for sophisticated cooling systems). This technology can be applied in scenarios where extreme performance is required, such as financial transactions that require millisecond-level responses, precise navigation for ocean-going vessels, and timely disaster warnings.
3. Opportunities for Domestic Computing Power
Domestic chips do not need to compete directly with NVIDIA in terms of general performance but can pursue a differentiated approach:
- Optimizing them for commonly used AI models to improve efficiency and cost-effectiveness;
- Using them in conjunction with NVIDIA chips through heterogeneous deployment, where domestic cards handle simpler tasks while NVIDIA cards handle more complex ones.
However, domestic computing power also faces several challenges, including:
- Systems engineering for large-scale clusters;
- Coordinating the performance of different types of chips;
- Managing energy consumption and heat dissipation;
- Lacking a unified software ecosystem (similar to NVIDIA's CUDA) that makes development easier for developers. Experts call for the establishment of a "Chinese version of the CUDA alliance" to standardize practices and promote collaboration.
4. Four New Directions in Computing Power Development
Experts from China Mobile have identified four key trends:
- From Focusing on Hardware to Emphasizing Software: Moving away from relying solely on single chips to using a combination of CPU, GPU, and specialized chips to maximize their respective strengths.
- Beyond Peak Performance: Shifting from comparing the highest computing power values to addressing bottlenecks in storage and communication speeds.
- From Single-Point Optimization to Full-Stack Optimization: Optimizing not only chips but also the entire chain, including hardware, software, algorithms, and operations.
- From Construction to Management: Focusing on improving the efficiency of existing computing power infrastructure rather than just building more centers.
Inference efficiency is crucial; only by ensuring fast and cost-effective inference can AI services be profitable. Space-based computing power plays a vital role in enabling collaborative inference across various platforms, including remote areas.
5. The Future of Computing Power Ecosystems: Integrating Terrestrial and Space-Based Resources
Experts agree that the future computing power ecosystem must combine terrestrial and space-based resources. Earth-based computing power will meet most regular needs, while space-based computing power will handle extreme scenarios. The systems engineering expertise gained in space (e.g., radiation resistance and energy management) can also help address challenges associated with heterogeneous clusters on Earth. A combination of hardware and software, along with the integration of terrestrial and space-based resources, is essential for the true advancement of the AI industry.