Editor's Note:

The rapid evolution of Generative AI and the soaring demand for computational power in autonomous driving and intelligent manufacturing are catalyzing a profound transformation in the valuation of computational resources. The traditional "stack computing power" model, anchored on hardware scale, has stumbled into a development quagmire characterized by high investment and low efficiency. The core proposition of the AI era has shifted to the "Efficiency Revolution"—centering on computational efficiency as the pivotal metric, and facilitating the transition from "scale-centric" to "efficiency-centric" value through technological innovation, architectural refinement, and scenario integration. This revolution is pivotal not only to the success of enterprises' intelligent transformation but also to reshaping the productivity landscape in the digital economy era.

The Paradox of Scale and Inefficiency

The explosive growth of Generative AI has spurred an unprecedented surge in computational demand. Recently, the "2025 White Paper on the Development of New-Quality Computing Power" jointly published by IDC and NTT revealed that in the first half of 2024, the market size of intelligent computing services (GenAI IaaS) for Generative AI in China's intelligent computing service market reached RMB 5.2 billion, with a year-on-year growth of 203.6%, marking a remarkable acceleration. However, this robust demand stands in stark contrast to the severe bottlenecks faced by traditional computational infrastructure in terms of technical architecture, return on investment, and usage efficiency, trapping it in a vicious cycle of "scale stacking - inefficiency - further scale expansion."

Many enterprises blindly pursue hardware parameters in building computational capacity, measuring strength solely by the number of servers acquired, neglecting deep alignment with business scenarios. This extensive approach has led to a vast amount of idle computational resources. IDC research indicates that the global resource utilization rate of traditional computational centers generally falls below 30%, with "computational waste" becoming a significant burden on enterprises' intelligent transformation.

Furthermore, traditional computational management and scheduling technologies lag far behind demand, making it impossible to achieve dynamic and optimized allocation of computational resources. In deep learning model training, the phenomenon of "computational stacking" is particularly pronounced—enterprises invest hundreds of GPUs to accelerate training, but due to issues such as memory access bottlenecks and data transmission delays, the actual computational efficiency is less than 40% of the theoretical value.

It is noteworthy that the construction of data centers, hardware procurement, and long-term operation and maintenance in computational infrastructure constitute substantial expenses, while accelerating technological iteration leads to high equipment replacement costs. More alarmingly, a large number of small, medium, and micro-enterprises are hindered from intelligent transformation due to the high cost of "computational stacking," exacerbating the imbalance in the development of the digital economy.

From a macro perspective, with the progression of "dual carbon" goals, the energy consumption issue of computational centers has become increasingly salient. Traditional air-cooling technology has reached its energy efficiency limit, with the Power Usage Effectiveness (PUE) value of data centers generally exceeding 1.5, and even surpassing 1.88 in some high-load scenarios. This implies that for every kilowatt-hour of electricity consumed for computation, more than 0.5 kilowatt-hours are consumed by auxiliary systems such as cooling, contradicting the concept of green development.

Breaking the Shackles of "Hardware Stacking"

Confronted with the dilemma of traditional computational power, the crux of the "Efficiency Revolution" lies in using "computational efficiency" (i.e., effective computational output per unit of energy consumption) as the core indicator to achieve a systematic enhancement in computational value through technological innovation, architectural refinement, and scenario integration. This transformation is not an upgrade of a single technology but a comprehensive reconstruction encompassing hardware design, software scheduling, and application optimization.

Hardware-level advancements focus on improving the "energy efficiency ratio" rather than merely pursuing peak computational power. For instance, liquid cooling technology can reduce the PUE of data centers to below 1.15, saving over 40% of energy compared to traditional air-cooling solutions while tripling the computing density per cabinet. In the realm of chip design, mixed-precision computing (such as dynamic switching between FP16 and INT8) has become mainstream. In large model training scenarios, adopting mixed-precision computing can increase computational utilization by 30% while sacrificing only 1% of accuracy, significantly reducing training costs. Additionally, the proliferation of heterogeneous computing architectures (CPU+GPU+FPGA, etc.) enables on-demand allocation of diverse computational resources. For example, utilizing dedicated acceleration chips in AI inference scenarios can enhance computational efficiency by 8-10 times compared to general-purpose CPUs.

The linchpin of the Efficiency Revolution is to transcend the mindset of "hardware stacking" and achieve optimal allocation of computational resources through intelligent software scheduling. The new generation of computational management platforms boasts three key features: Firstly, they dynamically perceive business needs, such as automatically adjusting computational allocation based on model training progress; secondly, they support fine-grained resource segmentation, upgrading the traditional "whole machine allocation" model to on-demand combinations of "cores/memory/storage"; thirdly, they integrate fault prediction and disaster recovery mechanisms. An autonomous driving company improved computational resource utilization from 28% to 65% through an intelligent scheduling platform while controlling fault recovery time within 10 seconds. Furthermore, the maturation of open-source frameworks and tools (such as optimization plugins for TensorFlow and PyTorch) enables enterprises to enhance computational efficiency without custom development, lowering the technical threshold.

The ultimate aim of the Efficiency Revolution is to foster deep integration of computational power and business scenarios to maximize "effective computation." In inference scenarios, efficiency optimization prioritizes low latency and high concurrency. A short video platform enhanced its recommendation model's lightweight transformation and inference engine optimization, increasing its daily processing capacity from 218 billion tokens to 350 billion tokens without augmenting computational investment, supporting a 200% increase in user growth.

In training scenarios, efficiency optimization emphasizes balancing data parallelism and model parallelism. A medical AI team adopted pipeline parallelism technology, reducing the training time of multimodal medical imaging models from 15 days to 4 days while lowering computational costs by 40%. More critically, efficiency optimization is extending from single scenarios to the entire chain. For instance, in the manufacturing industry, enhancing efficiency throughout the entire cycle, from product design simulation to production process optimization, can shorten the new product development cycle by over 30%.

The proliferation of Generative AI has positioned internet enterprises as pioneers of the Efficiency Revolution. In large model training scenarios, techniques such as mixed-precision training and gradient compression have reduced the training cost of a language large model by 60% while maintaining model performance; in inference, adopting model distillation and quantization techniques has reduced end-side computational requirements by 80%, supporting tens of millions of concurrent users.

From "Quantity Owned" to "Value Created"

The essence of the Efficiency Revolution is an inevitable requirement for productivity development in the digital economy era. As computational power transforms from a "scarce resource" to a "general capability," its value assessment criteria will inevitably shift from "quantity owned" to "value created."

Traditional IT investment emphasizes "hardware configuration," whereas the Efficiency Revolution urges enterprises to shift to "business value"-oriented computational assessment. The PEEIE (Product, Efficiency, Engineering, Industry, Ecology) computational facility selection criteria proposed by IDC encapsulate this transformation—no longer focusing solely on hardware indicators such as processor performance and storage capacity, but evaluating the actual value of computational solutions to the business from five dimensions: product richness, computational efficiency indicators, engineering capabilities, industry adaptability, and ecological synergy. After applying this standard for selection, a financial enterprise reduced its computational investment by 30% but increased its business processing capacity by 50%, proving that efficiency-oriented value assessment is more aligned with the needs of the AI era.

The Efficiency Revolution will propel the transformation of the computational industry from "hardware sales" to "efficiency services." On one hand, hardware vendors need to shift from "selling equipment" to "selling efficiency," with server vendors not only providing hardware but also software tools and services for efficiency optimization. On the other hand, new efficiency service providers will emerge, specializing in niche areas such as computational scheduling and application optimization. This refined division of labor will enhance the overall industry efficiency. A computational service company provides full-cycle services of "computational diagnosis - solution design - continuous optimization" to manufacturing clients, improving their computational efficiency by over 40% and securing stable service revenue for itself, fostering a win-win scenario.

A leading cloud service provider estimates that if the efficiency of the entire industry is improved by 50%, global computational center energy consumption in 2030 can be contained within 1.2 times that of 2025, while computational supply capacity can increase by 5-8 times, achieving sustainable development with "computational growth and energy consumption control."

Conclusion

Looking back from the threshold of the AI era, the transition from "computational stacking" to the "Efficiency Revolution" is not merely a choice of technological path but also an innovation in development philosophy. As computational efficiency becomes the core metric for measuring digital productivity, and efficiency enhancement becomes the key driver for enterprises' intelligent transformation, the value of computational power no longer hinges on cold hardware parameters but on the actual value it creates for business innovation, industrial upgrading, and social progress.

The wave of the Efficiency Revolution has arrived, and only by embracing this trend can one gain a competitive edge in the AI era, making computational power the core engine driving the high-quality development of the digital economy.