Outlook 2024: Nine Major Trends in Computing Power Development

9 min readFeb 7, 2024

The computing power infrastructure and industrial scale are growing rapidly, with significant empowerment effects. Computing power has become a major driving force for economic growth, and countries around the world are continuing to increase their investment in computing power infrastructure. Morgan Stanley predicts that the capital expenditure of the top ten global cloud computing service providers will reach $200 billion in 2024, with additional investment focusing on the artificial intelligence field.

In China, the development of computing power infrastructure is accelerating its transition from “supply-side optimization with communication computing as the main focus” to “demand-driven with intelligent computing at the core,” increasingly highlighting its role in supporting the emergence of new economic development dynamics. According to data released by the Ministry of Industry and Information Technology and CCID Consulting, in 2022, China’s share of high-performance computing power was nearly 20%, with the explosion of large models driving exponential growth in high-performance computing power. In the first half of 2023, smart computing accounted for over 50% of the newly added computing power facilities in China, with the overall computing power scale reaching 197 EFLOPS. It is expected to drive the core industrial scale of computing power in China to exceed 2 trillion yuan.

Taking into account the external environment of the comprehensive information and communication industry, with favorable policies and strong demand combined with technological innovation, it is expected that the computing power industry centered on intelligent computing will continue to maintain a high-speed growth trend in 2024, presenting nine major new trends overall.

Rapid Growth in Market Scale

The acceleration of ToB/C large-scale model applications and the rapid evolution of multimodal large-scale models further drive the robust development of intelligent computing industries.

In 2023, large language models such as ChatGPT have sparked a ‘chip grabbing war,’ leading to explosive growth in computing power demand driven by AI. In 2024, as large models accelerate penetration into vertical industries and domains, platforms like GPT Store further promote the landing of AI-native applications, and the rapid development of multimodal models, the intelligent computing market will continue to maintain a high-growth trend.

In the overseas market, FactSet and Bloomberg predict that by 2024, the revenue from cloud business and the investment growth in cloud infrastructure of the top three cloud providers in North America (AWS, Microsoft, and Google) will reach 22.5% and 16.6%, respectively, a slight increase compared to 2023, overall indicating optimism.

In the domestic market, according to the development goals outlined in the ‘Action Plan for High-Quality Development of Computing Power Infrastructure’ jointly issued by six departments including the Ministry of Industry and Information Technology, from 2023 to 2025, China’s computing power scale will have a compound growth rate of 18.5%. The newly added computing power scale in 2024 is expected to approach 40 EFLOPS, and the core industrial scale of computing power is expected to exceed 2.4 trillion yuan.

Accelerated Formation of New Industrial Patterns

AI computing power promotes the upgrading of IDCs and cloud services, driving the formation of a new pattern for computing power services called ‘three-three-one.’

In 2023, third-party computing power leasing companies specializing in algorithm construction and leasing solutions, with abundant GPU resources, emerged as new forces in the computing power service market, attracting capital attention.

By 2024, IDC service providers, cloud service providers, and third-party computing power leasing companies will become the ‘three major operational entities’ in the computing power market, dividing the overall computing power service market and providing ‘three types of computing power service models’: computing power leasing, computing power + platform services, and computing power + platform + model services. In particular, the third-party computing power leasing model is expected to replicate the high-speed growth path of third-party IDC service providers. Under the drive of cooperation with high-quality customers and utilizing their own channel and resource integration capabilities, they will build a ‘customer-capital-AI computing power’ closed-loop expansion.

Meanwhile, the computing power sharing alliance model will gradually emerge, with multiple startups jointly purchasing and sharing GPU computing power, reducing the overall cost of computing power usage. The industry market will form a ‘three major operational entities,’ ‘three types of computing power service models,’ and ‘a new type of computing power joint purchase and sharing alliance’ pattern.

Highlighted Scale and Intensive Construction

Breaking through the scale of large models drives the clustering of intelligent computing, with E-level or above large-scale intelligent computing centers becoming mainstream.

The breakthrough of large model parameter scale crosses orders of magnitude, leading to a multiplication of data set size and an urgent need for massive computing power support, which promotes the continuous improvement of AI server performance, presenting a clear trend of large-scale and clustered intelligent computing layout.

Firstly, the improvement of the performance of individual servers can effectively reduce the transmission delay of parameters and data between servers, improve computing efficiency, and it is expected that in the second half of 2024, high-performance GPU cards surpassing the existing H series will be launched.

Secondly, relying on high-performance GPU cards to build supercomputers (server clusters), E-level intelligent computing clusters will become mainstream. According to CCID’s report, by the end of 2024, 5% to 8% of enterprises in China will upgrade their large model parameters from the hundred billion level to the trillion level, and the demand for computing power will increase by 320%. Large-scale models parameters introduced by Google, Microsoft, and others will evolve to the scale of hundreds of billions or trillions, focusing on building E-level intelligent computing clusters for large model training.

According to public statistics, there are only five super E-level intelligent computing centers in China. It is expected that by 2024, the proportion of super-large-scale intelligent centers will steadily increase.

Deepening Layout of Computing Power

AI large models extend to the edge and end, and intelligent computing infrastructure accelerates penetration into cities and edges.

As multimodal large models and underlying computing power technologies continue to improve, deploying large models on the edge and mobile ends becomes an inevitable trend, and the layout of intelligent computing centers towards cities and edges will become more apparent.

On the one hand, the expansion of AI large models brings a surge in inference computing power demand, driving the sinking of distributed inference computing centers, and local or nearby deployment of computing power can effectively alleviate cost pressure.

On the other hand, AI large models are gradually sinking to edge terminals such as smart cars, computers, and mobile phones, integrating into terminal intelligent bodies. Edge computing power can effectively meet the rapid response requirements of low-latency AI applications. Following cities like Chengdu, Beijing, Shanghai, Shenzhen, more cities will introduce policies for the high-quality development of computing power in 2024, coordinating urban and industrial smart computing resource demand, and accelerating the upgrade of urban computing power infrastructure.

Integration of Cloud Intelligence and Training and Inference

Integration becomes the mainstream service model of intelligent computing, realizing efficient collaboration between computing power, algorithms, and data.

Based on data, cloud computing-based integrated intelligent computing services will become mainstream, realizing efficient collaboration between computing power, data, and algorithms to meet the requirements of data processing, storage, transmission, and other aspects in intelligent computing application scenarios.

The integration of AI and cloud computing development has become a consensus among leading cloud providers. AWS and NVIDIA have carried out full-stack cooperation in AI infrastructure, acceleration libraries, and basic models, aiming to make AWS the best cloud environment for running GPUs. Recently, Alibaba Cloud has successively launched functions such as one-click deployment of large models to databases and function computing to optimize the AI development process on the cloud. Baidu has upgraded its Lingjing Matrix platform to an intelligent body platform, shifting its focus from model layer to ecosystem and application cultivation.

In 2024, the focus will be on the full-stack integrated service of ‘AI + cloud + data,’ and the development paradigm and industrial ecology of ICT will be further reconstructed. A series of products such as cloud hosts, storage, and databases will be fully upgraded for AI, and the entire process of model use, including data processing, training, fine-tuning, and inference, will tend to be realized in the same service environment.

Acceleration of Ubiquitous Computing Power Network

New technologies for computing power networking are flourishing, and computing power center interconnection and internal network urgently need to break through bandwidth bottlenecks.

Recently, the National Development and Reform Commission and other five departments jointly issued the ‘Implementation Opinions on Deepening the Implementation of the ‘East Number West Calculation’ Project and Accelerating the Construction of a National Integrated Computing Power Network,’ which clearly outlines the creation of hierarchical network latency circles to meet the requirements of multi-level computing power service systems for differentiated urban-regional-national levels, providing guidance for computing power networking.

In 2024, in terms of computing power center interconnection networking, with the release of new generations of high-performance chips, continuous demand for 800G and 1.6T will be generated, and key technologies such as long-distance RDMA, hundred-petabyte-level all-optical interconnection, and new optical fibers will further breakthroughs, leading to continuous improvement in the low-latency and deterministic guarantees of computing power interconnection. In terms of internal networking of computing power centers, non-blocking and high throughput are the core demands for accommodating large model training, which will drive the maturity of RoCEv2-related algorithms.

Unified Public Computing Power Scheduling

The continuous breakthrough of computing power inclusive services, and the emergence of regional and urban-level service platforms.

The construction and trial operation of unified computing power scheduling platforms and interconnection platforms, operated by governments, operators, and cloud providers, to support the integration of computing, networking, and cloud services, will realize the transformation from ‘task finding computing power’ to ‘computing power adapting to tasks,’ effectively solving problems such as dispersed computing power resources, multiple subjects, imbalanced supply and demand, and high usage costs.

In 2024, public computing power unified scheduling services will undergo new changes. On the one hand, public computing power platforms based on IXP will become the main form, and additional computing power scheduling, supply-demand docking, etc., will accelerate pilot applications. On the other hand, the aggregation effect of computing power ecology around hub nodes will be further strengthened, and regional-level and urban-level public computing power service platforms will initially emerge, forming a joint operation body with industry influence.

Acceleration of Diversification and Localization

Chip types, architectures, and supply show a trend towards diversification, and domestic chip autonomous ecosystems are accelerating construction.

Under the background of widespread application of new technologies and global computing power shortage, chip types, architectures, and suppliers are all showing a trend towards diversification, and the ability of independent and controllable domestic chips will be further enhanced. Regarding chip types, the industrialization of emerging technologies such as 5G, artificial intelligence, and autonomous driving is accelerating, and industry and policy attention are expanding from CPU-centric to high-performance computing chips, storage chips, etc.

In terms of chip architecture, RISC-V, with its excellent flexibility and scalability, breaks through the industrial pattern dominated by x86 and ARM architectures. The industry has successfully explored the application of RISC-V in the field of AI.

Regarding chip manufacturers, mainstream cloud providers are developing their own chips to balance the monopoly of companies such as Intel and NVIDIA through software and hardware coordination. Huawei, SMIC, Changxin, and other domestic manufacturers have made breakthroughs in recent years.

In 2024, the focus will be on the construction of the domestic chip ecosystem, including chip design, manufacturing, packaging, and related software, systems, frameworks, and other key links.

Evolution of Traditional DCs to AIDCs

AI computing power leads the rapid development of data centers towards greening, modularization, and intelligent operation and maintenance.

Data centers are rapidly evolving into intelligent computing centers, with the introduction of liquid cooling, modular construction, and intelligent operation and maintenance.

Firstly, the application of liquid cooling will gradually shift from local pilot to climbing period. When the power density of a single cabinet reaches 20KW, the investment cost of liquid cooling is basically equivalent to that of air cooling. Liquid cooling applications will accelerate popularization in 2024, and more than 50% of domestic data center projects will be covered by 2025.

Secondly, the engineering implementation capability of intelligent computing centers such as flexible adaptation and rapid delivery will be enhanced. Modular design and standardized packaging around cooling, power supply, and intelligent systems, joint industry innovation, and system solutions will become more abundant.

Thirdly, the level of intelligent operation and maintenance of data centers is expected to reach a new level, especially through the use of AI and other technologies, fully leveraging the advantages of automation and intelligence in power supply and air conditioning equipment to achieve precise control of operation energy consumption and smart operation.

What is Utility?: Aimed at providing various utility services such as computing resources, data storage, algorithm models, etc. It is a decentralized computing service platform built on blockchain technology and chip consensus. https://utnet.org/learning/utility_explain

Utility Whitepaper download link: http://oss2.xuanwoo.com/UtilityNetWhitePaper.pdf

Detailed explanation of Utility chip parameters:

https://mp.weixin.qq.com/s/roJHlQw9E12lUEQUdAgAHg

About Utility

Utility Net (UNC) is a decentralized digital chip network composed of chips defined by BDC (Blockchain Define Chip), which uses chip self-verification to prove and efficiently provide powerful on-chain resources at zero cost.

*Utility official website🔗: https://utnet.org/

*Utility Wallet

🔗https://utnet.org/solutions/utility_wallet

*Utility (AI deployment) services

🔗https://utnet.org/solutions/ai_model

*Utility Container Cloud Services

🔗https://cloud.utlab.io/admin/#/?redirect=%2Findex

*Utility Faucet

🔗https://faucet.utlab.io/

*Block Explorer: Under development

🎬 Utility explanation video: To be updated

📣 Twitter: https://twitter.com/UtilityNet_