Computing Power Storm: Decoding the Future Wave of Computing

Author: Iris Chen, Dr. Nick

1. Demand and Challenges Coexist

The "2022-2023 Global Computing Power Index Assessment Report" points out that against the backdrop of slowing global GDP growth, the digital economy continues to maintain a strong growth momentum. The proportion of the digital economy in the GDP of major countries around the world is rising year by year, with the overall proportion of sample countries expected to increase from 50.2% in 2022 to 54.0% in 2026. Computing power has gradually become a major driving force for economic growth. For every 1-point increase in the computing power index, the digital economy and GDP of a country will grow by 3.6‰ and 1.7‰, respectively. More critically, when the computing power index exceeds 40 points, each additional point will drive GDP growth 1.3 times more than when below 40 points, and when exceeding 60 points, it can even reach 3 times. The advantages of regions that are early adopters of computing power will be strengthened with the increase in the proportion of computing power investment, further widening the gap with latecomer regions, highlighting the importance of developing computing power.

1. The AIGC Wave is Coming, Huge Demand in the Computing Power Industry

With the application and development of key technologies such as artificial intelligence, blockchain, the Internet of Things, and AR/VR, the demand for computing power will increase in the future. It is expected that by 2030:

• Artificial Intelligence: Deeply penetrating all industries, requiring 16,000 EFLOPS of computing power (equivalent to embedding 160 billion Qualcomm Snapdragon 855 NPUs in smartphones).

• Blockchain: Supporting areas such as cryptocurrency, requiring 5,500 EFLOPS of computing power (equivalent to 1.3 billion AntMiner V9s).

• Internet of Things: Connecting all devices in factories and homes, requiring 8,500 EFLOPS of computing power (equivalent to using 7.9 billion chips in high-end IoT edge devices).

• Space computing/AR/VR/Metaverse: When fully leveraged, requiring 3,900 EFLOPS of computing power (equivalent to 2.1 billion SONY PS4 consoles).

At the same time, the explosive popularity of ChatGPT in 2022 has triggered a wave of AIGC, further increasing the demand for computing power. In the GPT series released by OpenAI, GPT-3 is a language model composed of 175 billion parameters, while GPT-4's parameters have reached the trillion level. As the parameter count of large models continues to grow, the computing power required to train an AI model will increase by 275 times every two years. This will push the global AI computing market scale growth to new heights. IDC predicts that by 2026, the global AI computing market will reach $34.66 billion, with the generative AI computing market growing from $820 million in 2022 to $10.99 billion in 2026, increasing its share of the AI computing market from 4.2% to 31.7%. Under this development trend, the future demand for computing power is enormous.

2. Security Threat Costs Difficult to Reduce, Challenges Facing the Computing Power Industry

(1) Security: Flexible access to computing power networks, numerous distributed resource nodes

• The computing power network consists of five main parts: computing network service layer, computing network scheduling layer, computing power center, edge computing/user center, and computing power carrying network. However, this architecture, while providing efficient and flexible computing services, also brings a series of security challenges:

• The computing power network features ubiquitous computing power and flexible access, and frequent resource connections will increase the attack exposure of resources.

• A vast amount of data, involving confidential privacy, circulates within the computing power network. Tampering or leaking during transmission can lead to severe consequences.

• Computing power services are end-to-end, with a large user base and numerous distributed resource nodes, making data information management complex and evidence tracing difficult.

• The new architecture of the computing network introduces new network elements such as computing network perception units and computing network control units, increasing management complexity.

(2) Cost: GPU Supply Shortage, Severe Idle Computing Power

With the prosperity of AI, the demand for GPUs has surged. Currently, most of the GPU market is dominated by NVIDIA, but the supply of NVIDIA chips is tight, and prices have skyrocketed. For example, the market price of the A100 GPU has reached 150,000 yuan, with an increase of over 50% in two months. At the same time, the application of large models will further increase computing power costs. It has been calculated that 10,000 NVIDIA A100 chips are the threshold for achieving AI large models, with the single training cost of GPT-3 exceeding $12 million.

At the same time, there is a problem of idle computing power with GPUs. Training a model with 175 billion parameters for GPT-3 requires storing over 1TB of data in memory, which exceeds the capacity of any existing GPU today. Limited by memory, more GPUs are needed for parallel computing and storage, leading to low GPU utilization and idle computing power. Moreover, due to memory limitations, the relationship between model complexity and the number of GPUs required is not linear, exacerbating the issue of low GPU utilization. GPT-4, trained on approximately 25,000 A100 GPUs for 90 to 100 days, had a computing power utilization rate of only 32% to 36%. Additionally, much computing power exists in independent data centers, cryptocurrency miners, and consumer devices like MacBooks and gaming PCs, making it difficult to aggregate and utilize these resources.

With the vigorous development of computing power, electricity demand will also grow rapidly. It is expected that the global electricity demand for data centers from 2023 to 2027 will be 430-748 terawatt-hours, equivalent to 2-4% of global electricity demand from 2024 to 2027, posing challenges to electricity infrastructure. Morgan Stanley predicts that under the baseline scenario of increasing GPU utilization from 60% to 70%, the total electricity capacity of global data centers will reach 70-122 gigawatts from 2023 to 2027, with a compound annual growth rate of 20%. Specifically:

• In a bull market scenario (90% chip utilization): The global electricity demand for data centers from 2023 to 2027 is expected to be 446-820 terawatt-hours.

• In a bear market scenario (50% chip utilization): The global electricity demand for data centers from 2023 to 2027 is expected to be 415-677 terawatt-hours.

Therefore, companies that can meet the rapidly growing electricity demand for computing power will benefit from this trend, especially those power solution providers that can reduce power supply delays for data centers.

2. Development Trends and Project Introduction

1. Decentralized Computing Provides Secure and Low-Cost Computing Solutions for Web 3

The essence of Web 1 is union, where web pages are "read-only," and users can only search and browse information; the essence of Web 2 is interaction, where websites are "readable and writable," and users are not just recipients of content but can also participate in content creation. Web 3 represents an era of interconnectedness, where websites are "readable, writable, and ownable," and users have ownership and control over the digital content they create, with the ability to choose to enter agreements with others for distribution. As a representative of the next generation of the internet, Web 3 emphasizes decentralization, openness, and user sovereignty. Decentralized computing, distinct from traditional cloud computing, effectively meets the computing demands driven by modern technologies and has become the core of Web 3 infrastructure. With the development of new internet technologies and the further expansion of data volume, the decentralized application market has broad prospects. According to Zhiyan Consulting, the global decentralized application market is expected to reach $1,185.54 billion by 2025.

In the face of security, cost, and electricity challenges in the computing power industry, building a decentralized distributed computing power network is an important direction for the development of AI infrastructure. Decentralized computing provides a secure, low-cost, and uninterrupted computing solution for various applications in the Web 3 ecosystem by comprehensively utilizing existing computing resources through leasing, sharing, and scheduling of computing power. Compared to traditional centralized systems, the specific advantages of decentralized computing are as follows:

》Security

• All participants have processing capabilities. If one participant is threatened, others can respond.

• Allows for distributed control and decision-making. This helps ensure that no single entity can exert complete control over the internet or its users, making it less likely for users to be monitored or censored, thus enhancing online privacy and freedom of speech.

》Low Cost: Decentralized computing distributes costs and responsibilities across multiple entities, making it more affordable and sustainable in the long run. Currently, Web 3 decentralized computing platforms in the market can offer prices that are generally 80-90% lower than centralized computing platforms.

• Cheaper computing power. In traditional data centers, the cost composition includes servers (30%), housing (12%), networking (15%), AC (21%), power (17%), and labor (5%). Decentralized computing relies on users sharing resources and contributing computing power in a mutually beneficial manner, theoretically saving 70% of costs.

• Cheaper training costs. Decentralized computing allows for the parallel threading of thousands of serverless technologies, enabling GNN training to scale to billions of edge models. According to research from UCLA, for large models, decentralized computing can provide 2.75 times the performance of traditional systems per dollar, and for sparse large models, it is 1.22 times faster and 4.83 times cheaper.

• Cheaper deployment costs. Traditional AI solutions require significant investments in software development, infrastructure, and talent. Decentralized computing allows developers to leverage existing resources and infrastructure, making it easier to build and deploy AI applications. It also democratizes AI development, allowing users to share computing resources and collaborate on developing AI solutions.

• Infrastructure more suitable for AI. By reducing the costs of training and computing, decentralized computing has the potential to enable more organizations and individuals to use AI, driving growth and innovation across numerous industries.

》Uninterrupted Service: Decentralized network nodes are distributed, theoretically never going down, with no single point of failure.

Project Introduction

Akash Network: A decentralized cloud computing marketplace that allows users to securely and efficiently buy and sell computing resources. Unlike other decentralized platforms, users can run any cloud-native application on Akash-hosted containers without needing to rewrite the entire internet in a new proprietary language, and there is no vendor lock-in preventing switching cloud providers.

io.net: A decentralized computing network that allows machine learning engineers to access distributed cloud clusters at a lower cost than centralized services. It features products like IO workers, IO cloud, and IO browser, with a valuation of over $1 billion on Solana.

2. AI Drives High-Performance Computing, High-Performance Computing Empowers AI

High-performance computing refers to computing systems that use supercomputers and parallel computer clusters to solve advanced computing problems. Such systems are typically over a million times faster than the fastest desktop computers, laptops, or server systems, and have widespread applications in established and emerging fields such as autonomous vehicles, the Internet of Things, and precision agriculture.

High-performance computing accounts for only about 5% of the total available market in data centers, but with the rapid development of AI and the application of large models, the increase in AI and high-performance data analysis workloads is driving changes in HPC system design, and HPC is also empowering AI, with both driving each other's development. Global HPC spending was approximately $37 billion in 2022, and Hyperion predicts it will reach $52 billion by 2026, while the AI market empowered by HPC is expected to have a compound annual growth rate of 22.7% from 2020 to 2026.

Project Introduction

Arweave: The newly proposed AO protocol, which uses a non-Ethereum modular architecture, can achieve ultra-high-performance computing on a storage public chain, even achieving near-Web2 experiences, providing a solid new infrastructure for Web3 x AI.

iExec: A decentralized cloud computing platform providing high-performance computing services, allowing users to rent computing resources to perform compute-intensive tasks such as data analysis, simulation, and rendering.

CETI: Founded by the former CEO of crypto.com, targeting enterprise-level high-performance computing centers.

3. Turning Point in Human-Computer Interaction: Spatial Computing

Spatial computing refers to the use of AR/VR technology to integrate users' graphical interfaces into the real physical world, thereby changing human-computer interaction. The release of Microsoft's MR headset Hololens in 2015 was a milestone in modern spatial computing. Although it has not become widespread, it demonstrated the potential of spatial computing. This year, Apple's release of Vision Pro brought more precise spatial awareness technology and deeper user interaction experiences, pushing spatial computing into the spotlight.

In fact, we are reaching a turning point in human-computer interaction: shifting from traditional keyboard and mouse setups to touch gestures, conversational AI, and enhanced visual computing interactions at the edge. According to IDC's predictions, global VR device shipments are expected to reach 9.17 million units in 2023, a year-on-year increase of 7%, while AR device shipments will be 440,000 units, a year-on-year increase of 57%. It is expected that over the next four years, the VR market will grow at an annual rate of over 20%, while the AR market will exceed 70%. The development of AR/VR technology will significantly enhance the importance of spatial computing. Following PCs and smartphones, spatial computing has the potential to drive the next wave of disruptive change—making technology a part of our daily behavior, connecting our physical and digital lives with real-time data and communication.

Project Introduction

Clore.ai: A platform connecting tenants and users needing GPUs, aimed at allowing users to access powerful computing resources at competitive prices and flexible terms. Its powerful GPUs enable users to render movies at a professional level, significantly reducing the time required, and are compatible with various rendering engines, also usable for AI training and mining.

Render Network: A decentralized GPU rendering platform aimed at advancing the next generation of rendering and AI technology, allowing users to scale GPU rendering tasks on-demand to high-performance GPU nodes around the world.

4. Edge Computing Becomes an Important Supplement to Cloud Computing

Edge computing refers to processing data at physical locations closer to end devices, where the "edge" is within a maximum round-trip time of 20 milliseconds to the end user. Edge computing deploys computing resources closer to end devices, allowing data to be processed locally, thereby reducing the latency and network bandwidth pressure of transmitting data to the cloud for processing. Therefore, it has advantages in terms of latency, bandwidth, autonomy, and privacy.

Tech giants such as Facebook, Amazon, Microsoft, Google, and Apple are investing in edge computing and edge locations (from internal IT and OT to external, remote sites) to be closer to end users and data generation points. Bank of America predicts that by 2025, 75% of enterprise-generated data will be created and processed at the edge, and by 2028, the market size of edge computing will reach $404 billion, with a compound annual growth rate of 15% from 2022 to 2028.

Project Introduction

Aethir: A cloud computing infrastructure platform, launching Aethir Edge in April 2024 as Aethir's only authorized mining device, leading the development of decentralized edge computing and paving the way for the democratization of edge computing's future.

Theta Network: A decentralized video transmission service platform aimed at addressing bottlenecks such as high costs and low efficiency in existing video transmission systems. A hybrid cloud computing platform based on a fully edge architecture, Theta EdgeCloud, is planned for launch in the second quarter of 2024.

5. AI Training Expected to Fully Shift to AI Inference

In the trend of decentralization, AI training is not currently the best landing scenario for DePIN. The computing power requirements for AI production mainly revolve around AI inference and AI training. AI training refers to feeding a large amount of data to train a complex neural network model, while AI inference refers to using the trained model to infer various conclusions from large amounts of data. Therefore, the difficulty of combining decentralization with computing power decreases layer by layer from training to fine-tuning training to inference. If a decentralized computing application for GPT is built on Ethereum, a single matrix multiplication operation could consume gas fees as high as $10 billion and take a month, with the training cost for each token (approximately 750 words for 1000 tokens) typically around 6N (where N is the number of parameters in the large language model), while the inference cost is only about 2N, meaning the inference cost is roughly one-third of the training cost.

At the same time, compared to AI training, AI inference is more closely linked to the demand from large-scale application terminals such as consumer electronics. Counterpoint Research predicts that global PC market shipments will return to pre-pandemic levels in 2024, with AI PCs expected to grow at a compound annual growth rate of 50% from 2020 and dominate the PC market after 2026. With the emergence of new consumer electronics products that integrate AI, such as AI PCs and AI smartphones in 2024, the trend of large-scale applications of edge AI large models and AI software will become increasingly evident, which also means that the importance of AI inference is becoming more pronounced, becoming the core technology behind the efficient operation of large models and AI software at the edge. The focus of the AI industry is expected to shift from training to inference.

Project Introduction

Nosana: A blockchain-based distributed GPU resource sharing platform aimed at addressing the GPU shortage in the market. In 2023, it shifted to AI inference, launching a large-scale GPU computing grid for AI inference, which is a deliberate integration of blockchain technology into AI, making it an ideal tool for handling AI's demanding computational requirements.

Exabits: A decentralized AI and high-performance computing service platform aimed at building a fair, user-friendly, and inclusive AI ecosystem, providing affordable accelerated computing for AI model training and inference.