How to Build a Decentralized Computing Power Platform with io.net

Background

With the launch of OpenAI's GPT-4 LLM, the potential of various AI Text-to-Image models has been witnessed, and the increasing applications based on mature AI models have led to a soaring demand for computing resources such as GPUs.
An article from GPU Utils in 2023 discussing the supply and demand situation of Nvidia H100 GPUs pointed out that large enterprises venturing into AI have a strong demand for GPUs. Tech giants like Meta, Tesla, and Google have purchased large quantities of Nvidia GPUs to build AI-focused data centers. Meta owns about 21,000 A100 GPUs, Tesla has about 7,000 A100s, and Google has also made significant GPU investments in its data centers, although specific numbers were not provided. Driven by the demand for training large language models (LLMs) and other AI applications, the demand for GPUs (especially H100s) continues to grow.
At the same time, according to data from Statista, the AI market size grew from $134.8 billion in 2022 to $241.8 billion in 2023, and it is expected to reach $738.7 billion by 2030. The market value of cloud services also increased by about 14% from $633 billion, with a portion of this attributed to the rapid growth in AI market demand for GPU computing power.
For the rapidly growing AI market with immense potential, what angles can we explore to deconstruct and uncover related investment entry points? Based on a report from IBM, we summarized the infrastructure needed for creating and deploying AI applications and solutions. It can be said that AI infrastructure primarily exists to handle and optimize the vast datasets and computing resources required for training models, addressing issues of dataset processing efficiency, model reliability, and application scalability from both hardware and software perspectives.

The substantial computing resources required for training AI models and applications prefer low-latency cloud environments and GPU computing power, and the software stack also includes distributed computing platforms (Apache Spark/Hadoop). Spark distributes the workflows that need to be processed across large computing clusters and has built-in parallel mechanisms and fault tolerance designs. The inherently decentralized design of blockchain makes distributed nodes the norm, and the POW consensus mechanism established by BTC requires miners to compete for block results through computing power (workload), which has a similar workflow to AI needing computing power to generate models/inference problems. Thus, traditional cloud server providers have begun to expand new business models, such as renting out GPUs like servers, selling computing power. Mimicking the blockchain approach, AI computing power adopts a distributed system design, utilizing idle GPU resources to lower the computing costs for startups.

IO.NET Project Introduction

Io.net is a distributed computing power provider that combines the Solana blockchain, aiming to leverage distributed computing resources (GPU & CPU) to address the computational demand challenges in AI and machine learning. IO integrates idle GPUs from independent data centers and cryptocurrency miners, collaborating with crypto projects like Filecoin/Render, gathering resources from over one million GPUs to solve the shortage of AI computing resources.
On the technical side, io.net is built on the machine learning framework ray.io for distributed computing, providing distributed computing resources for AI applications ranging from reinforcement learning, deep learning to model tuning and model execution. Anyone can join io's computing network as a worker or developer without additional permission, while the network adjusts computing power prices based on the complexity of computational tasks, urgency, and supply of computing resources, pricing according to market dynamics. Based on the characteristics of distributed computing power, io's backend also pairs GPU providers with developers based on GPU demand types, current availability, requesters' locations, and reputations.
$IO is the native token of the io.net system, serving as a medium of exchange between computing power providers and computing power service purchasers. Using $IO can reduce the order fee by 2% compared to $USDC. At the same time, $IO plays an important incentive role in ensuring the normal operation of the network: $IO token holders can stake a certain amount of $IO to nodes, and nodes can only earn corresponding rewards during idle machine periods if they have $IO tokens staked.
The current market value of the $IO token is approximately $360 million, with an FDV of about $3 billion.

IO Token Economics

The maximum total supply of $IO is 800 million, of which 500 million were allocated during the token TGE, and the remaining 300 million tokens will be gradually released over 20 years (with a monthly release decrease of 1.02%, approximately 12% per year). The current circulation of IO is 95 million, consisting of 75 million unlocked for ecological development and community building during TGE and 20 million mining rewards from Binance Launchpool.
The reward distribution for computing power providers during the IO testnet is as follows:

• Season 1 (up to April 25) - 17,500,000 IO
• Season 2 (May 1 - May 31) - 7,500,000 IO
• Season 3 (June 1 - June 30) - 5,000,000 IO

In addition to the testnet computing rewards, IO also provided some airdrops to creators participating in community building:

• (First round) Community / Content Creators / Galxe / Discord - 7,500,000 IO
• Season 3 (June 1 - June 30) Discord and Galxe participants - 2,500,000 IO

The Season 1 testnet computing rewards and the first round of community creation/Galxe rewards were completed during the TGE.
According to official documentation, the overall distribution of $IO is as follows:

$IO Token Destruction Mechanism

Io.net executes the buyback and destruction of $IO tokens according to a fixed preset program, with the specific buyback and destruction quantities depending on the $IO price at the time of execution. The funds used for buying back $IO come from the operational revenue of IOG (The Internet of GPUs - GPU Internet), collecting a 0.25% order reservation fee from both computing power purchasers and providers in IOG, as well as a 2% fee for purchasing computing power using $USDC.

Competitive Analysis

Similar projects to io.net include Akash, Nosana, OctaSpace, Clore.AI, etc., which focus on solving the computational demands of AI models in a decentralized computing power market.

• Akash Network utilizes a decentralized market model to aggregate and rent out excess computing capacity by leveraging idle distributed computing resources, addressing supply-demand imbalances through dynamic discounts and incentive mechanisms, and achieving efficient, trustless resource allocation based on smart contracts, thus providing secure, cost-effective, and decentralized cloud computing services. It allows Ethereum miners and other users with underutilized GPU resources to lease these resources, creating a cloud service market where service pricing is conducted through a reverse auction mechanism, allowing buyers to bid for resource rentals, driving competitive price reductions.
• Nosana is a decentralized computing power market project within the Solana ecosystem, primarily aimed at forming a GPU grid using idle computing resources to meet the computational demands of AI inference. The project defines its computing power market operations through programs on Solana and ensures that GPU nodes participating in the network complete tasks reasonably. Currently, it provides computing power services for the inference processes of LLama 2 and Stable Diffusion models during its second phase of testnet operations.
• OctaSpace is an open-source, scalable distributed computing cloud node infrastructure that allows access to distributed computing, data storage, services, VPNs, etc. OctaSpace includes CPU and GPU computing power, serving disk space for ML tasks, AI tools, image processing, and rendering scenes using Blender, among others. Launched in 2022, OctaSpace operates on its own Layer 1 EVM-compatible blockchain. This blockchain employs a dual-chain system, combining proof of work (PoW) and proof of authority (PoA) consensus mechanisms.
• Clore.AI is a distributed GPU supercomputing platform that allows users to access high-end GPU computing resources from nodes providing computing power globally. It supports various uses such as AI training, cryptocurrency mining, and film rendering. The platform offers low-cost, high-performance GPU services, and users can earn Clore tokens by renting GPUs. Clore.ai emphasizes security, complies with European laws, and provides robust APIs for seamless integration. In terms of project quality, Clore.AI's website is relatively rough, lacking detailed technical documentation to verify the authenticity of the project's self-introduction and data, leading us to remain skeptical about the project's GPU resources and actual participation levels.

Compared to other products in the decentralized computing power market, io.net is currently the only project that allows anyone to join and provide computing resources without entry barriers. Users can contribute computing power to the network using consumer-grade GPUs starting from the 30 series, as well as resources from Apple chips like Macbook M2 and Mac Mini. The more abundant GPU and CPU resources and rich API construction enable IO to support various AI computing demands, such as batch inference, parallel training, hyperparameter tuning, and reinforcement learning. Its backend infrastructure consists of a series of modular layers, allowing for effective resource management and automated pricing. Other distributed computing power market projects often collaborate on GPU resources aimed at enterprises, presenting certain barriers for user participation. Therefore, IO may possess the ability to leverage token economics to unlock more GPU resources.
Here is the current market value/FDV comparison between io.net and its competitors:

Review and Conclusion

The launch of $IO on Binance marks a fitting conclusion to a project that has drawn significant attention from the start, with a bustling testnet and gradually facing attacks and doubts regarding the transparency of its scoring rules during the delay of practical testing. The token launched during a market correction, opening low and rising high, ultimately returning to a relatively rational valuation range. However, for testnet participants drawn by io.net's strong investment lineup, some were pleased while others were disheartened. Most users who rented GPUs but did not persist in participating in every season of the testnet did not achieve the desired excess returns, instead facing the reality of "reverse pulling." During the testnet, io.net divided each season's prize pool into two pools for GPU and high-performance CPU to calculate separately. In Season 1, due to a hacking incident, the announcement of scores was delayed, but the final score exchange ratio for the GPU pool was determined to be nearly 90:1, with the costs for users renting GPUs from major cloud platform providers far exceeding the airdrop benefits. During Season 2, the official fully implemented the PoW verification mechanism, with nearly 30,000 GPU devices successfully participating and passing PoW verification, resulting in a final score exchange ratio of 100:1.
After such a highly anticipated start, whether io.net can achieve its stated goal of providing various computational demands for AI applications and how much real demand remains after the testnet may only be proven by time.
References:
https://docs.io.net/docs
https://blockcrunch.substack.com/p/rndr-akt-ionet-the-complete-guide
https://www.odaily.news/post/5194118
https://www.theblockbeats.info/news/53690
https://www.binance.com/en/research/projects/ionet
https://www.ibm.com/topics/ai-infrastructure
https://gpus.llm-utils.org/nvidia-h100-gpus-supply-and-demand/
https://www.statista.com/statistics/941835/artificial-intelligence-market-size-revenue-comparisons/
https://www.grandviewresearch.com/press-release/global-cloud-ai-market