io.net is a network based on the Solana blockchain that allows users with idle computing power to provide computing power to resource-intensive artificial intelligence companies. Previously, io.net completed its Series A financing, with a valuation of $1 billion and raised $30 million, led by Hack VC, with participation from Multicoin Capital, Animoca Brands, Solana Ventures, Aptos, OKX Ventures, etc. In this episode of Wu Blockchain English Podcast, we invited Garrison Yang (0xHushey), Chief Strategy Officer and Chief Marketing Officer of io.net, to discuss io.net's technical advantages, future development plans, and how to solve the pain points of the artificial intelligence and cloud computing industries.
The content of this article only represents the views of the interviewees, does not represent the views of Wu Blockchain, and does not provide any financial advice. Readers are advised to strictly abide by local laws and regulations.
The text summary is processed by GPT and may contain errors. Please listen to the podcast for the full text:
Youtube: https://youtu.be/HSBBGT5Vqvg
What is io.net?
io.net is building the world’s largest AI computing network, and our goal is essentially to build a decentralized competitor to AWS (Amazon Web Services). The project began in 2020 when our founder and CEO sought more economical computing resources while developing quantitative models for algorithmic trading. Even in 2020, the cost of purchasing computing power from AWS and Azure proved to be too expensive. Therefore, he explored computing resources around the world, including independent data centers in Saudi Arabia and Asia Pacific, and at crypto miners in use. He interconnected these resources to obtain GPUs from various geographically distributed locations to form a large computing network. This move was successful. Initially, he used Ray, which was not widely known at the time because most people preferred Kubernetes, which was popularized by Amazon Web Services. He independently operated this network to support his trading model for two and a half years.
In 2023, Open AI launched ChatGPT, and suddenly the world turned its attention to AI, especially the huge computing power it requires. ChatGPT announced in 2023 that it was spending $700,000 per day to train their ChatGPT model. People also questioned how it could provide real-time inference - responses to queries without charging users. Ahmad recognized that Open AI used the Ray architecture to scale their computing network, which led to the beginning of our business. Ahmad realized that he had a resource that was highly sought after by the AI industry: the ability to scale decentralized and distributed nodes to provide computing power to AI companies at a lower cost, so he founded io.net. Notably, we were one of the winners of the Solana hackathon in April 2023. The io.net platform was officially released to the public in November 2023, and in just over four months, we have witnessed rapid progress.
Another reason for developing io.net, in addition to the desire to create the largest AI computing network, is that we want to ensure that a decentralized, community-driven and community-owned computing network exists globally. By providing a decentralized source of computing power, we aim to ensure that applications, data sources, AI models, and AI model reasoning remain globally accessible, censorship-free, and usable across borders in the future.
How does io.net solve the pain points of the AI and cloud computing industries?
AI has been in our lives for quite some time. The really new developments are consumer-facing AI and enterprise AI that is visible to consumers. For example, ChatGPT opens a lot of doors, just like general AI image and video platforms, which have become tools we use in our daily lives.
What most people don’t realize is that we’ve been using AI for a long time. Facebook uses AI to scan the pictures you post, Google uses AI to help power their search engine, and Siri uses AI to understand and respond to our queries. Now, as AI becomes more popular and more companies want to develop using AI, their costs have increased. Computing power has become one of the most important and increasingly scarce resources in the world. Today, AI relies on GPU computing instead of CPU computing, making GPU requirements very high.
The problem is that the supply of computing power is far from enough - the world's computing power demand is 2.5 times the total computing power that centralized cloud service providers like AWS, Azure, and Google Cloud can provide. Although these large companies have stockpiled many graphics cards, they do not have enough GPU computing power to serve everyone. Then, there is a lot of computing power that exists in independent data centers, crypto miners, and our consumer devices (such as MacBooks and gaming PCs). The problem is that these resources are difficult to aggregate and utilize.
What io.net does is aggregate GPU computing power from a variety of sources. Some sources are higher quality, more reliable, and more valuable, some are not, but we aggregate them all to provide AI companies with a cluster of computing power. The result is that AI companies can access this computing power at a price that is up to 90% cheaper than the current price of centralized cloud service providers, and it is faster. With io.net, there are no contracts to sign, no KYC, no long-term commitments, and you can choose where your nodes come from and what their speed is. This approach helps us solve the main pain points of the AI and cloud computing industries by providing a decentralized, cost-effective, and easily accessible computing platform.
Can io.net match enterprise-level GPUs for AI model training?
In short: absolutely no problem.
By aggregating multiple slightly lower-end graphics cards together, the same level of computing power can be provided at a lower cost. What io.net needs to solve is how to scale up these physical infrastructures.
Enterprise-class graphics cards can also be aggregated by us. Independent enterprise-class data centers may have five graphics cards here and ten graphics cards there, and usually these data centers will not coordinate with each other to serve a single customer. But if they put these resources on io.net, we can aggregate them together. This not only maximizes the utilization of these enterprise-class GPUs, but also enhances their monetization capabilities. In essence, io.net acts as a platform that facilitates the aggregation of computing power from different sources, creating a more efficient and cost-effective computing environment for AI model training.
Idle GPUs everywhere are just inefficiently utilized resources. We can pass them on to customers in need, and end customers can use them at a cheap price.
But I want to be clear that while our goal is to build a decentralized alternative to centralized cloud service providers, I don't expect it to completely replace existing infrastructure. The best comparison is to think about the grid and renewable energy. We are not replacing base generation (like nuclear power), which generates a lot of energy over a period of time, is stable, cheap, and efficient. But nuclear power, as a "base load power source", cannot be adjusted flexibly. That's why solar and batteries are also needed to supplement the grid.
The same thing with AWS, if an enterprise signs a five-year contract with AWS and pays for a certain amount of computing power, they will continue to pay during this period regardless of whether they use it or not. This is generally cost-effective to purchase fixed computing power from centralized cloud service providers. However, when you have demand spikes, io.net, which can be used on demand, is valuable. For example, if you want to run an experiment today, or you know that your load will surge in seven days or a month, you can purchase additional computing power from io.net in a more economical way. As a decentralized network, you don’t need contracts or KYC, you can pick and choose different types, different sources, different sizes, and customizable computing resources at will, it’s that simple. This provides enterprises with a lot of different options and gives people choice and pricing power.
What are io.net's challenges and strategies in balancing individual and corporate contributors?
There are two types of people who are our ideal providers. One is an enterprise data center with surplus resources, typically this is a high-quality supply resource with high-speed internet bandwidth that is best suited to serve enterprise customers first. The second type falls into the realm of crypto miners - someone who buys a dedicated rig to provide power or run a GPU for crypto mining or similar activities, typically with above-average connection speeds and specialized hardware.
We liken GPUs to houses as investment assets. You invest in a house to rent it out and make money, and the same is true for GPUs. Despite the expected depreciation due to newer chips, we have seen prices appreciate due to the current shortage. Due to high demand, an A100 GPU can now pay back its investment in about nine months. This is a very attractive ROI compared to what is expected from traditional crypto mining economics.
The above is just about stablecoins, which is direct income, not including tokens. When you introduce crypto economy and tokens, the incentives are even greater. This is similar to the validator economy in blockchain, where you earn tokens with variable value while also earning fixed income from customers.
In addition, there are other opportunities in the large decentralized network that io.net is building. For example, there will always be some computing nodes that are not capable of AI/ML related tasks, perhaps because the connection speed is not ideal, or they are simply not fully utilized, resulting in their "unit economic model" not working. Despite this, these devices still have great potential to become relatively high-end validators. They can also perform other types of work and provide different levels of computing services.
Because the unit economics of AI/ML are so strong, we chose to tackle challenges in this area first before expanding our computing network to broader uses. This strategy not only solves the problem at hand, but also lays the foundation for the future development of the network.
When it comes to opportunities to build blockchains, I don’t think it necessarily has to be our own chain. Perhaps in the future the SVM (Solana Virtual Machine) decides to offer an appchain, or all validators directly support Solana, these are all possible developments. Each scenario adds utility to io.net, especially if we can build an appchain similar to Solana, that would be extremely compelling because it means we have both a blockchain and a computing layer that can provide infrastructure for decentralized applications (DApps), thus achieving true decentralization.
Who do you think is suitable to use io.net?
On the demand side, io.net is currently mainly used for end-to-end AI/ML workloads, which are mainly divided into two categories of demand: model training and model reasoning.
The first is the demand for model training. Although model training currently uses GPUs located in the same location, model training does not have strict latency requirements like model inference. In fact, it is not necessary to use aggregated GPUs, that is, by using the io.net extension architecture, it takes longer to train, and ultimately achieves lower-cost model training. After all, cost is the most important part of model training.
The second is the demand for model reasoning. Cost and dynamic load are the two main issues, both of which are very suitable for io.net. On the one hand, the cost of model reasoning is very high, and as explained before, the cost of io.net is very low. On the other hand, the load demand for model reasoning tends to increase unpredictably and quickly. If you publish your ChatGPT model, you need to consider all unpredictable loads. But io.net is very flexible and you can use it on demand at any time. You can completely purchase basic services from centralized cloud service providers, and then let io.net handle load peaks. We provide enterprises with the flexibility to manage demand peaks in a cost-effective manner, and provide additional load computing power with decentralized, on-demand resources.
I think not many people are training their own models now, after all, it's expensive, Facebook, Apple, and Microsoft have spent millions to train models. People are more likely to directly use the basic models developed by large technology companies, and then adjust and infer the existing models. I think this is the direction of the industry: a small amount of model training, a certain amount of model adjustment, and a lot of model inference. So io.net focuses on model inference, which is the "last mile" for developers and is easier to monetize. We built a very popular product called BC8.AI on the platform, which has about 25,000 transactions a day, all of which are model inference.
From a broader perspective, we chose to start the project from AI/ML because of its strong unit economics, but the development prospects of io.net are not limited to being the computing layer of AI. We also want to become the largest decentralized general computing layer. With the development of technology and the market, there will be more opportunities, allowing us to use this huge decentralized network to support a variety of computing needs, and ultimately realize a more open and free digital world.
Which parts of the io.net stack are relevant to blockchain?
The permissions that govern permissionless transactions are on the blockchain. We basically allow you to join the network in a permissionless way. There is no human intermediary in matching supply and demand. You use, you pay, and you know that people will receive the money and will provide you with a computing queue. There is actually a proof here that the supply side promises to provide capacity for a certain time. If you rent for an hour, I promise to provide it for an hour. If I don't provide that capacity for the entire hour, I will be fined. The economics and the mechanism are very similar to validators in this regard.
We also store computational proofs on the chain, and have a running log like a ledger, so if you look at a model inference result, you can always track exactly which node provided the computation, when it provided the computation, how much computation was provided, and how much you paid for it.
In the future, this type of transparency and data storage will be very helpful in tracking how these computations were produced. You might look at an image and always be able to track who influenced it, who provided the computation, and what model it was inferred from. I think this type of transparency is going to be very important as the AI industry grows.
Additionally, we chose Solana for our blockchain needs because it is fast and cheap. Despite some recent scalability challenges, it has proven to be able to handle the transaction load we’ve seen. Our goal is to maintain as much of our core infrastructure on Solana as possible, leveraging its capacity and upcoming updates like Firedancer for even greater scalability.
How does the matching process work for buyers selecting compute clusters on io.net?
Buyers don't need to go into too much detail. They simply indicate the type of device they want, the speed of their connection, and their preferred location. This gives buyers all the information they need to determine how many GPUs they need based on the capacity they are looking for. As the industry evolves, when latency and capacity are no longer the only determining factors, I expect buyers will make more nuanced choices. They may choose certain locations for specific reasons, or choose between low-end and high-end devices based on economic value. Right now due to scarcity, if a buyer needs capacity, they will take whatever is available that meets their needs. They will check our network, check if there are enough 4090s available, and then choose and use them.
There is currently a market education process, and I believe that as decentralized cloud service providers grow and develop, industry demand will continue to develop and expand.
Another aspect is the management of low-quality nodes. This problem arises once a token airdrop or mining program starts. So we introduced "time score" and "reputation score". Each node on io.net has a reputation score, which allows customers to see how long a node remains available, its online time, and other performance indicators to help make decisions. We constantly ping each node, and if a node does not respond, it is considered unavailable. If it is unavailable, it will not receive rewards. The crypto-economic incentive is straightforward: if a node is available, it provides better service to the demand side, and is hired more frequently, earning more rewards. As long as the node remains available and performant when hired, the demand side will get the required computing power, a win-win situation.
How do you see io.net developing in the next five years?
We launched the coin on April 28th, and I am sure the world will change after that, of course the extent depends on its subsequent developments.
At present, everyone knows that computing power is very valuable, but as a consumer, you can't really trade and use it as a commodity except buying GPUs. What we envision is to build an ecosystem of products and services based on "computing power" as an accessible commodity. By creating a tokenized way for people to trade computing power and further build other things. Our goal is to turn $IO into a computing power currency. If computing power is digital oil, then it needs a petrodollar, and we are trying to build that petrodollar. We have created a market for it where people can trade, exchange assets for value, and use DeFi products built on it. In the future, when we enter the field of general computing, more things can happen because more industries can participate.
Next, we need to increase the supply side by 10 times, improve the demand side self-service experience, and expand use cases such as cloud gaming, image rendering, video rendering - these may not belong to traditional AI workloads. In addition, by expanding on both the supply and demand sides, allowing low-end graphics cards to meet low-end needs, we can further expand customer types to increase market share.
Right now, our daily turnover is about seven or eight thousand dollars, about $2.3 million annualized, and this is only four months. So we just need to continue to grow both the supply and demand sides of the market and get the flywheel spinning. I hope that in the next four or five years, the platform will become very large and move a lot of computing power to the network. Our goal is to gradually transition DePIN to the community. DePIN should always be a decentralized asset, and the different applications built on it, whether it is io.net or other protocols and services driven by DePIN, are completely separate. But the gradual decentralization of DePIN is very important.
What is the relationship between io.net, Render, and Filecoin?
io.net starts with having a UI, which makes the user experience better. Then we bring two things to Render, Filecoin and other DeFINs: network architecture and operating software, which operates the client running on the device. The network architecture is critical because not all DePINs can aggregate thousands of geographically dispersed GPUs into a single cluster. By placing GPUs on our DePIN, they can be accessed by other networks and used on demand, mainly AI.
Render excels at rendering images - that's what its network was built and tooled for. Filecoin, similarly, is built for storage. These companies excel at their respective functions, but we are better at serving AI workloads, and whether it's Render, Filecoin, or other DeFINs, we need to enable clients on hardware capable of satisfying those AI workloads, which can access our network and use our tools to make money for AI clients, after all, AI does have strong unit economics.
In the future, io.net does not intend to replace platforms like Render or Filecoin, but provides more possibilities for cooperation. If someone wants to use their computing power to serve Filecoin or projects like Render, they can do so freely using io.net as a platform.
When it comes to market competition, we have our own thoughts. The unit economics in this field are so good that it is inevitable that there will be competition. However, the advantage of building a market business is that we have a first-mover advantage. We are earlier, we are bigger, and we actually have a 4-5 year first-mover advantage, which will bring huge inertia. Our network architecture - such as deploying a fully geographically distributed cluster with more than 6,000 nodes - is also difficult for others to do. In addition, competitors have to overcome surging costs.
What are the unique technical advantages of io.net?
In fact, all the strategies for building DePINs are not much different - one key factor is how to aggregate GPUs that are physically separated from each other; another key factor is how many GPUs to aggregate, and the details include whether they must be the same model, the distance allowed between them, the size of the cluster, how much latency there is, and so on.
These may seem simple, but implementing these network capabilities and the tradeoffs they make are nontrivial. If someone copied our approach, forking Ray and offering Ray Cluster, they would find that they cannot guarantee the same latency as we do. And in a world where computing power is a commodity, competitiveness depends on small differences in latency, availability, and connection speeds.
This is where our technical barriers come in. It includes the network architecture and the construction of the orchestration layer, the ease of deployment for developers, the simplicity of connection for workers, and many UI/UX decisions.
In addition to the technical barriers, there is also a market strategy barrier: deciding whether to start in areas such as image rendering, video rendering, storage, AI workloads, simultaneous cloud gaming, or pixel streaming. These are not technical barriers in the traditional sense, but they are major strategic choices based on resource constraints.
How can users who like io.net participate?
If you are interested in io.net, there are several ways to get involved. For those of you who have consumer devices at home, such as a MacBook or a 4090 gaming PC, I encourage you to connect to the network and provide computing power. This is a great way to understand how decentralized physical infrastructure networks work, and they exist to optimize the world's inefficient resources.
For miners, io.net presents an exciting opportunity to contribute and benefit. If you are interested in the idea of GPUs as assets to earn yield, this is worth considering. GPUs are yield-producing assets that often provide faster returns than traditional investments like houses or cars. This is not investment advice, but it is an angle worth considering - the financialization of computing power. We will turn $IO into a digital petrodollar, support a complete ecosystem of financialized products, tools, and services, and build the entire DeFi ecosystem on top of it.
Developers can try to deploy clusters on io.net. It took me three minutes to deploy 6,000 nodes today. It was difficult for an ordinary person to obtain such a large scale of computing power in the past. This is a novel and fascinating experience.
Even if you don’t provide computing power directly, there are still many opportunities to participate. You can join our community discussions, such as on topics such as data availability, protecting censorship-resistant AI models, and inference. Please follow us on Twitter or join our Discord, our Discord is moderated by some very smart and wise thinkers, please join in the conversation, we will be here.
