Author: LFG Labs
Original publication date: June 3, 2024
Original link: https://www.chaincatcher.com/article/2127297
If the core background of io.net is "grassroots", what do you think?
With $30 million in financing and the attention of top investors such as Hack VC, Multicoin Capital, Delphi Digital, and Solana Lab, it does not seem so "down-to-earth" at all, especially when coupled with the labels of GPU computing power/AI revolution, which are all synonymous with high-end.
However, amidst the clamor of community discussions, key clues are often overlooked, especially the profound changes that io.net may bring to the global computing network. Unlike the "elite" positioning of AWS, Azure, and GCP, io.net is essentially taking a civilian route:
It supplements the neglected "waist + long tail" computing power needs by gathering idle GPU resources, creating an enterprise-level, decentralized distributed computing network, empowering AI innovation for a wider range of small and medium-sized users with more incremental/existing computing power resources, and realizing the "re-liberation of productivity" of global AI innovation with low cost and high flexibility.
Behind the AI wave, the neglected undercurrent of computing power production relations
What are the core productivity resources of this round of AI wave and the future digital economy era?
Without a doubt, computing power.
According to Precedence Research, the global AI hardware market is expected to grow at a compound annual growth rate (CAGR) of 24.3% to exceed $473.53 billion by 2033.
Even if we put aside the forecast data, from the perspective of the logic of incremental and stock, we can clearly see that in the future development of the computing power market, there will be two major contradictions that are destined to exist for a long time:
In terms of increment, the exponential growth of computing power demand is bound to be much greater than the linear growth of computing power supply.
In terms of stock, the computing power is "cut to the top" under the head effect, and the mid-tier and long-tail players are left with nothing to eat. However, a large number of distributed GPU resources are idle, resulting in a serious mismatch between supply and demand.
Incremental dimension: computing power demand far exceeds supply
The first is the incremental dimension. In addition to the rapid expansion of the AIGC large model, countless AI scenarios in the early stages of an explosion, such as medical care, education, and smart driving, are being rapidly rolled out. All of these require massive computing resources. Therefore, the current market gap in GPU computing resources will not only continue to exist, but will also continue to expand.
In other words, from the perspective of supply and demand, in the foreseeable future, the market demand for computing power will definitely be far greater than the supply, and the demand curve will still show an exponential upward trend in the short term.
On the supply side, due to the constraints of physical laws and actual production factors, whether it is process technology improvement or large-scale factory construction to expand production capacity, it can only achieve linear growth at best. This also means that the computing power bottleneck in the development of AI is doomed to exist for a long time.
Stock dimension: Serious mismatch between supply and demand of mid-tier and long-tail players
At the same time, with limited computing resources and facing serious growth bottlenecks, Amazon Web Services (AWS), Microsoft Azure, and Google Cloud Platform (GCP) together account for more than 60% of the cloud computing market share, which is clearly a seller's market.
They hoard high-performance GPU chips and monopolize a large amount of computing power resources. The small and medium-sized computing power demanders in the middle and long tail not only have no bargaining power, but also have to face many problems such as high capital costs, KYC entry barriers, and unfair leasing terms. In addition, traditional cloud service giants, due to profit ratio considerations, inevitably ignore the differentiated business demands of "mid-level + long-tail" users (such as shorter, more immediate, smaller leasing needs, etc.).
But in fact, outside the computing power network of cloud service giants, a large amount of GPU computing power is idle and unused. For example, there are hundreds of thousands of third-party independent Internet data centers (IDCs) around the world that have small training tasks and waste resources. There are even massive amounts of idle computing power in crypto mining farms and crypto projects such as Filecoin, Render, and Aethir.
According to official estimates by io.net, the idle rate of IDC graphics cards in the United States alone is as high as over 60%. This has created an ironic paradox of supply and demand mismatch: more than half of the computing power resources of tens of thousands of small and medium-sized data centers, crypto mining farms and other operators are wasted on a daily basis and cannot generate effective revenue, but mid-level and long-tail AI entrepreneurs are enduring the high-cost, high-threshold computing power services of cloud giants, and even more diverse innovation needs cannot be met.
Those who are thirsty will die of drought, and those who are flooded will die of floods. With these two basic premises clarified, we can actually see at a glance the core contradiction between the current global AI development and the global computing power market - on the one hand, AI innovations are everywhere, and the demand for computing power is constantly expanding; on the other hand, a large number of "waist + long tail" computing power demands and idle GPU resources cannot be effectively met, and are outside the current computing power market.
This problem is not only the contradiction between the growing computing power demand of AI entrepreneurs and the lagging computing power growth, but also the contradiction between the vast number of "mid-level + long-tail" AI entrepreneurs, computing power operators and the unbalanced and insufficient supply and demand mismatch, so it is far beyond the solution capabilities of centralized cloud service providers.
For this reason, the market needs new solutions. Imagine if these operators with computing power can flexibly choose to rent out computing power when idle, can they obtain a computing cluster similar to AWS at a low cost?
It is important to know that building a new data network with such large computing power is extremely expensive. This has given rise to a computing power matching platform specifically for idle computing power resources in the middle and tail ends and small and medium-sized AI entrepreneurs, in order to mobilize these scattered idle computing power resources and specifically match them with small and medium-sized model training and large models in medical, legal, financial and other segmented scenarios.
It can not only meet the diversified computing power needs of the middle and tail, but also will be a dislocated supplement to the existing computing power service pattern dominated by centralized cloud giants:
Cloud service giants with massive computing resources are responsible for "urgent, difficult and dangerous needs" such as large model training and high-performance computing;
Decentralized cloud computing markets such as io.net are responsible for more diversified "flexible and low-cost needs" such as small and medium-sized model calculations, large model fine-tuning, and reasoning deployment;
In fact, it is to provide a more inclusive dynamic balance supply and demand curve between cost-effectiveness and computing power quality, which is more in line with the economic logic of market optimization of resource allocation.
Therefore, distributed computing networks such as io.net are essentially a solution that integrates "AI+Crypto", that is, using a distributed collaboration framework combined with basic economic means of token incentives to meet the needs of the mid- and tail-end AI market, which has huge potential but is in an exiled state. It allows small and medium-sized AI teams to customize their combinations and purchase the required GPU computing services that large clouds cannot provide, thereby achieving the "re-liberation of productivity" in the global computing market and AI development.
To put it bluntly, io.net is not a direct competitor of AWS, Azure, and GCP. Instead, it is a "complementary comrade-in-arms" who works with them to optimize the allocation of global computing resources and jointly expand the market pie. It is just a front in charge of different levels of "cost-effectiveness & computing power quality" needs.
It is not even ruled out that io.net will recreate a market share that is no less than that of the current top three cloud giants by aggregating "mid-level + long-tail" supply and demand players.
io.net: A matching trading platform for global GPU computing power
Precisely because io.net is based on Web3 distributed collaboration + token incentives to reshape the production relations of the mid- and tail-end computing power markets, we can actually see the shadow of the sharing economy such as Uber and Didi in it, that is, a matching and trading platform for GPU computing power similar to Uber and Didi.
As we all know, before the emergence of Uber and Didi, the "call and get a taxi" experience in a broad sense did not exist for users, because a number of private cars formed a huge and disorderly network of idle vehicles. If you wanted to take a taxi, you could only wave and wait on the roadside, or apply for dispatch from the corresponding taxi center company in each city. It was time-consuming, highly uncertain, and a seller-dominated market, which was not friendly to most ordinary people.
This is actually a true reflection of the supply and demand sides of the entire computing power market. As mentioned above, the small and medium-sized computing power demanders in the middle and long tail not only have no bargaining power, but also have to face many problems such as high capital costs, KYC entry barriers, and unfair leasing terms.
Specifically, how does io.net achieve its positioning as a "global GPU computing power distribution center + matching market", or in other words, what kind of system architecture and functional services are needed to help mid- and long-tail users obtain computing power resources?
Flexible and low-cost matching platform
The biggest attribute of io.net is that it is a light-asset computing power matching platform.
That is to say, like Uber and Didi, it does not involve the actual operation of heavy assets such as GPU hardware, which is extremely risky. Instead, it connects the supply of mid- and long-tail retail computing power (many of which are regarded as second-class computing power in large clouds such as AWS), and through matchmaking, it activates the previously idle computing power resources (private cars) and the mid- and tail AI needs that are in urgent need of computing power (taxi riders).
One end of io.net is connected to thousands of idle GPUs (private cars) in small and medium-sized IDCs, mining farms, encryption projects, etc., and the other end is connected to the computing power needs of hundreds of millions of small and medium-sized companies (taxi riders). Then io.net acts as a matching platform for intermediate scheduling, just like a broker matching countless buy and sell orders one by one.
This helps entrepreneurs train more personalized small and medium-sized AI models by gathering idle computing power at low cost and in a more flexible deployment configuration form, greatly improving resource utilization. The advantages are obvious. Regardless of whether the market is overheated or overcooled, as long as there is a mismatch of resources, the demand for platforms to achieve matchmaking is the strongest:
On the supply side, suppliers of idle computing power resources such as small and medium-sized IDCs, mining farms, and crypto projects only need to connect with io.net. They do not need to set up a dedicated BD department, nor are they forced to sell their computing power at a discount to AWS and other companies due to their small scale. Instead, they can directly match their idle computing power to suitable small and medium-sized computing power customers at market prices or even higher prices at extremely low friction costs, thereby gaining profits.
On the demand side, small and medium-sized computing power demanders who originally had no bargaining power in front of large clouds such as AWS can also use the io.net resource pipeline to connect to smaller-scale computing power that does not require permission, waiting, KYC, and has more flexible deployment time. They can freely select and combine the chips they need to form a "cluster" and complete personalized computing tasks.
The suppliers and demanders of computing power in the middle and tail end have similar pain points such as weak bargaining power and low autonomy in front of large clouds such as AWS. io.net activates the supply and demand of the middle and long tail ends and provides such a matching platform, allowing both supply and demand parties to complete transactions with better prices and more flexible configurations than large clouds such as AWS.
From this perspective, similar to platforms such as Taobao, the emergence of inferior computing power in the early stage is also an inevitable development law of the platform economy. io.net has also set up a reputation system for both suppliers and demanders, accumulating points based on computing performance and network participation to obtain rewards or discounts.
Decentralized GPU Clusters
Secondly, although io.net is a matching platform between retail supply and demand parties, current computing scenarios such as large models require several graphics cards to perform calculations together - it not only depends on how much idle GPU resources your matching platform can aggregate, but also on how closely the connections between the distributed computing power on the platform are.
In other words, this distributed network that covers small and medium-sized computing power in different regions and of different scales needs to implement a "decentralized but centralized" computing architecture: according to the flexible computing needs of different scenarios, several distributed GPUs can be put into the same framework for training, and ensure that the communication and collaboration between different GPUs are very fast, at least to achieve low latency and other characteristics that are sufficient for use.
This is completely different from the dilemma that some decentralized cloud computing projects are limited to the use of GPUs in the same computer room. The technical implementation behind it involves the "three pillars" of the io.net product portfolio: IO Cloud, IO Worker, and IO Explorer.
The basic business module of IO Cloud is Clusters, which is a group of GPUs that can self-coordinate to complete computing tasks. AI engineers can customize the desired cluster according to their needs. It is also seamlessly integrated with IO-SDK to provide a comprehensive solution for expanding AI and Python applications.
IO Worker provides a user-friendly UI interface that allows both supply and demand parties to effectively manage their supply operations on a web application, including functions related to user account management, computing activity monitoring, real-time data display, temperature and power consumption tracking, installation assistance, wallet management, security measures, and profit calculation;
IO Explorer mainly provides users with comprehensive statistics and visualizations of various aspects of the GPU cloud. It enables users to easily monitor, analyze and understand the details of the io.net network by providing complete visibility into network activity, important statistics, data points, and reward transactions;
Because of the above functional architecture, io.net allows computing power suppliers to easily share idle computing resources, greatly lowering the entry threshold. Demanders do not need to sign long-term contracts or endure the long waiting time common in traditional cloud services. They can quickly build a cluster with the required GPUs and obtain services such as super computing power and optimized server response.
Lightweight elastic demand scenarios
To be more specific, when talking about the misaligned service scenarios of large clouds such as io.net and AWS, they are mainly concentrated in some lightweight elastic demands that are not cost-effective for large clouds. Among them, the diversified scenarios that everyone can think of, such as segmented field model training for small and medium-sized AI startup projects, large model fine-tuning, etc., are all included.
In addition, there is another commonly used scenario that is easily overlooked: model reasoning.
As we all know, the early training of large models such as GPT requires the use of tens of thousands of high-performance GPUs, super computing power, and massive data for long-term high-quality calculations. This is also the absolute advantage area of large clouds such as AWS and GCP.
But after training, the main computing power demand becomes the long-term model reasoning. The computing power demand at this stage is also much higher than that of the training stage. Reasoning based on the trained model, that is, the daily conversation and interaction scenarios between ordinary users and models such as GPT, accounts for 80%-90% of AI computing share.
Interestingly, the overall computing power of the reasoning process is more stable, and it may only take dozens of GPUs a few minutes to get the answer, and the requirements for network latency and concurrency are lower. At the same time, most AI companies may not train their own large models separately, but only choose to optimize and fine-tune a few large models such as GPT. These scenarios are naturally suitable for io.net's distributed idle computing resources.
In addition to the minority of high-intensity and high-standard application scenarios, the broader, daily lightweight scenarios are also a virgin land that needs to be developed. It looks fragmented, but the market share is even larger - according to the latest report from Bank of America, high-performance computing accounts for only a small part of the total available market (TAM) of data centers, only about 5%.
In short, it’s not that AWS, GCP, etc. are unaffordable, but that io.net is more cost-effective.
Web2 BD’s winning hand
Of course, in the end, the core competitiveness of platforms such as io.net that are oriented to distributed computing resources still lies in BD capabilities, which is the key to success or failure.
In addition to the strange phenomenon of graphics card brokers spawned by Nvidia's high-performance chips, the biggest problem that plagues many small and medium-sized IDC and other computing power operators is that "even good wine needs no bush."
So from this perspective, io.net actually has a unique competitive advantage that is difficult to replicate in the same field - it has a Web2 BD team directly based in Silicon Valley. They are all veterans who have been immersed in the computing power market business for many years. They not only understand the diversified scenarios of small and medium-sized customers, but also grasp the terminal needs of a large number of Web2 customers.
According to official disclosure by io.net, 20 to 30 Web2 companies have expressed their willingness to purchase/lease computing power. They are willing to try or make mistakes because of the lower cost and more flexible computing power services (some may not be able to wait for computing power on AWS at all). For trial and error, a single customer needs at least hundreds or thousands of graphics cards (equivalent to hundreds of thousands of US dollars in computing power orders per month).
This real willingness to pay on the demand side will essentially attract more idle computing resources to actively flow in on the supply side, making it easier to break through the circle of Web2&Web3 and form a first-mover network scale effect.
Tokenized computing power ecological settlement layer may
As mentioned above, io.net is based on Web3 distributed collaboration + token incentives to reshape the mid- and tail-end computing power market, the latter of which mainly adopts the dual-token model of IO and IOSD:
Token IO, the utility includes paying computing power rental fees, giving IO Worker allocation incentives, rewarding AI and ML deployment teams for continued use of the network, balancing some demand and supply, pricing IO Worker computing units, and community governance;
The stablecoin IOSD, which is pegged to the US dollar and can only be obtained by destroying IO, is designed to provide a stable value storage and transaction medium for the io.net platform;
In addition, io.net is also considering supporting suppliers to increase the probability of being rented by pledging IO. Demanders can also pledge IO to give priority to the use of high-performance GPUs, thereby allowing the development of a complete ecosystem around the pledge function to capture the incremental value generated by the entire computing power ecosystem.
This actually brings up another question. Since io.net has gathered a huge amount of idle computing power resources, can it go a step further and directly combine Crypto's computing power tokenization method to give GPUs greater on-chain financial possibilities?
For example, in the future, it is entirely possible that io.net will build a dedicated computing power chain based on a huge computing power network. By providing permissionless and barrier-free tokenized infrastructure services, anyone and any device will be able to directly tokenize computing power (for example, converting A100 and H100 into standardized tokens or NFTs), thereby allowing transactions, staking, borrowing, lending, and leverage.
This is equivalent to creating a vast on-chain GPU computing power market for users, where users and funds from all over the world can enter freely and efficiently. We can simply imagine two scenarios to get a glimpse of the imagination space that the on-chain financial market centered on computing power will have in the future.
1. Security-type computing power token
For example, if a computing power operator on io.net owns several A100 or H100 graphics cards, but he needs funds or wants to cash in early, he can package the computing power value corresponding to these graphics cards into an NFT or FT on io.net - the Token represents the discounted computing power cash flow of the corresponding graphics card in the next year (or a certain period of time), which can be priced in IO.
Since most ordinary investors do not have the opportunity to directly purchase, hold, and operate AI computing power, this type of token provides market traders with an opportunity to gamble on the rise and fall of computing power prices in the future. Operators with a large amount of computing power but tight cash flow also gain financial leverage and can flexibly realize liquidity anytime and anywhere according to actual needs.
During this period, the graphics card behind the token will be operated by io.net, and the subsequent cash flow earned by the corresponding computing power will be divided proportionally (Token holders receive 0.9 and operating nodes receive 0.1).
And because it is a standardized token, it can also be freely circulated and traded in CEX or DEX like other tokens. This will further form real-time computing power pricing with free entry and exit, and truly transform GPU computing power into a resource with global liquidity.
2. Bond-type computing power token
In addition, you can also raise funds to purchase high-performance graphics cards and increase network computing power by issuing bond-type tokens. The bond principal is equivalent to the value of the graphics card device itself, and the bond interest is the cash flow income from leasing the graphics card computing power in the future, which means that the potential rental value of the graphics card's computing power and future income are the market value of the token. Holding the token can obtain real RWA income.
This is equivalent to creating a vast GPU computing power market for global users. Users and funds from all over the world can enter the GPU computing power market freely and efficiently without worrying about high thresholds and high funds. It will further fully connect real graphics cards and various decentralized financial products, laying the foundation for more and richer supporting services for users in the future.
More importantly, the entire process uses IO as the main transaction/circulation currency, making io.net/IO likely to become the settlement layer/settlement currency of the entire global computing power ecosystem. The vision of the on-chain financial market centered on the tokenization of computing power can almost recreate a valuation space similar to the io.net decentralized computing power network narrative.
summary
In general, Web3, as a new type of production relationship, is naturally adapted to AI, which represents a new type of productivity. This is also a simultaneous advancement of technology and production relationship capabilities. From this perspective, the core logic of io.net is to change the production relationship between traditional cloud service giants, medium and long-tail computing power users, and global idle network computing resources by adopting the economic infrastructure of "Web3+token economy":
Provide solutions to the real pain points of AI computing power supply and demand, build a two-sided market that encompasses and serves "waist + long tail" GPU computing power resources & user needs, optimize the supply and allocation of computing power resources, and bring about a great liberation of productivity in global AI development, especially small and medium-sized AI innovations.
The vision is undoubtedly grand. If successful, it will most likely become the core matching infrastructure and value settlement layer of the global GPU computing power ecosystem. It is expected to obtain the best valuation premium and has great potential for imagination, but it is undoubtedly full of challenges.
Original link: https://www.chaincatcher.com/article/2127297
The io.net Binance Square community reprints the original article. The copyright and content responsibility of the original article belong to the original author. The reprinting of io.net and Binance Square does not represent certification or support for part or all of the views of the reprinted content.