Multicoin Capital: Why We Invested in io.net

Author: Shayon Sengupta, Multicoin Capital
Compiled by: JIN, Techub News
On June 6, 2024, Binance announced that Launchpool will launch the io.net token IO. Users can stake BNB and FDUSD in the IO mining pool on the Launchpool website starting at 8 AM HKT on June 7 to earn IO rewards. The mining period will last for a total of 4 days. The website is expected to update approximately five hours after this announcement, prior to the start of the mining activity.

Additionally, Binance will list the io.net token IO at 8 PM HKT on June 11, opening trading markets for IO/BTC, IO/USDT, IO/BNB, IO/FDUSD, and IO/TRY.

IO Token Unlocking and Rewards

According to the official documentation from io.net, the total supply of IO tokens is 800 million, with 500 million IO released at launch and the remaining 300 million gradually issued over the next 20 years until the cap of 800 million tokens is reached. The initial 500 million supply unlocking and rewards are shown in the figure below, divided into five categories: seed investors, Series A investors, core contributors, research and development, and ecosystem and community;

IO Token Unlocking and Rewards

Estimated Distribution of IO Tokens

• Seed Investors: 12.5%

• Series A Investors: 10.2%

• Core Contributors: 11.3%

• Research and Development: 16%

• Ecosystem and Community: 50%

Estimated Distribution of IO Tokens

The following introduction to io.net was written by Multicoin Capital, a participant in the previous io.net $30 million Series A financing:

We are excited to announce our investment in io.net, a distributed network providing AI computing power rental services. We not only led the seed round but also participated in the Series A financing. io.net has raised a total of $30 million, with participants including Multicoin, Hack VC, 6th Man Ventures, Modular Capital, and a consortium of angel investors, aiming to build an on-demand, always-available market for AI computing power.

I first met io.net's founder Ahmad Shadid at the Austin Hacker House during the Solana hackathon in April 2023, and I was immediately drawn to his unique insights into the decentralization of ML (machine learning) computing infrastructure.

Since then, the io.net team has demonstrated strong execution capabilities. Today, the network has aggregated tens of thousands of distributed GPUs and provided over 57,000 hours of compute time for AI companies. We are thrilled to partner with them to support the AI renaissance over the next decade.

1. Global Computing Power Shortage

The demand for AI computing is growing at an astonishing rate; this demand is currently unmet. In 2023, data center revenues providing computing power for AI demands exceeded $100 billion, but even in the most conservative scenarios, the demand for AI exceeds chip supply.

During periods of high interest rates and cash flow shortages, new data centers capable of accommodating such hardware require significant upfront investment. The core issue lies in the limited production of advanced chips like the NVidia A100 and H100. While GPU performance continues to improve and costs steadily decrease, the manufacturing process cannot be accelerated due to shortages of raw materials, components, and capacity.

Despite the promising future of AI, the physical space required to support its operation is increasing daily, leading to a significant rise in demand for space, power, and cutting-edge equipment. io.net opens a pathway for us, where computing power is no longer constrained by these limitations.

io.net is a classic example of DePIN applied in the real world: by using token incentives to structurally lower the cost of acquiring supply-side resources, it reduces costs for end GPU computing power users. It aggregates idle GPU resources scattered around the globe into a shared pool for AI developers and companies to use. Today, the network is supported by thousands of GPUs from data centers, mining farms, and consumer-grade devices.

While it is possible to consolidate these valuable resources, they do not automatically scale to a distributed network. Throughout the history of cryptocurrency technology, there have been several attempts to build distributed GPU computing networks, all of which failed due to not meeting the demands of the demand side.

Coordinating and scheduling computing power across heterogeneous hardware with different memory, bandwidth, and storage configurations is a key step in achieving a distributed GPU network. We believe the io.net team has the most practical solution in today's market to make this hardware aggregation useful and economically viable for end customers.

2. Paving the Way for Clusters

In the history of computer development, software frameworks and design patterns self-adjust around the hardware configurations available in the market. Most frameworks and libraries used for AI development heavily rely on centralized hardware resources, but in the past decade, distributed computing infrastructure has made significant progress in practical applications.

io.net leverages existing idle hardware resources by deploying custom networks and orchestration layers to connect them, creating a super-scalable GPU internet. This network utilizes Ray, Ludwig, Kubernetes, and various other open-source distributed computing frameworks, enabling machine learning engineering and operations teams to scale their workloads on the existing GPU network.

ML teams can parallelize workloads on io.net GPUs by launching clusters of computing devices and utilize these libraries to handle orchestration, scheduling, fault tolerance, and scaling. For example, if a group of graphic designers contributes their home GPUs to the network, io.net can build a cluster designed so that image model developers around the world can rent collective computing resources.

BC8.ai is an example, a fine-tuned stable diffusion variant model that was entirely trained on the io.net network. The io.net browser displays real-time inference and incentives for network contributors.

AI Supercomputer

The generation information for each image is recorded on-chain. All fees are paid to 6 RTX 4090 clusters, which are consumer-grade GPUs used for gaming.

Today, there are tens of thousands of devices on the network, spread across mining farms, underutilized data centers, and Render Network consumer nodes. In addition to creating new GPU supply, io.net is also able to compete on cost with traditional cloud service providers, often offering cheaper resources.

They achieve cost reductions by outsourcing GPU coordination and operations to decentralized protocols. On the other hand, cloud service providers mark up their products due to employee expenses, hardware maintenance, and data center operational costs. The costs of consumer-grade GPU clusters and mining farms are far lower than what hyperscalers are willing to accept, creating a structural arbitrage that allows resource pricing on io.net to be dynamically lower than the rising cloud service rates.

3. Building the GPU Internet

io.net has a unique advantage in maintaining a light asset operation and reducing the marginal cost of serving any specific customer to nearly zero, while directly establishing relationships with both demand and supply sides of the market, capable of serving thousands of users who need access to GPUs to build competitive AI products that everyone will interact with in the future.