By Minhyun Kim, Founder and CEO of AI Network
With AI expected to add trillions of dollars to global economic growth by 2030, the question on everyone’s mind is how to deal with the uncertain supply of graphics processing unit (GPU) resources. As we know, GPUs are essential for training AI models and powering services like ChatGPT, but frequent supply issues that threaten the very evolution of the technology. It is a scenario that sends stakeholders scrambling to secure enough GPU resources and find ways to increase the efficiency of their AI models, to varying degrees of success.
One solution that often gets overlooked is tapping into the vast pool of unutilized GPUs—those owned by individuals and enterprises that are only used some of the time. There are enough GPUs out there to legitimately address the GPU resource bottleneck. But unlike the filesharing apps of the past, which leveraged unutilized storage resources without compensation, the only way to make this solution work is by incentivizing GPU owners to rent out their unutilized resources.
Potential of Unutilized GPUs
It is difficult to quantify an exact number of unutilized GPUs in existence, but it is likely in the single-digit billions if we factor in gamers and large enterprises. If we estimate somewhere between a 15-25% usage rate, and perhaps lower for gamers, we start to get an understanding of just how much computing power could potentially be unlocked.
Until now, however, this has not been a viable solution primarily due to the need for compensation. Filesharing apps worked, in my mind, because individuals did not require compensation for hosting files. They seemed content to contribute to the community with the expectation that they would also benefit from downloading content from other hosts. GPUs are different. Running one 24 hours a day is a lot more expensive. Like with cryptocurrency miners, there needs to be a pot of gold at the end of the rainbow for people to seriously consider “renting” their GPUs.
The Airbnb Model
If there was one thing Airbnb taught us at the beginning, it was the potential for monetization of an unutilized resource: your home when you were out of town. The idea that you could rent out your home while you were away allowed you to turn your domicile into a revenue-generating asset, even if only for a short period of time. And all handled via an app.
The same model can be used to turn GPUs into revenue-generating assets. When not in use, owners can connect their GPUs to a resource pool and begin accepting processing tasks such as AI model training. For each completed task, the owner is compensated based on market-determined rates for resources—similar to how apps like Uber calculate ride pricing. A tiered approach could ensure that owners with better hardware earn a higher rate.
It is this kind of model that will incentivize individuals and enterprises to make their unutilized GPU resources available for public consumption and, ultimately, maximize AI growth in the face of a resource shortage. The challenge is ensuring that it maximizes value for the GPU owners while keeping resource rates affordable and competitive relative to other GPU resource providers. In my mind, this can only be done using a decentralized approach.
Role of Blockchain
I do not wish to overstate its importance, but blockchain does three things that are needed for the above model to be viable with GPUs. The first is supporting decentralization. We know centralized systems are prone to high fees in order to cover overhead and reach revenue expectations. A decentralized system underpinned by blockchain, smart contracts, and digital assets reduces the need for overhead by abstracting decision-making to community decentralized autonomous organizations (DAOs) and automating processes.
The second is the ability to account for and compensate GPU usage. Blockchains are designed to be very good at this exact task. Each GPU that connects to the system can be represented by a non-fungible token (NFT) and each task can be a smart contract with financial terms and GPU resource ownership embedded. Every time a task is completed, the smart contract triggers an automatic payout to the owner(s) of the GPU resources. It is a highly automated and trustworthy system that has already been demonstrated successfully with decentralized storage and computing usage.
The third is cheap value transfer. Unlike Airbnb and Uber, the average individual will likely earn relatively small amounts per day. Cashing out frequently, therefore, would be prohibitively expensive via existing bank transfer services. Cryptocurrencies, on the other hand, offer a much cheaper alternative, not to mention access to decentralized financial services and the ability to keep revenue in stablecoins back by fiat currencies that are less prone to inflation.
In closing, we know that GPU resources have the potential to be a bottleneck for the evolution of AI. Producing more chips is the obvious solution, but we also need to consider alternatives, especially during times when chip production fails to meet demand. Incentivizing GPU owners to turn their GPUs into revenue-generating assets is one such alternative. If implemented in a decentralized way via blockchain, it could create a fair and open market for GPU resources while keeping fees negligible.
About the author:
Minhyun Kim is the CEO of Common Computer and co-founder of AI Network. He began his career as a software developer at Google and contributed to building Google’s key services for eight years. He then founded Common Computer Inc. in 2019 with the mission of making it into “The Internet for AI.” Successfully raising $19 million through several rounds of venture capital has set the company on a clear path to eventually provide the resources to launch open-source AI for everyone. Minhyun graduated from the Korea Advanced Institute of Science and Technology with a bachelor’s degree in computer science.
The views and opinions expressed herein are the views and opinions of the author and do not necessarily reflect those of Nasdaq, Inc.