One of the biggest challenges when it comes to training artificial intelligence (AI) models is the cost of the graphical processor units (GPUs) employed. Most organizations wind up using GPUs accessed via a cloud service because the cost of acquiring GPUs can be prohibitive.
However, Monster API this week launched an alternative decentralized approach that enables organizations to leverage under-utilized GPU resources that can be found everywhere from individual desktop systems and gaming systems to servers in on-premises IT environments and Tesla vehicles at one-tenth the cost of a cloud service.
In addition to providing access to GPUs, Monster API provides access to a set of application programming interfaces (APIs) to access an orchestration framework based on Kubernetes, a containerized implementation of the Compute Unified Device Architecture (CUDA) software developed by NVIDIA and a set of inference engines it makes available.
Fresh off raising $1.1 million in seed funding, Monster API CEO Saurabh Vij said the goal is to sharply reduce the cost of AI experimentation via a subscription service. Developers can readily access the resources required to build and deploy AI models in a matter of minutes, he adds. Developers are also fine tuning AI models such as Stable Diffusion, Whisper AI, StableLM using their own data, says Vij.
Monster API said it can already provide access to more than 30,000 GPUs, including enabling consumption of those resources in a way that dynamically scales from one to 100 across multiple geographies. That approach also enables Monster API to provide billing based on the number of API calls to make billing for consumption of those resources more predictable.
There are obviously GPUs that are under utilized almost everywhere. The challenge is finding a way to harness them at distributed scale. The decentralized approach being advocated by Monster API in many ways is conceptually similar to how many organizations are starting to harness compute resources to drive blockchain applications, notes Vij. “The future is decentralized,” he says.
It’s not clear to what degree large enterprises are comfortable with any form of decentralized computing, but many individual developers and data scientists building AI models are starving for GPU resources. A lot of potentially innovative applications might never be developed simply because the cost of accessing GPUs in the cloud is prohibitive. In the absence of any alternative the only organizations capable of building AI models will be technology companies and large enterprises, says Vij.
Of course, there’s a lot more that goes into building and maintaining AI models than access to GPU infrastructure but it’s already apparent there needs to be a way to democratize the building of AI models in a way that doesn’t require expensive cloud resources. The challenge and the opportunity now is making it possible for almost anyone to build an AI model for hundreds of thousands of use cases that are not likely to be addressed by a major technology company any time soon.