AI Keeping a Lid on the Cost of Building and Deploying AI Applications

When it comes to building and deploying artificial intelligence (AI) applications it is quickly becoming apparent that organizations will need to rely heavily on graphical processor units (GPUs). The challenge is that GPU-based cloud services are expensive to consume and at any given point in time one cloud service provider might, from a pricing perspective, provide a better option over another.

Being able to deploy an AI application that depends on multiple GPU resources, however, requires an ability to manage application development and deployment spanning multiple clouds. Personal AI, a provider of a chat application based on a generative grounded transformer (GGT-P) engine, turned to CodeNow to achieve that goal using a platform that makes it practical to deliver AI applications that span multiple cloud services.

The CodeNow platform takes advantage of Kubernetes clusters running in multiple cloud computing environments to provide development teams with a consistent set of application programming interfaces (APIs) for deploying applications. Developers are also provided with insights into application runtime behavior via a set of observability tools that CodeNow has embedded into the platform, says Codenow CEO Petr Svoboda. “The focus needs to be on change management,” he says. “It’s all about developer productivity.”

Most AI applications are based on a microservices architecture that enables them to take advantage of containers running on Kubernetes clusters to better orchestrate the processing of data. Otherwise, the AI application would simply be too unwieldy to build, deploy and manage.

With the rise of AI, it’s become apparent there is a need to distribute the processing of data across multiple Kubernetes clusters. CodeNow is making a case for an approach to software delivery that gives organizations more pricing control over cloud services that can more easily be woven together.

Of course, each cloud service provider would prefer builders of AI applications to run all their software on their cloud, so they offer discounts based on the volume of infrastructure resources provided. The issue that organizations encounter is pricing of cloud services can be volatile so there are savings to be had by pitting one cloud service provider against another. The challenge has been that managing software delivery across multiple cloud services has historically been too challenging for many organizations to implement across platforms that had very different APIs. Kubernetes, however, presents a uniform set of APIs across multiple cloud services that significantly simplifies software delivery.

It’s too early to say how whether GPUs will become commoditized just like instances of virtual machines running on X86 processors already are. The current demand for GPUs exceeds supply so GPUs will remain comparatively expensive to invoke in the cloud. As such, however, that higher level of cost makes it more important during challenging economic times to maximize returns on investment (ROI) in GPUs whenever possible.

In the meantime, it’s apparent that a cultural divide has emerged between application developers and the data scientists that build AI models. Most data scientists have little to no expertise when it comes to building and deploying an application to if applications are to be infused with AI; it will come down to how simple it becomes for development teams to deploy those applications.