Avoiding GenAI Disillusionment to Make Magic in the Cloud

Research from Gartner shows investing in artificial intelligence has reached new heights, driven by a focus on generative AI (GenAI). Still, in many cases, the technology hasn’t produced the hoped-for business value. So, while AI may have weathered Gartner’s hype cycle, it’s about to enter a disillusioning period often marked by ROI-lowering performance. It’s no wonder the Harvard Business Reviews reports that 80% of AI projects fail.  

AI shortcomings are often caused by cloud infrastructure that just isn’t capable of powering GenAI research and development. With the right resources, you can cull value from unstructured data to strengthen decision-making, fortify product quality, bolster marketing, raise customer experiences and identify industry trends. The potential is endless -- as long as the organization’s environment is optimized for AI. 

Cloud Control 

Some believe cloud-based GenAI is not cost-effective because it’s less expensive to deploy the necessary high-end processing and networking on-premises. However, operating GenAI on-premises requires GPUs, which are both hard to find and expensive, and you need to run workloads 24x7 at a 90% resource utilization. This isn’t effective for organizations that want, and financially need, to develop incrementally. The cloud provides that controlled growth, and when it comes to unpredictable workloads, its elasticity offers a far better path. 

Another benefit of the cloud is the types of GenAI models being used: Open-source versus closed-source. While closed-source models resoundingly outperform open-source ones, they can’t be used on-premises -- for this you need the cloud. What’s more, the cloud offers a lower cost of entry, accompanied by a healthy community of managed services and expert partners for support.  

Structure and Support 

There are many ways organizations can ensure their computing and storage infrastructure cost-effectively supports its GenAI projects. For starters, they can modernize by fine-tuning applications for better performance, while ensuring files and metadata are in the right place so that cost-effective scaling can follow. 

Cleaning, combining, and consolidating large data sets from various sources is important. This makes data more usable, while generating stronger insights because a complete data set is analyzed, not just a subset. Organizations can also optimize so AI can cross-reference and validate information, resulting in a better quality of data, especially when positioned around collection and analytical resources. 

Correctly configuring compute and storage can prevent surprise costs. This requires understanding the size of a model in order to feed it into the correct GPU. And when it comes to the storage side, optimizing the input data size for the model can improve latency. However, GenAI apps and models do necessitate ongoing optimization and tuning.  

When it comes to cloud providers, note they’ll usually provide a number of evaluation models, helping customers to choose the right model and lower their testing costs. Also, don’t forget that cloud providers offer redeemable credits that can significantly lower the costs of computing services.  

Making Magic 

While many feel GenAI issues are a technology problem, they’re actually a business one. Organizations have to pinpoint the bottlenecks and breaks, and then use the correct tools to fix issues. Further, some feel they need to work through technical issues before finding a GenAI use case. Actually, this should be flipped: The use case should be determined first in order to gain a clear picture of goals and expected ROI.  

Not understanding the cause of your work, and the objective to meet, is what makes GenAI projects needlessly complex. Because every model and workload is different, it’s best to set output and performance benchmarks and then work backward. Start with a proof of concept including a minimum of 10 users to get the feedback rolling. Watch every input and output your GenAI creates and see how they stack up to those benchmarks.   

Finally, realize that there are experts and tools that can help, so you don’t have to do it by yourself. Managed service providers can have built-in security to keep inappropriate content from getting into your data set. Amazon and Google offer tools that can provide guard rails for GenAI projects. Additionally, there are consultancies with irreplaceable hands-on expertise that can create the optimal approach for an organization.  

With an infrastructure that is AI-ready, you’ll avoid disillusionment and be in a position to make some GenAI magic.