Four critical AI use cases built with GPUs
Author: David Immerman, Consulting Analyst, S&P Global Market Intelligence
Not every AI use case requires significant compute and power, and CPUs are sufficient for some. However, a growing set of workloads demands far more computing and power. In these cases, the underlying physics of the problem benefits from large-scale parallel processing, and GPUs often provide a more practical way to run them.
As part of the Agentic AI Matures research project, 451 Research by S&P Global interviewed 16 AI leaders across several industries. Those conversations surfaced hundreds of AI use cases spanning traditional machine learning, generative AI, LLMs and agentic workflows. Many require modest compute resources, but a smaller subset has greater needs — workloads where the physics of the problem benefit from parallel processing at scale. The following are four of the most compute-hungry use cases from those interviews, along with the AI infrastructure criteria each highlights.
Access the full AI Maturity Report 2026, commissioned by Vultr, along with a webinar from Vultr CMO Kevin Cochrane and 451 Research by S&P Global Chief Analyst Eric Hanselman analyzing the findings.
Use case 1: Real-time computer vision on the plant floor
A CIO at an aerospace and machinery manufacturer described deploying computer vision at the edge — cameras running AI inference to flag quality defects and safety violations as they happen. Alerts need to surface within seconds, so the compute must live close to the action.
"As we get more smart manufacturing in place, we are going to have more edge computing requirements, and you know, the GPUs are something that we may need to be considering for some... to enable some of those, like, computer vision type of activities on the plant floor for quality or whatnot." – CIO, aerospace machinery manufacturing, 1,001-5,000 employees
The infrastructure requirement is computing proximity to the physical site. Sending raw video frames to a distant data center introduces latency that makes real-time alerting impractical. Inference needs to happen close to where the work is — on hardware without virtualization overhead adding to the round trip — and ideally within the same jurisdictional boundary as the operational data.
Use case 2: Scientific R&D and high-resolution image analysis
An advanced materials company described using machine learning to analyze high-resolution imagery of EV battery cross-sections — each image is 5 gigabytes — to optimize battery chemistry and improve performance.
"So on each one of our batteries, images are 5 gigabytes each. So when we do machine learning and we look at the image, that's where the GPU is... that's where a lot of GPU is actually required to go." – Sr. director, IT security, industrial manufacturing, 500-1,000 employees
The bottleneck here tends to be memory bandwidth rather than raw compute clock speed. Multi-gigabyte tensors are well served by accelerators with large, high-throughput memory to avoid constant data movement stalling the pipeline. These workloads are also irregular — intensive during an R&D sprint and idle in between — which makes on-demand access to burst capacity far more economical than owning hardware that sits unused most of the year.
Use case 3: Large-scale model training and batch processing
Not every company building on AI has the capital budget to own a GPU cluster. A director at a financial regulatory technology firm — whose core business involves evaluating and auditing AI models — described a procurement reality that's widely shared but rarely discussed openly.
"For our batch processing and our model training, we can't afford to have a huge cluster of GPUs sitting in our data center. We just don't have the capital for that, so we use a combination of GPUaaS providers ... it's just much more cost-effective for us to rent it for a few hours a week than to own it." – Director, finance & insurance services RegTech, 11-50 employees
This is the GPU-as-a-service use case in its clearest form. The critical infrastructure criterion is cost per GPU-hour — a gap that compounds quickly when GPUs are only needed for a few hours a week. The secondary element is availability: reserved or on-demand access to large clusters without the capital commitment of ownership. For most firms at this scale, renting is not a compromise; it is often the more rational choice.
Use case 4: Quantum computing simulation
At the far end of the compute spectrum, a chief information officer from a large defense and technology organization described building a framework for quantum computing simulation — workloads so demanding that performance, not cost, is the dominant variable.
"Making sure that... our return of information is at least 10 times faster than the normal sequential. So, quantum computing... we are building a framework for quantum computing, so that's expensive to do." – CIO, technology/defense, 50,000+ employees
Quantum simulation on classical hardware is among the most GPU-intensive workloads that exist. Achieving the order-of-magnitude speedup this leader described is well-suited to raw GPU density — large clusters of high-end accelerators working in parallel, with high-bandwidth interconnects to minimize communication overhead between nodes. Cost is secondary to the ability to run the simulation at target performance.
What these workloads have in common
Each of these four use cases points to something different from the underlying infrastructure: compute proximity for computer vision, memory bandwidth for R&D imaging, low cost-per-GPU-hour for batch training and raw cluster density for simulation. Few run efficiently in a traditional general-purpose CPU environment, and most are poorly served by infrastructure designed for a different class of workload.
The leaders we spoke with are building these systems now, often by stitching together offerings from multiple providers because no single platform has fully met all their needs. The common thread: They are looking for the right GPU hardware, close enough to where the work needs to happen, at a cost structure that matches how intermittently or continuously the workloads actually run. However, computing is only part of the equation: The organizations getting real results pair the right infrastructure with clean data, disciplined model training, reliability controls, a clearly defined business outcome and the production processes to sustain performance at scale.
Access the 2026 AI Maturity Report and a webinar from Vultr and 451 Research by S&P Global here.
Copyright © 2026 S&P Global.
The content of this artifact is for educational purposes only. S&P Global does not endorse any companies, technologies, products, services, or solutions. Permission to reprint or distribute any content from this artifact requires the prior written approval of S&P Global.
