1.5 Key AI Use Cases Across Industries
Key Takeaways
Blueprint objective 1.5 requires you to name representative AI use cases and the industries that run them, and objective 1.6 expects you to attach the right NVIDIA platform to each.
Five horizontal workload families cover almost every vertical example: computer vision, natural language and LLMs, speech and conversational AI, recommendation systems, and anomaly or fraud detection.
Healthcare uses medical imaging, genomics, and clinical documentation; NVIDIA aligns Clara for healthcare workflows and Holoscan for real-time medical-device sensor streams.
Financial services concentrate on fraud detection, risk modelling, and document processing, where RAPIDS accelerates the tabular feature engineering that dominates the pipeline.
Manufacturing, retail, telecom, automotive, and energy map to Omniverse and Isaac, Metropolis and Merlin, Aerial, DRIVE, and Earth-2 respectively.
1.5 Key AI Use Cases Across Industries
Blueprint anchor: Domain 1 objective 1.5 — "Explain the key AI use cases and industries." The NVIDIA course that maps to this objective opens with a unit titled AI Transformation Across Industries, so expect scenario stems that describe a business problem in a named vertical and ask which workload type or which NVIDIA platform fits.
1. Start With the Five Horizontal Workload Families
Verticals are long lists; workload families are short. Almost every industry example on the exam reduces to one of five patterns, and each pattern has a distinct infrastructure signature you can reason about.
| Workload Family | What It Does | Infrastructure Signature | Typical NVIDIA Software |
|---|---|---|---|
| Computer vision | Classification, detection, segmentation, OCR, video analytics | High ingest bandwidth; many concurrent streams; often edge-deployed; benefits from hardware video decoders (NVDEC) | Metropolis / DeepStream, TAO, TensorRT |
| Natural language & LLMs | Summarisation, chat, code generation, document understanding, RAG | Very large model weights; memory-bandwidth-bound decoding; KV-cache pressure; needs NVLink for tensor parallelism | NeMo, TensorRT-LLM, NIM |
| Speech & conversational AI | Speech recognition, translation, text-to-speech, voice agents | Strict real-time latency budgets; small models, many concurrent sessions — an ideal MIG or fractional-GPU workload | Riva |
| Recommenders & personalisation | Ranking, next-best-offer, search relevance, ad targeting | Enormous sparse embedding tables that exceed GPU memory; heavy host-memory and I/O traffic | Merlin, RAPIDS |
| Anomaly & fraud detection | Payment fraud, intrusion detection, predictive maintenance | Streaming, latency-sensitive scoring on tabular or telemetry data | RAPIDS, Morpheus |
If you can classify the scenario into one of these five, the correct infrastructure answer usually follows without needing vertical-specific knowledge.
2. Industry Mapping
| Industry | Representative Use Cases | Why Acceleration Is Required | Associated NVIDIA Platform |
|---|---|---|---|
| Healthcare & life sciences | Medical image reconstruction and triage, genomic variant calling, protein and molecule modelling, ambient clinical documentation | 3D volumetric data and genomic pipelines are enormous; device-side inference must run inside a procedure's real-time window | Clara (healthcare workflows), Holoscan (real-time medical-device sensor streaming), BioNeMo (biomolecular models), Parabricks (genomics) |
| Financial services | Card-fraud scoring, anti-money-laundering, credit and market risk, algorithmic trading, claims and document processing | Fraud decisions must land in milliseconds inside the authorisation path; risk simulation is compute-bound | RAPIDS (accelerated dataframes and ML), Morpheus (cybersecurity/anomaly), NeMo for document AI |
| Retail & e-commerce | Product recommendation, demand forecasting, dynamic pricing, shrink and loss prevention, conversational shopping | Recommender embedding tables and real-time video analytics both exceed CPU throughput at store or catalogue scale | Merlin (recommenders), Metropolis (in-store vision) |
| Manufacturing & industrial | Automated optical inspection, predictive maintenance, robotics, factory digital twins | Defect detection runs at line speed; simulation and digital twins are graphics- and physics-heavy | Metropolis (inspection), Isaac (robotics), Omniverse (digital twins), PhysicsNeMo/Modulus (physics-ML surrogates) |
| Telecommunications | RAN signal processing, network optimisation and fault prediction, customer-care assistants | Software-defined RAN requires deterministic, sub-millisecond signal processing | Aerial (accelerated RAN), NeMo/Riva for care agents |
| Automotive & transport | Autonomous-driving perception and planning, in-cabin monitoring, fleet simulation and validation | Perception must be real time and safety-certified; validation requires replaying millions of simulated miles | DRIVE (in-vehicle AV platform), Omniverse for simulation |
| Energy & climate | Seismic imaging, reservoir and grid simulation, weather and climate forecasting | Physics simulation is the classic HPC workload; AI surrogates cut turnaround from hours to seconds | Earth-2 (climate/weather), PhysicsNeMo |
| Public sector & smart cities | Traffic and transit analytics, public-safety video, citizen services | Thousands of concurrent camera streams need distributed decode plus inference | Metropolis, NIM-packaged assistants |
| Logistics & supply chain | Route and fleet optimisation, warehouse robotics, demand planning | Combinatorial optimisation grows factorially; GPU solvers make near-real-time re-planning feasible | cuOpt (route optimisation), Isaac |
Answering tactic: vertical names in a stem are usually flavour. Ask yourself which of the five families is this? — a "hospital reading 3D scans" question is a computer-vision question with an edge-latency constraint, and a "bank scoring transactions inline" question is an anomaly-detection question with a millisecond budget.
3. Where the Workload Runs: Data Center, Cloud, and Edge
The same use case demands very different infrastructure depending on where inference happens, and the exam does test this distinction.
- Data center / cloud training. Multi-GPU, multi-node, NVLink plus InfiniBand, scheduled through Slurm, Kubernetes, or Run:ai. Latency is irrelevant; throughput and checkpoint durability dominate.
- Data center inference. Triton or NIM behind a load balancer, often on MIG partitions so that one physical GPU serves several models with hardware-guaranteed isolation.
- Edge inference. Retail cameras, factory lines, hospital devices, and vehicles run on compact accelerated platforms (Jetson-class modules, IGX for functional-safety industrial and medical use). Constraints flip: power envelope, physical security, intermittent connectivity, and remote fleet management matter more than raw FLOPS.
A useful rule for scenario questions: training gravitates to where the data already lives; inference gravitates to where the decision must be made. A retailer trains a recommender centrally on months of transaction history, then serves shelf-camera inference in the store because backhauling raw video is neither affordable nor fast enough.
4. Generative AI as a Cross-Industry Layer
Generative AI does not replace the five families — it sits on top of them and appears in every vertical in roughly the same three shapes:
- Assistants and copilots grounded in enterprise documents through RAG (customer care, clinical notes, field-service manuals).
- Content and synthetic-data generation (marketing assets, and synthetic training images for defect classes that are too rare to collect).
- Agentic workflows where a model plans and calls tools or APIs rather than only producing text.
All three inherit the same infrastructure profile as the LLM family: large weights, memory-bandwidth-bound decoding, and a strong preference for NVLink-connected GPUs when the model does not fit on one device.
A hospital wants an AI system that processes a live endoscopy video feed and highlights suspected lesions inside the procedure's real-time window. Which NVIDIA platform is purpose-built for real-time sensor streaming on medical devices?
NVIDIA Merlin
NVIDIA Holoscan
NVIDIA cuOpt
NVIDIA Aerial
A retailer's recommendation engine has embedding tables far larger than the memory of any single GPU. Which characteristic of this workload family should drive the infrastructure design?
It is compute-bound on FP64 arithmetic and needs double-precision Tensor Cores.
It requires the strictest possible per-request latency and should be deployed on edge devices only.
It is dominated by very large sparse embedding lookups, so host memory capacity and I/O bandwidth between host and GPU matter as much as raw FLOPS.
It never benefits from GPU acceleration and should remain entirely on CPUs.
A manufacturer wants a physically accurate virtual replica of a production line to test layout changes and train robots before touching the real factory. Which NVIDIA platform matches this use case?
NVIDIA Omniverse
NVIDIA Morpheus
NVIDIA Parabricks
NVIDIA Riva
Sections you finish are checked off in the contents.