Cloud Computing

Cloud computing is the on-demand delivery of computing resources, including servers, storage, databases, networking, GPUs, and complete software applications, over the internet from a provider's data centers, billed by how much is used. The US National Institute of Standards and Technology defined it in 2011 in NIST SP 800-145 through five essential characteristics: on-demand self-service, broad network access, resource pooling across many customers, rapid elasticity, and measured service. Amazon Web Services launched its S3 storage and EC2 compute services in 2006, and AWS, Microsoft Azure, and Google Cloud are now the largest providers. For artificial intelligence, the cloud turns expensive accelerators into a resource that can be rented for an afternoon, which is how most teams train large models without owning a data center.

Service and Deployment Models#

The service models differ in how much of the stack the provider manages:

• Infrastructure as a Service (IaaS): Virtual machines, block storage, and networks, with the customer managing the operating system, drivers, and software. GPU virtual machines such as Amazon EC2 P5 instances are IaaS.
• Platform as a Service (PaaS): Managed runtimes, databases, and Kubernetes clusters, where the customer deploys code and the provider handles servers and patching. Managed ML platforms such as Amazon SageMaker, Google Vertex AI, and Azure Machine Learning add training jobs and model endpoints on top.
• Serverless computing: Functions or containers that start on each request, scale to zero when idle, and bill per request and execution time, a later model that hides servers entirely.
• Software as a Service (SaaS): Complete applications used in a browser or through an API, with nothing for the customer to operate.

NIST also names four deployment models: public clouds shared by many customers, private clouds run for one organization, community clouds shared by organizations with common requirements, and hybrid clouds that combine them. In every model, providers follow a shared responsibility model: the provider secures the physical data centers and the virtualization layer, and the customer is responsible for its identities, configurations, and data security.

Cloud Computing for AI#

Training is the most compute-hungry part of AI. Cloud providers rent single GPUs or clusters of NVIDIA H100, H200, and B200 accelerators joined by high-bandwidth interconnects for distributed training, and Google Cloud also offers its own TPUs. Spare capacity is cheaper: AWS Spot Instances cost up to 90% less than on-demand prices but can be reclaimed with a two-minute warning, so jobs that use them save checkpoints of their model weights often enough to resume.

Data placement shapes cloud costs. Datasets live in object storage, often organized as a data lake, and providers charge egress fees for data leaving their network while ingress is usually free, so training compute tends to run in the same region as the big data it reads. On the inference side, models packaged as Docker containers run on managed endpoints for model serving or on Kubernetes clusters that add or remove replicas as traffic changes, which is the rapid elasticity in NIST's definition, and regions close to users keep inference latency low. Cloud ML platforms also host the surrounding MLOps tooling for experiment tracking, model registries, and model monitoring.

Ultralytics Platform applies this rent-by-use model to YOLO training: cloud training offers 26 GPU types, from a 16 GB RTX 2000 Ada to a 288 GB B300, bills the GPU time a run actually uses, and, when the selected GPU is out of capacity, proposes the closest available one, matching VRAM first, then speed, then price.

Cloud Computing vs. Edge Computing#

Edge computing processes data on or near the devices that produce it, such as cameras, robots, and gateways, which avoids the network round trip to a data center and keeps working when the connection drops. Cloud computing centralizes elastic capacity for the jobs that need it most: training, large batch processing, long-term storage, and services whose users can tolerate tens of milliseconds of network delay. Most computer vision systems combine the two. A model trained on cloud GPUs is exported to a compact format such as ONNX or TensorRT and runs as edge AI on the device, and the device sends detections and selected images back to the cloud, where they become training data for the next model version.