Cloud platforms for AI/ML provide on-demand GPU compute and managed services for training and deploying machine learning models — offering instances with A100s, H100s, and other accelerators alongside managed ML platforms like SageMaker, Vertex AI, and Azure ML, enabling teams to scale AI workloads without owning hardware.

Why Cloud for AI/ML?

• No Capital Investment: Pay for GPUs as needed, no $40K H100 purchases.
• Elastic Scale: Scale from 0 to 1000 GPUs for training, back to 0.
• Managed Services: Training, serving, monitoring handled by platform.
• Latest Hardware: Access H100s, H200s as they release.
• Global Availability: Deploy close to users worldwide.

GPU Instance Comparison

High-End Training Instances:

Instance          | GPUs      | GPU Memory| $/hr (On-Demand)
------------------|-----------|-----------|------------------
AWS p5.48xlarge   | 8× H100   | 640 GB    | ~$98
GCP a3-megagpu-8g | 8× H100   | 640 GB    | ~$100
Azure ND H100 v5  | 8× H100   | 640 GB    | ~$98
Lambda Cloud 8xH100| 8× H100  | 640 GB    | ~$85

Inference Instances:

Instance          | GPUs      | GPU Memory| $/hr (On-Demand)
------------------|-----------|-----------|------------------
AWS g5.xlarge     | 1× A10G   | 24 GB     | ~$1.00
GCP g2-standard-4 | 1× L4     | 24 GB     | ~$0.70
Azure NC A100 v4  | 1× A100   | 80 GB     | ~$3.67
AWS inf2.xlarge   | 1× Inferentia2| 32 GB | ~$0.75

Cost Optimization

Spot/Preemptible Instances:

Type          | Discount | Risk            | Use For
--------------|----------|-----------------|------------------
Spot (AWS)    | 60-90%   | Interruption    | Training w/checkpoints
Preemptible   | 60-80%   | 24hr max        | Batch jobs
Spot Block    | 30-50%   | 1-6hr guaranteed| Short jobs

Reserved/Committed:

Commitment    | Discount | Best For
--------------|----------|------------------
1-year        | 30-40%   | Steady inference workloads
3-year        | 50-60%   | Long-term production
PAYG fallback | 0%       | Burst capacity

Managed ML Services

AWS SageMaker:

Component     | Purpose
--------------|----------------------------------
Studio        | IDE for ML development
Training      | Managed training jobs
Endpoints     | Model serving
Pipelines     | ML workflow orchestration
Ground Truth  | Data labeling

GCP Vertex AI:

Component      | Purpose
---------------|----------------------------------
Workbench      | Managed notebooks
Training       | Distributed training
Prediction     | Serving endpoints
Pipelines      | Kubeflow-based workflows
Feature Store  | ML feature management

Azure Machine Learning:

Component      | Purpose
---------------|----------------------------------
Designer       | Drag-and-drop ML
AutoML         | Automated model selection
Compute        | Managed clusters
Endpoints      | Deployment targets
MLflow         | Experiment tracking

Decision Framework

Use Case                  | Provider Strength
--------------------------|------------------
Existing AWS shop         | SageMaker
Google ecosystem          | Vertex AI
Microsoft shop            | Azure ML
Cost-sensitive            | Lambda, RunPod, Vast.ai
Simplest experience       | Replicate, Modal
Maximum control           | Raw GPU instances

Storage Options

Service        | Provider | Use Case           | Cost
---------------|----------|--------------------|---------
S3             | AWS      | Datasets, artifacts| $0.023/GB
GCS            | GCP      | Same               | $0.020/GB
Azure Blob     | Azure    | Same               | $0.018/GB
EFS/Filestore  | Various  | Shared model access| Higher
FSx for Lustre | AWS      | High-perf training | $0.14/GB/mo

Cloud Architecture for LLM Training

<svg viewBox="0 0 760 470" xmlns="http://www.w3.org/2000/svg" font-family="-apple-system,Segoe UI,Roboto,sans-serif"><rect width="760" height="470" fill="#0d1117"/><defs><marker id="arrow" viewBox="0 0 10 10" refX="8" refY="5" markerWidth="6" markerHeight="6" orient="auto-start-reverse"><path d="M0 0L10 5L0 10Z" fill="#60a5fa"/></marker><filter id="glow"><feGaussianBlur stdDeviation="7"/></filter></defs><text x="380" y="34" fill="#e6edf3" font-size="21" font-weight="700" text-anchor="middle">Cloud AI — Elastic Inference Across a Region</text><text x="380" y="56" fill="#8b98a5" font-size="13" text-anchor="middle">requests route to accelerator pools while queues, replicas, and telemetry absorb demand</text><path d="M54 190q0-42 46-42 18-49 73-31 37-42 83-6 50-9 60 40 39 6 39 45H54Z" fill="#10233b" stroke="#60a5fa" stroke-width="2"/><text x="184" y="173" fill="#93c5fd" font-size="12" font-weight="700" text-anchor="middle">cloud region</text><g transform="translate(72 218)"><circle cx="18" cy="18" r="15" fill="#1e3a5f" stroke="#60a5fa"/><circle cx="62" cy="18" r="15" fill="#1e3a5f" stroke="#60a5fa"/><circle cx="106" cy="18" r="15" fill="#1e3a5f" stroke="#60a5fa"/><text x="62" y="55" fill="#8b98a5" font-size="10" text-anchor="middle">API requests</text></g><path d="M198 236H268" stroke="#60a5fa" stroke-width="2" marker-end="url(#arrow)"/><g transform="translate(276 188)"><path d="M0 24h104v92H0Z" fill="#171421" stroke="#a78bfa"/><path d="M0 24L52 0l52 24" fill="#211936" stroke="#a78bfa"/><g fill="#c4b5fd"><circle cx="25" cy="55" r="5"/><circle cx="45" cy="55" r="5"/><circle cx="65" cy="55" r="5"/><circle cx="85" cy="55" r="5"/></g><text x="52" y="86" fill="#c4b5fd" font-size="11" text-anchor="middle">request queue</text><text x="52" y="104" fill="#8b98a5" font-size="9.5" text-anchor="middle">backpressure</text></g><path d="M382 246H428" stroke="#a78bfa" stroke-width="2" marker-end="url(#arrow)"/><g transform="translate(438 108)"><text x="116" y="0" fill="#6ee7b7" font-size="11" text-anchor="middle">accelerator replica pool</text><g fill="#123c35" stroke="#34d399"><rect x="0" y="18" width="105" height="78" rx="8"/><rect x="127" y="18" width="105" height="78" rx="8"/><rect x="0" y="116" width="105" height="78" rx="8"/><rect x="127" y="116" width="105" height="78" rx="8"/></g><g fill="#6ee7b7" font-size="11" text-anchor="middle"><text x="52" y="51">GPU 1</text><text x="179" y="51">GPU 2</text><text x="52" y="149">GPU 3</text><text x="179" y="149">GPU N</text></g><g fill="#8b98a5" font-size="9" text-anchor="middle"><text x="52" y="72">model shard</text><text x="179" y="72">model shard</text><text x="52" y="170">model shard</text><text x="179" y="170">model shard</text></g></g><path d="M555 310V353" stroke="#34d399" stroke-width="2" marker-end="url(#arrow)"/><g transform="translate(438 360)"><path d="M0 0h232" stroke="#3a4453"/><path d="M0 26l20-9 20 5 20-15 20 9 20-4 20-9 20 18 20-8 20 5 20-12 32 7" fill="none" stroke="#38bdf8" stroke-width="2"/><text x="116" y="48" fill="#8b98a5" font-size="10" text-anchor="middle">latency · errors · utilization · tokens/s</text></g><path d="M438 398C366 414 318 352 330 305" fill="none" stroke="#fbbf24" stroke-width="2" stroke-dasharray="6 4" marker-end="url(#arrow)"/><text x="358" y="396" fill="#fbbf24" font-size="10">autoscale</text><text x="380" y="452" fill="#6b7684" font-size="11.5" text-anchor="middle">Elasticity works only when admission control, model placement, warm capacity, and telemetry are designed together.</text></svg>

Quick Starts

AWS (Launch GPU instance):

aws ec2 run-instances \
  --image-id ami-xxx \
  --instance-type p4d.24xlarge \
  --key-name my-key

GCP (Create GPU instance):

gcloud compute instances create gpu-instance \
  --zone=us-central1-a \
  --machine-type=a2-highgpu-1g \
  --accelerator=type=nvidia-tesla-a100,count=1

Cloud platforms are the infrastructure foundation for AI at scale — providing the elastic GPU compute and managed services that enable teams to train frontier models and deploy production AI systems without massive capital investment.