How to Use CoreWeave GPU Cloud for AI Training: 2026 Guide
CoreWeave is a Kubernetes-native GPU cloud — meaning you interact with it using kubectl rather than a web console wizard. This gives developers tremendous power and flexibility, but requires a brief orientation if you're coming from AWS EC2 or Google Colab. This guide walks through every step from account creation to your first training job running.
💡 What You'll Need: A CoreWeave account (sign up at coreweave.com), kubectl installed locally, basic familiarity with Docker and Python/PyTorch. No Kubernetes expertise required — this guide covers everything.
Account and kubectl Setup
Create a CoreWeave Account
Go to coreweave.com → "Get Started". Enterprise sign-up requires a conversation with sales. Developer access is available with a credit card — you get $50 in free credits to start. Complete identity verification (usually takes a few hours).
Download kubeconfig
In the CoreWeave dashboard: Settings → API Access → Download kubeconfig. Save to ~/.kube/config. This file authenticates all your kubectl commands to CoreWeave's Kubernetes clusters.
Install and Verify kubectl
Install kubectl for your OS (kubernetes.io/docs/tasks/tools). Verify: kubectl get nodes — you should see CoreWeave GPU nodes listed with their GPU type.
Choosing the Right GPU Instance
| GPU | VRAM | Price/hr | Best For |
|---|---|---|---|
| H100 SXM5 | 80GB HBM3 | $2.06 | Large model training (7B+) |
| A100 80GB | 80GB HBM2e | $2.21 | Training, fine-tuning |
| H100 NVL | 94GB HBM3 | $3.54 | Very large models (70B+) |
| RTX A6000 | 48GB GDDR6 | $1.28 | Fine-tuning, inference |
| L40S | 48GB GDDR6 | $1.33 | Inference serving |
For the theoretical background on why H100s are so much faster than older GPUs, see our guide on Groq vs Nvidia AI inference 2026 which explains GPU memory bandwidth architecture in detail.
Submitting a Training Job
Cost Optimization Tips
- Delete jobs immediately after completion — you're billed per-minute, idle GPU pods still cost money
- Use spot/preemptible instances for non-critical training runs — 50-70% cheaper, can be interrupted
- Right-size GPU count — 8x H100 for large models, 1-2x A100 for fine-tuning smaller models
- Store data efficiently — use CoreWeave's high-speed NVMe storage close to GPU nodes, not external S3 buckets which add latency and transfer costs
For full cost breakdown and startup-specific pricing, see our CoreWeave pricing guide for startups.
FAQs
Basic kubectl knowledge is sufficient. CoreWeave provides well-documented YAML templates for common AI workloads — you don't need to understand Kubernetes deeply to submit training jobs. The key commands are apply, get, logs, and delete. CoreWeave's documentation includes copy-paste examples for PyTorch, TensorFlow, and JAX training patterns.
Three options: (1) CoreWeave PVCs (Persistent Volume Claims) — fastest, data lives on CoreWeave's NVMe storage. (2) Mount S3-compatible buckets directly in pods — works well for large datasets already in S3. (3) Rclone/rsync to copy data to PVC before training. Option 1 gives the best training throughput since storage is local to GPU nodes.