hivenet run --gpu a100 --image pytorch/pytorch:latest --volume ./my_model:/workspace In 11 seconds, she had a shell. No SSH key management. No waiting for “provisioning.” She was inside the container. nvidia-smi showed a glorious, cold A100 staring back at her.

Most tutorials start with “Verify your identity.” Hivenet’s tutorial began with a download button. She installed the Hivenet CLI via a single curl command:

But then a warning popped up: “Provider has a 4-hour uptime guarantee. Session is ephemeral.” Panic. “What if Iceland goes offline?” She read the rest of the tutorial: State management. She learned to use Hivenet’s native volume snapshots. Every 10 minutes, her checkpoints automatically streamed to a decentralized IPFS-backed store.

The tagline read: “Decentralized GPU compute. No hidden cloud tax.”

She copied her training script over. It ran. It screamed. 1,200 tokens per second. At this rate, the 72-hour job would finish in 40 minutes .

Thirty-eight minutes later, the console printed: Training complete. Accuracy: 94.2% She paid $0.56. No egress fee to download the model. She shut down the instance, and the A100 in Iceland immediately returned to its owner for someone else to use.

Maya stared at the timer on her local laptop. 72 hours left until her grant proposal deadline. Her personal RTX 3060 had been chugging for 14 hours just to complete 3% of the LLM fine-tuning. At this rate, her model would finish training sometime next winter.

It didn’t mention that she would later use Hivenet to spin up 10 H100s for a distributed training run across three continents for less than the price of a pizza. But that’s a story for another deadline. Moral of the tutorial: Hivenet turns “I can’t afford an A100” into “I just borrowed one from Iceland.”